This PDF file contains the front matter associated with SPIE Proceedings Volume 13181, including the Title Page, Copyright information, Table of Contents, and Committee Page.

To address the issue of sparse semantic annotation in images under complex weather conditions, a common approach is to adapt models trained on datasets collected under normal weather conditions to datasets collected under complex weather conditions. This adaptation often involves introducing reference images to enhance segmentation accuracy. However, aligning only the target and reference images neglects contextual knowledge and fails to alleviate the domain gap caused by complex weather conditions. In contrast to previous domain alignment approaches, this paper focuses on learning domain-invariant features from the source domain. It aims to reduce domain discrepancy at both the input and output levels. Firstly, at the input level, the quality of target images is dynamically improved using certain prior information. Secondly, at the feature level, domain-specific knowledge, such as contextual relevance, is extracted. Finally, at the output level, a learnable guided filter is introduced to enhance target saliency and preserve segmentation edges. Without increasing training time, this paper further improves performance, and experimental results validate the effectiveness of the proposed method, achieving a mIoU of 66.9% from Cityscapes to ACDC.

So far, pottery to create art has a history of thousands of years. Ceramics, sculpture art not only conforms to the trend of the times, but also into the characteristics of the times, in their own creation and development of the brand of the times. This paper summarizes the style of Chinese contemporary ceramic sculpture by analyzing the development history and current situation of Chinese ceramic sculpture, and giving in-depth thoughts to the future development trend. Combining factor analysis and K-mean algorithm, this paper can comprehensively identify and categorize the styles of ceramic sculpture. Firstly, the factor analysis method is used to find out the potential correlation between different styles, and then the K-mean algorithm is used to classify the sculptures of different styles, so as to realize the comprehensive identification and classification of ceramic sculpture styles, and to provide more accurate and comprehensive data support for the research of culture and art. The experimental results proved that the accuracy, checking rate and F1 value of K-mean algorithm are 98%, 87% and 56% respectively, which are better than the classification effect of SVM (support vector machines) algorithm.

Elevating energy efficiency (EE) poses significant challenges, primarily associated with implementation complexity and energy consumption. In this paper, we explore a type of hybrid precoder under grouped-adaptive-connected (GAC), In this configuration, divide the antennas into groups and switches dynamically control the links between Radio Frequency (RF) chains and antenna groups. Within the GAC structure, we formulate the design of hybrid precoder as an optimization problem aimed at maximizing EE. Additionally, we divide the task of maximizing EE into two problems of spectral efficiency (SE) maximization and combinatorial optimization. To address the issue, this paper introduces an LEO satellite hybrid precoding algorithm. Additionally, building upon the solution obtained from last problem, we propose a GAC hybrid precoding algorithm (GACHP) based on cross-entropy. Results of simulation indicate that under the GACHP algorithm, system parameters can be configured and adapted to maximize EE of system.

In recent years, the ultraviolet (UV) band in the solar blind zone has become an important research point in the field of wireless optical communications due to its low background noise, non-line-of-sight (NLOS) transmission and high confidentiality. However, UV communication faces challenges such as low data transmission rates and high bit error rates (BER). This study, drawing on the features of layered space-time code detection algorithms, proposes a hybrid space-time coding scheme suitable for UV communication. The scheme is evaluated through simulation analysis of various detection algorithms and modulation methods, demonstrating its effectiveness in reducing BER and enhancing overall system performance.

The arrayed waveguide grating (AWG) is a promising device which can be integrated on chip to achieve multi-wavelength optical processing. Herein, two kinds of 32-channel AWGs with 100 GHz channel spacing and operating in C-band are designed, namely the arc-type AWG and rect-type AWG. Different device structures are introduced including the arrayed waveguide quantity, arrayed waveguide width, transition connector, and routing strategy. The performance is investigated by conducting numerical simulation. For both the AWGs, simulation results show that the crosstalk, insertion loss and channel uniformity are less than -32dB, 6 dB, and 2.5 dB. For the arc-type AWG, it shows a more compact footprint due to the use of 500-nm arrayed waveguides, and the 3 dB bandwidth of each channel is wider compared to the rect-type AWG. For the rect-type AWG, the arrayed waveguides are widened and straightened. It is expected to have higher process error tolerance. Euler bending is also applied to reduce bending loss. The simulation results show that the proposed devices are desirable to be put into use in dense wavelength division multiplexing scenarios such as optical communication and microwave photonics systems.

Facial expression recognition is cutting-edge research in the field of computer vision, aiming to deepen the understanding of human emotions. With the popularity of social media and virtual interactions, there is an urgent need to accurately capture facial expressions to improve the user experience. In this paper, we summarize the research status of facial expression recognition, discuss the challenges faced by related technologies, and propose an improved VGG facial expression recognition method based on attention mechanism (AMVGG) to improve the recognition performance and promote the further development of this field. In this paper, it is proposed to embed multiple CBAM modules in the improved VGG network as the backbone network, and introduce deep attention center loss combined with softmax loss function to optimize the network. The experiments on the classical dataset FER2013 and RAF-DB achieved 73.81% and 88.76% accuracy, respectively, which were higher than those of other typical recognition algorithms.

Satellite Internet technology has emerged as a critical player in providing global connectivity, particularly in areas where traditional terrestrial infrastructure is challenging to deploy. Satellite Internet enables high-speed data transmission by leveraging communication satellites orbiting the Earth. These satellites serve as crucial nodes in a global network, facilitating internet access in remote or underserved regions. However, the inherent challenges of satellite communication, such as latency and energy limit, have prompted the development of advanced routing techniques to optimize data flow. To improve satellite Internet performance, this paper proposes a geolocation-based routing system. Furthermore, based on the system, we propose an energy-efficient distributed routing algorithm. The integration of Georoute in satellite Internet networks contributes to improved load balancing and fault tolerance. The dynamic adaptation of routing paths based on real-time conditions ensures efficient data delivery and minimizes the impact of disruptions. As the demand for reliable and high-performance global connectivity continues to grow, satellite Internet combined with Georoute emerges as a viable solution to bridge the digital divide and provide seamless internet access across diverse geographical landscapes. To demonstrate the effectiveness of our algorithm, we build a testbed driven by public constellation information and implement the prototype on top of the testbed. Extensive realistic-data-driven simulations demonstrate that our algorithm can significantly improve the network throughput and reduce the satellite energy consumption as compared to other approaches.

In the era of we-media big data and AIGC, various kinds of network false information emerge in an endless stream. How to effectively control the spread of false information and its variants in the Internet has become the focus of attention of network management agencies at all levels and new media in all walks of life. This paper starts with various models and methods to maximize the influence of network communication, studies the key nodes and critical paths in network information communication, and designs a dynamic multi-objective control optimization algorithm for false information propagation based on such key nodes and paths, so as to effectively control the spread of network false information in the Internet we-media era. By testing the influence of network communication of specified information, dynamic multi-objective optimization control algorithm can effectively reduce the penetration rate and propagation speed of false information. Practice has proved that the relevant propagation control algorithm has a remarkable inhibition effect on the network transmission of specific information, and the inhibition effect is more than 15% higher than that of other similar algorithms.

With the increasing demand for computing power in machine learning tasks, the training of deep neural network models has been pushed to multi-GPU training or even larger scale distributed training. However, the acceleration effect and scalability of model training are largely limited by the communication efficiency between GPUs. In order to improve the communication efficiency of domestic GPU accelerator, this paper studies and analyzes the communication performance of Allreduce operator widely used in deep learning tasks. Based on a data compression algorithm and multi-stream parallel technology, this paper ports and optimizes the Allreduce operator on domestic heterogeneous platform. The experimental data results show that compared with RCCL, the ported optimized version of Allreduce operator achieves different degrees of performance improvement such as 30% to 90% under different data sizes. This paper implements the transplantation and optimization of Allreduce operator on the domestic accelerated heterogeneous platform, and achieves good acceleration results, which provides support for the efficient communication of domestic GPU accelerators and the ecological construction of domestic heterogeneous platforms

The underwater acoustic channel under ice cover is a complex and variable channel with various interferences, which greatly limits the transmission performance and capability of communication signals. This article introduces energy monitoring technology into conventional underwater spread spectrum communication systems and explores a robust underwater spread spectrum communication model under ice cover. Enhance the anti-interference and multipath resistance of underwater acoustic communication systems through dynamic energy monitoring, comparison, and judgment processes. The communication system experiment was carried out in the underwater acoustic channel covered by ice, achieving high-quality communication transmission with zero error code and verifying the reliability of the system.

The Time of Flight Diffraction (TOFD) method stands as a highly efficient, cost-effective, and environmentally friendly approach for defect detection within the domain of ultrasonic testing. This method proves particularly efficacious in discerning weld seam defects within mechanical welded structures. The D-scan image assumes a pivotal role in detecting flaws within low-alloy thick plate butt welds. In D-scan images, a considerable portion of the data, under reasonable scanning depths, serves as background rather than conveying defect information. This background, primarily noise, acts as non-target interference and significantly disrupts defect identification. Particularly, when the probe detection angle approaches the surface, lateral waves emerge as relatively stable and continuously distributed noise in the image, characterized by trailing tails in time. This characteristic makes them susceptible to overlapping with signals from near-surface defects, resulting in blind spots during detection. To enhance the clarity of D-scan images, this paper introduces an expeditious and potent technique for lateral wave straightening and removal. the proposed method relies on the peak detection to A-scan and correlation analysis of A-scan signals for the meticulous rectification of D-scan signals. Moreover, it incorporates an insightful examination of image energy distribution and one-dimensional wavelet transform to effectively remove lateral waves, facilitating the generation of clear and discernible images of weld seam defects.

This paper elaborates the basic principle of direct sequence spread spectrum communication, analyzes and compares the efficiency of typical jamming methods such as aimed jamming and barrage jamming to direct spread spectrum communication system, and carries out corresponding simulation using Simulink platform. It is concluded finally that the optimal jamming method is the barrage jamming with the optimal jamming bandwidth, which is of guiding significance for direct sequence spread spectrum communication without narrowband suppression technology in practical communication confrontation.

Seed quality is a key factor in wheat germination rate, total yield and other indicators, and is of great importance for food security. For a long time, wheat seed selection and breeding have been heavily dependent on manual experience selection, which is characterized by high labor costs, unstable detection accuracy and high subjectivity. In this paper, we propose a machine vision recognition algorithm model WSEED-YOLOV8 for healthy wheat seeds improved from YOLO V8, which significantly improves the model's ability to extract features for wheat seed categories and whether wheat seeds suffer from Sunn Pest (Eurygaster integriceps) damage by introducing Deformable Convolution V2 (DCNv2) and encapsulating it in the C2f_DCNv2 module instead of the second, third and fourth C2f modules in YOLO V8n. By introducing the Polarized Self Attention mechanism module, with a slight increase in the parameters, it reduces the omission of false detections in the case of dense stacking of wheat seeds in the same image; and by using the Shape Loss Shape-IoU function as the IoU calculation function of the bounding box, which improves the target detection ability in the case of tilted wheat placement. The experimental results show that the WSEED-YOLOV8 wheat healthy seed recognition algorithm model achieves a recognition precision of 95% on the same wheat seed dataset, which is 15.9% higher than the YOLO V8n baseline precision. An 11% improvement in recognition precision was achieved with a similar number of parameters as the YOLO V8s baseline.

With the acceleration of urbanization, the quality of groundwater has been seriously threatened, and impurity identification has become an urgent problem to be solved. Based on big data fusion technology, this study uses machine learning and deep learning algorithms to intelligently identify groundwater impurities. Firstly, groundwater quality data, including turbidity, total suspended matter, pH value, dissolved oxygen and other indicators, were collected, and pretreatment and feature extraction were carried out. Then, support vector machine, random forest and neural network algorithms are used to build a classification model to classify and predict the impurities in groundwater. The experimental results show that the groundwater impurity identification method based on big data fusion has high accuracy and stability, and can provide strong support for groundwater quality monitoring and treatment. This study not only enriched the theoretical system of groundwater impurity identification, but also provided an operable scheme for practical application.

For high resolution 3-D ultra-wideband(UWB) Through-Wall Imaging (TWI), one of the technical challenges is long data collecting time. In this paper, the application of the Compressive Sensing (CS) method to 3-D point-like targets imaging for stepped-frequency UWB TWI is discussed. By solving a constrained convex l1-norm minimization optimization problem, only a small number of data, which is collected by randomly selected frequencies and randomly selected antenna scan positions, is sufficient to obtain the sparse targets image. Simulations indicate that the targets can be truly imaged by the CS method with only measuring 7% of all spatial-frequency data even with 10% additive Gaussian white noise.

As a basic subtask in natural language processing, text enunciation recognition aims at determining the enunciation relationship between presupposition and hypothesis. At present, most deep learning text sensitivity models use single interaction and single inference to process the semantic features of the text, which is not sufficient and comprehensive for semantic mining of Chinese text. Therefore, a multi-round and multi-interaction inference model for Chinese text implication (MM-CIM) is developed in this paper. On the one hand, different mutual attention mechanisms are used for the premise and hypothesis to generate different interactive alignment semantic features to expand the model's exploration of text semantics. On the other hand, the coding features and mixed features of the premise and hypothesis are deduced in turn to deepen the model's exploration of text semantics. In addition, k-max pooling is introduced into the pooling part of the model to further expand the semantic mining ability of the model. The experimental results on Chinese text contained data show that all parts of the proposed model are optimized effectively and have better recognition effect than the basic model.

The Wi-Fi-based wireless sensing technology has emerged as a research hotspot in the field of perception in recent years, enabling intelligent sensing of human activities and the surrounding environment. However, existing wireless sensing models exhibit a high number of parameters, making real-time perception challenging, especially in scenarios with limited computational resources such as mobile devices. To address this issue, this paper proposes a classification recognition algorithm based on a hybrid approach that combines Spatially Separable Convolution (SSC) and Stacked Gate Recurrent Unit (SGRU). In the shallow layers of the hybrid model, the algorithm utilizes a feature extraction module composed of spatially separable convolution to capture spatial features of human gestures while maintaining the temporal consistency of features. In the deeper layers, SGRU network is employed to learn the spatiotemporal features of gestures. The SGRU consists of two layers, where the output of the first layer serves as the input to the second layer. Through validation on the open-source Widar dataset of human gestures, this paper demonstrates that the proposed SSC-SGRU, when compared to the comprehensive ResNet18 with a reduction of approximately 11.0M parameters, achieves an accuracy improvement of approximately 7.2%.

In recent years, deep learning has shown remarkable success in the field of object detection, with the YOLO series models known for their speed and effective recognition. A classroom behavior recognition model based on bidirectional feature fusion is proposed, enhancing the YOLOv5 model by incorporating bidirectional feature fusion, attention mechanisms, and modifying the loss function. Experimental results indicate that the improved model achieves a maximum average recognition accuracy of 97%, demonstrating a notable increase of 4.8% compared to the original model, highlighting a significant improvement in accuracy.

The demand for wide bandwidth and high efficiency RF transmitter modules is growing in various fields, and the difficulty of power amplifier research also lies in meeting the design requirements of wide bandwidth and high efficiency at the same time. In this paper, an ultra-wideband power amplifier operating at 0.5~2.5 GHz is designed based on the gallium nitride high electron mobility transistor CGH40010F. In order to achieve higher efficiency over a wide bandwidth, an improved simplified real-frequency method is used to optimize the design of the output matching circuit. Simulation results show that the saturated output power exceeds 40.4 dBm within 0.5~2.5 GHz, the drain efficiency ranges from 55.3% to 71.4%, the gain ranges from 12.4 dB to 14.4 dB, and the gain flatness is ±1 dB. it meets the requirements of the transmitter system

Aiming at the problems of cost limitation and high complexity in near-field security imaging for uniform array antennas, a strong sparse low sidelobe sparse array design algorithm is proposed. Firstly, according to the physical characteristics of sparse array, a joint optimization problem is established which minimizes the 1 _p (0< p <1) norm of array excitation to the objective function and is constrained by the level of main sidelobe. Then, the splitting variable of the main sidelobe constraint is introduced to separate the objective function from the level constraint of the complex main sidelobe, and the problem is simplified and solved by using the equality transformation relation. Then the complex derivative method is used to approximate the array excitation iteratively. Finally, the Alternating Direction Method of Multipliers (ADMM) is used to achieve the cooperative solution of the main sidelobe constraint and the array excitation minimization. The experimental results show that the method can obtain lower sidelobe level and fewer array elements without the main lobe broadening.

This paper takes a pioneering step by introducing Radio Frequency Identification (RFID) technology into the management of nautical charts, aiming to alleviate the challenges of labor-intensive processes, heavy workloads, and error-prone tasks. This technology involves equipping traditional nautical chart cabinets with RFID systems. These systems leverage binary tree anti-collision algorithms to accurately scan densely labeled chart tags, swiftly identifying nautical chart information. Experimental findings reveal that the main factor affecting the probability of successful detection of nautical charts is the transmit power of the reader rather than the tag sensitivity, when the RF transmit power is set to 27dBm, the success probability of nautical chart detection is optimal. And with the number of paper nautical charts stored in each drawer limited to 100 sheets or fewer, the inventory time for each drawer remains below 1.2 seconds, achieving a detection success rate exceeding 99.6%. The utilization of this technology satisfies the demand for rapid identification of paper nautical charts. It can remarkably contribute to the informatization construction of nautical chart management units and serve as a valuable model for managing data stored in dense configurations.

In the application of computer vision recognition, the recognition objects of some tissue-like structures present highly complex and variable characteristics. For this practical usage scenario, traditional classification algorithms usually cannot achieve the desired recognition effect. To achieve better recognition performance in this direction, this study integrates ASFF into the YOLOv5 model and introduces the SimAm attention mechanism module. The improved network structure is more lightweight, has fewer network parameters, and the attention mechanism is more effective. We trained the modified YOLO model on datasets, and experimental results show a significant increase in mAP value and a 20% increase in accuracy. This indicates that improvements to the model can significantly enhance performance.

Aiming at the problems that the existing navigation line extraction algorithm of the automatic navigation system based on machine vision is susceptible to the interference of the external environment and the slow processing speed, the paper proposes a navigation line extraction method based on the linear scanning of the image. The linear scanning navigation line extraction algorithm proposed in the paper takes the bottom edge of the image as the starting point selection range of the crop line straight line, and the top edge as the end point selection range of the straight line, selects two points at each of the top and bottom edges of the image, and then controls the upper and lower end point positions to produce straight lines with different slopes, counts the number of target points that fall within a certain region of the straight line, and finally takes the straight line that contains the largest number of target points as a straight line of the crop line and calculates a navigation line based on the crop line straight line to obtain a navigation line. Finally, the line containing the most target points is taken as the crop line line, and the navigation line is calculated according to the crop line line.

Currently, the identification and counting of birds in ecological monitoring are mainly done manually, which is a laborious and challenging process particularly when dealing with dense bird populations. In this paper, we propose an end-to-end multitasking neural network for identifying and counting dense bird populations. The network consists of three different branches, which respectively complete the tasks of population center position prediction, population boundary prediction, and counting. To achieve collaboration between multiple tasks and improve the performance of the model, a multitasking joint loss calculation strategy is proposed by utilizing the constraint relationships between multiple tasks. Additionally, we propose a method of population region mask, which can cope with the simultaneous emergence of multiple populations in actual scenarios. In order to validate the effectiveness of our method, we construct the first dense bird population identification and counting dataset and conduct thorough experiments. Compared to the baselines, our method has faster inference speed, lower resource consumption, and better identification and counting performance.

For improving the insufficient feature extraction and recognition capability of the traditional VGG16 network in the human facial expression recognition task, the study proposes a VGG16 network recognition model based on an SGE (Spatial Group Enhance) attention and feature fusion improvement, which firstly extracts the last three Block extracted from the VGG16 network and then introduces the SGE attention module to enhance the attention mechanism by spatial grouping. features are fused, and then the SGE attention module is introduced to enhance the attention mechanism by spatial grouping. Finally, the recognition effect of the improved model is analyzed, and the results show that the improved VGG16 network model improves the human facial expression recognition accuracy by 6.9%, 7.8%, and 7.4% on Cohn-Kanade, Multi-Media Interface, and AffectNet datasets, respectively, and the recognition accuracy is about 90% or more for eight common human facial expressions. The recognition accuracy for eight common human facial expressions is about 90% or more. The above data show that the research model can recognize human facial expressions more accurately.

Methods for recognizing 3D models from multiple views have shown remarkable performance in 3D model recognition. The primary challenge faced by these methods is integrating features from multiple views into a single, highly descriptive 3D shape descriptor. This paper proposes a Drop-based Multi-view Recognition Network (DropNet) for accurate recognition of multiple views of 3D models. The proposed DropNet first extracts multi-view features using a backbone network, and then randomly discards individual view features through the DropView module, which makes the number of view combinations that can be learned by the network grow at the level of combinatorial numbers, which can effectively improve the network's ability to recognize multi-views, and at the same time, improve the robustness of the network. Subsequently, the remaining view feature values are randomly discarded by the DropPoint module, which further enhances the robustness of the network and effectively prevents overfitting. The experimental results on benchmark datasets show that the DropNet outperforms several recent approaches in terms of multi-view recognition performance.

In the industrial field, Quick Response code has been widely used in the automatic management of production lines because of its high-density storage and fast reading characteristics. However, in real-world scenarios, the recognition success rate of Quick Response code often falls short of the desired outcome due to interference from factors such as lighting, noise, and scale variations. To solve this problem, this paper proposes an Image Enhancement-based Quick Response Code Recognition in complex working conditions (IE-QRCR). Firstly, the Quick Response code image is preprocessed through the joint optimization method of perspective transformation and low-light enhancement to improve image quality and reduce interference. Then, the optimized Quick Response code image is input into the recognition algorithm for recognition. The method can effectively improve the accuracy and robustness of recognition, especially in complex environments with low illumination and poor image quality. We verify the effectiveness of the method on some public Quick Response code datasets, which provides a useful reference for further research in the field of Quick Response code recognition.

With the continuous development of the film industry, audience expectations and demands are increasing. Film reviews, as crucial sources of feedback, play a significant role in the film market. Through data mining of film review data, valuable insights can be obtained regarding audience evaluations and preferences, facilitating informed decision-making in the film industry. This study focuses on analyzing film review data obtained through Scrapy technology from the TOP250 Douban movies. The Word2Vec and LSTM is employed to achieve sentiment classification of movie reviews. To evaluate the model's classification ability, a comparative analysis is conducted with several mainstream machine learning models. Experimental results demonstrate that the Word2Vec-LSTM model exhibits superior classification performance, providing film industry practitioners and researchers with valuable insights into audience perceptions of movies. This study contributes to a better understanding of audience emotions and offers meaningful contributions to the film industry.

This study discusses a new method combining image steganography technology with Natural Language Processing (NLP) large models, aimed at improving the accuracy and robustness of extracting steganographic text. Traditional Least Significant Bit (LSB) steganography techniques face challenges in accuracy and robustness of information extraction when dealing with complex character encoding, such as Chinese characters. To address this issue, this study proposes an innovative LSB-NLP hybrid framework. This framework integrates the advanced capabilities of NLP large models, such as error detection, correction, and semantic consistency analysis, as well as information reconstruction techniques, thereby significantly enhancing the robustness of steganographic text extraction. Experimental results show that the LSB-NLP hybrid framework excels in improving the extraction accuracy of steganographic text, especially in handling Chinese characters. The findings of this study not only confirm the effectiveness of combining image steganography technology and NLP large models but also propose new ideas for research and application in the field of information hiding. The successful implementation of this interdisciplinary approach demonstrates the great potential of integrating image steganography technology with natural language processing technology in solving complex information processing problems.

With the sharp change of global climate, NILM (non-intrusive load monitoring) has been a promising approach to reduce energy consumption and carbon emissions. This paper presents a new load identification method using sequence labeling to identify multiple variable-length events within the sliding window of aggregate power signal simultaneously. Furthermore, the proposed method integrates local features of events themselves with temporal features of the multiple events through CNN (convolutional neural network) and attention network. The experimental results on two public datasets REDD and UKDALE indicate that the proposed method achieves better performance than the existing methods.

It is a very worthy problem to efficiently identify a large amount of personal mobile image data. The core task of this problem is to automatically and efficiently screen out the pictures that may involve sensitive marks. This paper proposes a sensitive identification method based on image semantic segmentation to realize automatic detection of personalized image sensitive identification. This method utilizes the advanced visual features of images and obtains the description tag set of images through semantic segmentation. By using this tag set, the image sensitive identification is detected from two granularity of context and element. At the same time, the sensitive identification of users is mined. In order to test the validity of the method, 2493 pictures from users were collected and labelled, and the identification accuracy of image sensitive labels reached 89.3%. Finally, based on the above detection method of image sensitive identification, the user's personal photos are quantified and the image sensitive identification protection scheme for different scenes is proposed.

Edge computing is one of the important techniques in visual recognition system. In this paper, the application of edge computing in visual recognition system is introduced by taking the automobile water pump screw recognition as an example. The image is captured by an industrial CMOS camera, the workpiece image is adaptively filtered, and the image RGB to HSV conversion and equalization is carried out based on the feature color, which overcomes the interference of uncertain external factors; the edge computing method based on the feature color realizes the target localization, and the screw recognition method based on the Hough transform is designed, and tested and verified within the QtCreator platform. The results show that the proposed method can effectively recognize screws in the industrial production process with good detection accuracy and real-time performance.

Current work on Chinese medical name entity recognition focuses on extracting flattened entities, and when there are nested entities in the entities, the entity recognition is prone to errors and cannot accurately identify all the entities in the statement. In this paper, we construct an ALBIG model based on multilayer neural network and adversarial learning method. The model uses a pre-trained model RoBERTa-wwm stablish word embedding and connects IDCNN model to extract text features. Adds adversarial learning method to increase the robustness of the model and uses Global Pointer method based on Rotary Position Embedding as the output layer to obtain the final result. On the CMeEE dataset, the F1 values of the ALBIG model obtained higher scores compared to other experimental models, proving the validity of the model.

Person name is an important component of Tibetan named entities. Accurately identifying person name in text is a fundamental task in Tibetan natural language processing, and is a key issue in research tasks such as question answering systems, information extraction, machine translation, and knowledge graphs. This paper analyzes the characteristics of Tibetan person name and incorporates their morphological, word length, and positional features into the BiLSTM-CRF model. This model can effectively capture contextual information in sentences and annotate the sequence of person name syllables. The experiment was conducted on some corpora of Xizang Daily, and the results show that, compared with CRF model, the F₁ values have increased by approximately 5%. Indicating that the proposed Tibetan person name recognition method based on the BiLSTM-CRF model is effective.

In order to reduce the impact of public health emergencies on the normal life of residents, at the same time to prevent the secondary spread of diseases caused by improper methods in the management process. On the premise of reducing the overall operation cost as much as possible, based on the basic data such as the number of quarantinable population, the number of daily necessities distribution points, and transportation network of 9 administrative areas in Changchun during the epidemic isolation and prevention period in 2022, this paper uses the entropy weight method to score each community and set the number of delivery points, so as to judge whether the actual plan is reasonable. In order to further optimize, with the minimum service cost of all the selected delivery points and the premise of meeting the distribution demand, save as much as possible, reduce the number of sites, and minimize the site radius of each sorting point as the goal, the total number of selected delivery points should not exceed the set total. A multi-objective site selection planning model was established for constraints. Sensitivity analysis was performed on the number of sites to determine the optimal number of sites, and cluster analysis was used to obtain the optimal site selection coordinates and cell numbers corresponding to the drop points. As a result, the number of sites was greatly reduced and the efficiency of material distribution was improved. Considering that the less people flow in the context of the epidemic, the better, this paper also uses the improved k-means algorithm for the actual plan, and finally gets the following number of delivery points in each district: 168 in Erdao District, 53 in Jingyue District, 239 in Nanguan District, 292 in Kuancheng District, 303 in Chaoyang District, 25 in steam open district, 69 in economic open District, and 222 in Green Park.

Airborne front-end intelligent image recognition is one of the key technologies for unmanned aerial vehicle inspection of transmission lines. It is fundamentally different from traditional image recognition and judgment methods. This article implements convolutional neural network mechanical learning based on artificial intelligence technology. Four models were established for equipment identification, tower detection, wire detection, and insulator string detection, which were trained and tested separately. Taking into account both training time and actual usage efficiency. This article recommends a basic training sample size of 20000 for the model. Comparing two different neural network structures, it is found that the model with a pooling value of 2 and a 1.28 million pixel input has higher recognition accuracy. The training results of different models indicate that for different equipment recognition models, the training amount required to achieve stable recognition accuracy is from small to large, in order of wires, towers, equipment, and insulator strings.

In this paper, an human activity recognition model named Improved Time-Dependence Convolutional Neural Network is proposed to improve the human-robot interaction performance of limb exoskeleton robots. Firstly, select the time series signal of human lower limb knee joint as data set, use the One-Dimensional Convolutional Neural Network to extract the motion signal feature and reduce the dimension. Then, the Gate Recurrent Unit is combined to learn the long-term time dependence of motion signals and the relationship between potential features and target output. Finally, Residual network unit is introduced to help train the deep network and improve the stability of the network training process. Built the relevant model and compared their recognition performance. The results shown that the recognition rate of ITD-CNN recognition network model for different human activities is more than 99.97%, which has a high application value in the recognition of human activity state in the exoskeleton robot field.

In complex outdoor environments, processing a large amount of point cloud data and restoring the precise structure of real scenes is a key difficulty in real scene reconstruction. To this end, a laser 3D real scene reconstruction and lightweight access method based on LiDAR point cloud is proposed. Using region growing algorithm to segment laser foot points of buildings in 3D reality and obtain LiDAR point cloud data of 3D reality. Using the Euler Lagrange equation to establish a 3D data field for laser 3D real scene image reconstruction, and completing the 3D reconstruction of laser 3D real scene images. Using SSD lightweight algorithm to perform regression classification on the extracted target features, in order to locate targets of different scales and achieve lightweight access to 3D real-world image targets. The experimental results show that the research method can accurately reconstruct three-dimensional real-life buildings with lower reconstruction errors.

In response to the high manual classification intensity, low efficiency, and poor accuracy in current citrus defect detection, a rapid and accurate citrus defect detection solution based on machine vision technology is proposed. The main tasks include: (1) Creating a citrus dataset with defects on a clean background and performing white balance preprocessing. (2) Utilizing an improved deeplabv3+ network for recognition under the PyTorch framework. (3) Deploying the network to an embedded platform for practical detection. The results show that the citrus surface defect detection system achieves a defect recognition accuracy with a Mean Average Precision (MAP) value of 70.5%. The recognition accuracy is better than the pre-improvement effect without affecting the detection time. The proposed method effectively identifies defects on the surface of citrus. The deep learning algorithm is combined with embedded devices to realize the accurate identification of citrus defects.

Recent research on Intelligent Reflecting Surface (IRS), a promising emerging technology, has shown significant potential in positioning and sensing. This paper provides an overview of the current status of the IRS technology, including theoretical research, performance analysis, localization algorithms, and the interaction between localization and communication. Furthermore, a IRS-based single base station localization algorithm is proposed for single base station localization. The effectiveness of the algorithm is verified through simulations, demonstrating that sub-meter level localization can be achieved even under moderately favorable system parameter.

Semantic understanding of point clouds is a major task in 3D scenes. Due to the enormous number of points within point clouds, it is strait to process point clouds directly. Generally, previous work follow a sampler-extractor structure to get high-level semantic features. Previous sampling strategies in the sampler stuck in upholding local density while guaranteeing a predetermined number of sampled points. We propose a density-adaptive sampling strategy. In the extractor, PointNN introduces a non-parameter framework. However, it ignores the positional distance among points. We introduce a position-sensitive feature extraction module inspired by PointNN. We conduct comprehensive experiments to validate the effectiveness of our method.

With the development of modern technology, the application of hosting systems in hosting institutions has become widespread, and the check-in and check-out function as an important component of the hosting system. This paper analyzes and studies the check-in and check-out function of the hosting system, explores its design principles, implementation methods, and application effects, and introduces facial recognition technology to achieve the collection and verification of student identity information, ensuring the authenticity and accuracy of check-ins and check-outs.

Harmonic radar is a kind of radar that uses the higher-order harmonics reflected by the target for nonlinear target detection. Eavesdropping equipment usually has non-linear devices, so harmonic radar can be used in the field of anti-investigation to detect voice recorders and eavesdroppers. In this paper, a kind of harmonic radar based on dual-transmitter and dual-receiver is designed to meet the needs of anti-investigation of eavesdropping equipment, and three kinds of electronic targets such as recording pen are tested and verified. The verification results show that the device can detect nonlinear electronic targets well.

The core problem faced by coal mining is coal rock identification. In order to improve the accuracy of coal rock identification, this paper proposes that radar antenna frequency and detection height have an impact on the identification accuracy of coal rock interface. First of all, using matlab simulation, simulation results show that under different frequencies, the identification accuracy of air-coal seam interface increases with the increase of frequency, but the identification of coal rock interface only in the frequency is low only partial error. Secondly, under different detection heights, the identification accuracy of air-coal seam interface becomes more and more accurate with the increase of detection height, but in this experiment, the identification accuracy of the coal-rock interface is the same, which cannot be ruled out because of the insufficient experimental data, the problem still needs to be verified by a large number of experiments. The above research shows that the antenna frequency and detection height of ground penetrating radar have a certain impact on the identification of coal and rock.

In order to study the factors affecting the identification accuracy of coal-rock interface with ground penetrating radar, the finite-difference time-domain (FDTD) method is used to study the law of dust interference, water content of medium and dirt separation under different conditions. The excitation source is 900MHz Blackman-Harris pulse, and Mur twodimensional absorption boundary conditions are adopted to establish different radar forward modeling models. The simulation results show that: Air dust can absorb or scatter the energy of radar signal. The relative dielectric constant of air is set to 1, 2, 2.5, and the corresponding electrical conductivity is 1Ms/m, 4Ms/m, 10Ms/m. The results show that the amplitude of reflected wave decreases successively, and the accuracy of radar system detection decreases gradually. With the increase of coal water content, the energy of reflected wave decreases. The relative dielectric constant of coal seam is set as 3, 4 and 6. The results show that the amplitude of reflected wave decreases successively, which limits the detection range of radar system. When there is interference near the coal wall interface, the reflected wave of the coal wall interface and the reflected wave of the air-coal wall interface are superimposed, which brings challenges to the accurate identification of coal rock interface.

In recent years, urban traffic safety issues have increasingly garnered widespread attention, as traffic accidents pose significant threats to the safety of people's lives and properties. Driver's abnormal driving behaviors often occur in conjunction with traffic accidents, and the quantity of abnormal driving behavior data is significantly larger than that of accident data, making it a more comprehensive reflection of driving conditions and deserving in-depth analysis. Existing research rarely matches abnormal driving behavior data with map coordinates, and there is limited use of visualization methods to display the distribution of data points. This paper proposes an improved DBSCAN clustering algorithm based on abnormal driving behavior data and utilizes the minimum spanning tree algorithm to calculate the density of each clustering cluster. By computing the density of the clustering clusters, we can better understand the spatial distribution characteristics of abnormal driving behaviors, further analyze the degree of clustering of these clusters, and study the distribution of abnormal driving behavior.

In order to explore the characteristics of choosing passenger-seeking roads in the taxi passenger-seeking process and improve the operational efficiency of taxis, this paper proposes a method of mining the passenger-seeking road laws of high-occupancy taxis based on cross-tabulation analysis. The representative city of Chengdu is selected as the research area, and the stable taxi operation data in Chengdu is used as the research basis. Using the cross-tabulation analysis method, the correlation analysis between the number of passengers and the factors of road hotspot and road class is carried out in the form of Sankey diagrams, and it is found that the number of passengers is strongly correlated with the attraction of hotspot. In contrast, the correlation with the attraction of road class is weaker. We summarise the road selection rules for high occupancy taxis, such as "pay attention to road sections with many road hotspots," "pay attention to bus stop areas," and "pay attention to roads around primary roads," which are of strong practicality and reference value. It has strong practicality and reference value.

In which conditions should a Multiple Port Calling(MPC) mode switch to a Hub & Spoke (H&S) mode has brought about a huge challenge under the larger scale trend of container ships. The mode choice varying with the factors including portof- call frequency, ship size, and variable freight volumes at port. And there are trade-off relationships between the hub location choice and the freight rate of hubs together with the hub numbers. In the condition of remaining the service quality unchanged, considering the co-action complexity and trade-off effects mentioned above, a decision model for the evolution of container liner route mode under the upsizing of ships is proposed. Taking the mode choice, the ship size, and the hub numbers as decision variables, the maximum voyage profit rate as an objective, a three-stage mixed integer programming decision model is constructed. Also a heuristic three-layer cyclic iterative genetic algorithm is designed to solve the model. Finally, take a Far East -North American shipping route as an example, the optimized results and analysis verified the practicability and effectiveness of the model constructed and the algorithm designed in this paper.

With the surge of earth observation demands, the agile optical satellite scheduling has become a practical problem, which is hard to solve owing to the huge solution space and complex constraints. This paper proposes an agile optical satellite scheduling method based on Transformer networks to tackle this problem. A network model based on Transformer networks is proposed to generate high-quality solutions, and a masking mechanism is designed to improve the generation process. An Actor-Critic training algorithm is adopted to train the proposed network model. Experimental results show that the proposed algorithm outperforms three comparison algorithms in terms of training effect, solving ability, and computing efficiency.

The multiple uncertainties in real radar detection scenarios require that radar automatic target recognition systems must have efficient dynamic learning capabilities. And current deep learning methods are inefficient and impractical for incremental update learning of data. To address the problem of data incremental learning in radar target recognition, a network ensemble-based target incremental fusion recognition method is proposed by combining lightweight incremental learning with neural network fusion. Multiple convolutional neural networks are constructed to be trained and tested in the split-angle domains, and then integrated and fused to obtain a robust recognition performance close to that of large data volumes with significantly reduced data requirements. The experimental results based on the measured radar data show that the proposed method can efficiently accomplish the target incremental recognition task, and has strong engineering practicality.

After Bluetooth is turned on in electronic devices, when there are multiple devices for Bluetooth communication at the same time, signal interference may affect the accuracy of directional pairing. This is because the presence of interference sources may cause signal fluctuation, attenuation or distortion, which makes it difficult to judge the signal in the process of directional pairing. Therefore, the optimization method of directional pairing accuracy of Bluetooth communication between electronic devices is proposed. Using the delay of Bluetooth communication, the multi-path interference signal is identified, and the direct signal gain and Bluetooth communication interference are generated according to the signal direction and LCMV (Linearly Constrained Minimum Variance) signal waveform technology. Then the interference suppression of Bluetooth communication is realized. Based on this, the directional pairing of Bluetooth communication signals of electronic devices is completed through encryption and decryption of Bluetooth communication signals of electronic devices. The experimental results show that after interference suppression and encryption and decryption steps, the Bluetooth pairing of multiple electronic devices can be achieved with high precision.

In the era of sustainable intelligent agriculture, agricultural news appears to be intermingled with diverse, sparse, lack of datasets. In this context, the article proposes a text categorization model of ERNIE-BiLSTM-CNN incorporating an attention mechanism. The model adopts ERNIE as a language model, applies BiLSTM to extract the long-term dependent features of the text, adjusts its bias degree by introducing the attention mechanism, and finally uses CNN to extract the local features of the text. Experimental results demonstrate that the ERNIE-BiLSTM-CNN model is better for text classification.

As blockchain technology research continues to advance, named entity recognition within blockchain patent texts has emerged as a fundamental cornerstone for various other tasks. In light of the absence of a publicly accessible corpus within the current realm of blockchain technology, as well as the challenges posed by insufficient text semantic expression, the dearth of semantic features, and the intricacies associated with capturing diverse word vector representations. This paper us of annotated blockchain patent texts and proposes the MBBC MacBert-BiLSTM-CRF MBBC model for named entity recognition in blockchain patents. The model employs embedding the feature layer to vectorize the input text, enriching text information and mitigating polysemy issues. Text feature information is acquired through the long short-term memory network, and the conditional random field model is integrated to enforce label relationships, yielding an optimal label sequence. In this study, the accuracy rate of BAT (Blockchain Application Technology, BAT) on the self-constructed corpus achieves 91%, with a recall rate of 96% and an Fl value of 93%, representing a notable enhancement compared to mainstream models. Ablation experiments further demonstrate that the model in this paper effectively performs named entity recognition tasks associated with blockchain technology and subject content.

As one of the fundamental primitive operations, querying shortest distance between node pairs in a graph has gradually attracted great attention in both academia and industry. In recent years, with the rapid development of cloud computing, it allows clients to outsource their data to cloud servers. In this way, users can outsource large volumes of data which were difficult to store and compute on their own clients. However, it worth noting that for privacy and security reasons, user data needs to be encrypted before being transmission to the server. Meanwhile, it is crucial for the encrypted data transmitted to server to maintain availability in order to enable users' operations searching and querying. To preserve data privacy while enabling querying, clients must construct and transmit a secure index to the server for subsequent querying. When performing shortest distance queries on encrypted graph data outsourced in external storage, a significant challenge is computing the shortest distance in an efficient, accurate and secure manner. Furthermore, the challenge becomes more pronounced as the scale of the graph increases. In this paper, we propose a novel approach to address this issue by splitting the graph data into multiple segments and constructing 2HCL index LID within each subgraphs, as well as establishing an index LBD among these subgraphs. We also provide a comprehensive analysis of complexity and security of SSDAQ and present conclusion at the end of this paper. Compared to 2HCL, which is a classical and widely used algorithm, we reduce the time and space complexity from O(N2) time and O(kn|Lmax(sN)|) space to O(kn2) time and O(kn|Lmax(sn)|) space in setup phase.

In the context of the rapid development and widespread application of remote sensing technology, fine-grained aircraft classification is a research area with significant practical value. Most current classification algorithms, developed primarily for natural images, underperform when applied to remote sensing images. Additionally, fine-grained classification tasks inherently face challenges due to small inter-class differences and subtle discriminative features. To address these issues, this study designs an aircraft fine-grained network (AFG-Net) method, a network for fine-grained aircraft classification in remote sensing images. AFG-Net builds upon the ConvNext network, known for its superior classification capabilities over traditional CNNs and comparable to the Swin Transformer. Due to the influence of natural environment and complex imaging backgrounds on remote sensing images, this paper conducted data augmentation before training, which can help the network model better cope with interference and improve robustness. The following are the improvements made in this study: 1) Developing the ConvNext_s network, enhanced with the SimAm attention mechanism for better extraction of subtle, discriminative aircraft features. 2) Proposing a new composite loss function based on Mutual-Channel Loss, allowing the network to consider both global and local information more comprehensively, thereby improving aircraft classification performance and model robustness. 3) Demonstrating AFG-Net's applicability to fine-grained aircraft classification tasks in remote sensing images. Tested on the MTARSI and OPT-Aircraft_v1.0 datasets, AFG-Net achieves accuracies of 94.76% and 84.59%, respectively, outperforming existing advanced models in extensive experiments.

Well-trained models often face challenges in effectively processing data that diverges from their training distribution, a situation prevalent in real-world scenarios, particularly in the field of healthcare. In medical image segmentation, there is a pronounced need for domain adaptation due to the diversity of imaging modalities and variations in imaging conditions. This is where the concept of Test Time Adaptation (TTA) becomes crucial. TTA enables a model to access only online test samples and adapt to unknown distributions, specifically addressing the feature alignment challenge between the source and target domains within medical imaging contexts. Our approach involves a dual-path feature alignment structure to manage both domain-specific and domain-shared knowledge. The teacher model in our framework is responsible for adapting to the current domain distribution and applying various knowledge allocation strategies. Concurrently, the student model focuses on discovering shared representations between the source and target domains, maintaining feature consistency across different batches. A key aspect of our method is the use of a consistency loss to align the predictions of the teacher and student models, ensuring cohesive and reliable outputs. The final output of our system is a synergistic combination of insights from both the teacher and student models. We have rigorously tested our approach on medical image segmentation tasks. The results underscore the effectiveness of our method, demonstrating competitive performance in this challenging and critical domain.

In trained few-shot segmentation, an inevitable bias towards the base classes may occur. By leveraging the spatial consistency of organ distribution in medical image segmentation, spatial priors can be utilized to roughly predict the location of new categories. This helps to avoid bias towards the base classes. This paper introduces a novel prototype network that integrates prior spatial knowledge and grid attention to enhance the accuracy of few-shot medical image segmentation by predicting target locations and offering guidance for segmentation algorithms. We introduce dedicated spatial prior module to extract information about the location of targets. This module utilizes the spatial and shape consistency of supported and query targets in medical images to better guide the segmentation process. By incorporating a spatial prior module, the segmentation network is guided to follow the desired spatial structure of the target. Additionally, to emphasize essential local semantic information in medical images, we have devised a Grid Attention (GA) module. The GA module adaptively adjusts the importance of different positions in the feature map, allowing more attention to be placed on regions relevant to the segmentation task. By combining spatial priors with grid attention, our method can better capture target information in medical images and achieve more accurate segmentation results. This claim was substantiated through our experiments focusing on segmenting abdominal organs in MRI scans. Our method was objectively compared against several established models, demonstrating that the inclusion of spatial priors and grid attention greatly improved the segmentation results.

The problem of complex backgrounds, dramatic changes in the scale of objects and the presence of a large number of dense small targets in the UAV aerial images. In this study, a detection algorithm for small targets is proposed based on the YOLOv5s algorithm. Firstly, a new BottleNeck: C3LK module is designed for the C3 module using depthwise separable convolution to construct a large convolutional kernel on the YOLOv5 backbone network, which improves the ability to capture global information and compensates for the lack of local information for detecting small targets. Secondly, a small target detection layer and two feature fusion layers are added. At the same time, In order to better detect objects of different scales in complex backgrounds, a module for multi-scale feature extraction is designed by combining dilated convolution and attention mechanism. The experimental results show that the proposed algorithm improved 5.4% over YOLOv5s in terms of mAP@0.5 and over YOLOv7-tiny. The method meet the requirement of real-time detection in aerial image small target detection, and improve the detection performance in small target scenes.

Trusted computing is an effective means of system security protection, encompassing various security solutions. However, these solutions exhibit individual shortcomings, such as the co-location of protective code and system, a singular metric determination standard, and discontinuity in trusted measurements. Consequently, they fail to meet security requirements. Hence, we propose a Behavior-Based Dynamic Trusted Measurement scheme (BBDTM). Firstly, the scheme designs a dynamic measurement framework based on a dual-body architecture, segregating protective code from the operating system code to overcome the reliance of protective code on the host system. Subsequently, we assign corresponding weights to behaviors based on their impact on the system using information entropy theory, rendering the metric determination standards more rational. Next, a time window mechanism is designed to achieve a comprehensive trusted measurement of system behavior, addressing issues of trust decay and discontinuity in metric behavior. Finally, relevant experiments demonstrate that our scheme can achieve superior protective results.

For novice programmers, debugging is a crucial skill that can be challenging to master. In debugging education, debugging behaviors have attracted the interest of educators and researchers. They attempt to understand the objectives and logic of debuggers by collecting corresponding debugging data. Debugging targets inferred from this data can help educators understand students' debugging strategies and assess their skill development more efficiently. This paper designed a debugging task by inserting a series of bugs into a complete project, and collected fine-grained debugging data from novice programmers on this task. Thus, the problem of predicting debugging targets is transformed into a multi-label classification issue. Furthermore, due to the difficulty of annotating a large amount of debugging data, we propose a Semi-Supervised learning-based Debugging Target Prediction method (SSDTP). This method employs a self-training approach, fine-tuning on the labeled dataset to capture the textual and structural features of code differences before and after each debugging. It was applied to 14,080 debugging data instances, and experiments validated its effectiveness in debugging education. When the labeled dataset used accounted for 10% of the total data, the classification accuracy was 78%.

In China, coal-fired power plants, which mainly rely on coal, require timely inventory of coal storage to calculate economic benefits. How to achieve accurate measurement of coal storage has always been a difficult problem that coal-fired power plants urgently need to solve. For precise measurement of coal storage capacity of unmanned aerial vehicles (UAVs) in closed coal yards, the three-dimensional model through the three-dimensional (3D) reconstruction of oblique images is a highly valuable research path for engineering applications. This paper studies on the 3D reconstruction technology of oblique images based on UAVs vision. In terms of image enhancement, an enhancement function with offset is introduced to highlight image edges while preserving background information. For image feature extraction and matching, speeded up robust feature (SURF) image recognition and square filtering integral graph acceleration are adopted to meet real-time and robustness requirements. When extending dense point clouds with rich details in sparse point clouds, the area adaptive algorithm is used to meet the requirements of regional smoothing and surface changes. In the Poisson algorithm for 3D mesh reconstruction, Gaussian filtering is used instead of 3rd-order mean filtering to maintain the surface topology of the point cloud model. A 3D reconstruction of a closed coal yard in a coal-fired power plant was carried out using UAVs, and the results showed that the accuracy of the algorithm was improved by 4.5%, and the integrity by 2.22%. This technical solution meets the requirements of UAVs oblique image 3D reconstruction in closed coal yards, achieving accurate coal quantity statistics, timely inventory and transportation, stable boiler combustion, and successfully improving the accuracy and completeness of visual based 3D reconstruction.

The Wordle game offered daily by the New York Times has taken the United States by storm and is loved by many people and challenged by major internet users. In order to improve the user gaming experience, it is therefore necessary to analyse the game user data. This paper uses a time series model, a random forest regression model, and a random forest classification model, and constructs eight attribute features for the words, predicts the difficulty of guessing EERIE words as being difficult based on the attribute features, and visualises other features of the data. Firstly, the ARIMA(1,1,0) model was used and predicted an approximate prediction interval of [18578, 22562] for the outcome data reported on 1 March 2023. Next, we used a random forest regression model as well as a random forest classification model and predicted the percentage of players who could guess the word EERIE for the first time on March 1, 2023, respectively. The words were categorised by difficulty: easy, average and hard. We finally predicted that the word EERIE belongs to the difficult class. Finally, the following features were created based on the data obtained. The first feature: a time-series plot of the number of reported results; the second feature: a plot of the mean value of the number of attempts from 1 to 7, and the probability of success of the number of attempts by the player is considered to obey a normal distribution. The research in this paper can improve the Wordle game to a certain extent, enhance the players' gaming experience and attract more people to participate in the Wordle game.

Among various non-destructive testing techniques, ultrasonic non-destructive testing has gained widespread application domestically and internationally due to its advantages such as high portability, good quantitative and qualitative accuracy in defect detection, and ease of testing complex structures. Ultrasonic phased array detection, due to its characteristic of excitation delay, is capable of adapting to more complex workpiece inspections. In the field of ultrasonic phased array, total focus method imaging has the highest imaging accuracy and image quality. As an imaging method, the total focus method algorithm requires a large amount of storage space and computation time in data acquisition and image computation. It is challenging to be used in real-time detection and is generally applied to offline analysis. In order to narrow the gap between total focus method imaging in China and foreign countries and meet the demand for high-precision real-time non-destructive testing, this paper conducts research on a real-time imaging algorithm based on OpenCL hardware acceleration for ultrasonic total focus method. The analysis indicates that the method proposed in this paper provides a reference for heterogeneous acceleration of total focus method imaging. It also opens the possibility for its establishment in the field of industrial non-destructive testing.

Phased array technology primarily relies on precise control of the excitation time and sequence of each element in the ultrasonic transducer, generating beams with different deflection angles and focal points. The spatial superposition of these beams on the composite wavefront can achieve specific phased effects, enabling comprehensive scanning of the tested object. In ultrasonic phased array sector scanning, the sampling space of raw data is established in polar coordinates, while imaging display is set in Cartesian coordinates, requiring coordinate transformation. However, transformed images often exhibit significant imaging distortions and lower clarity. This paper proposes an algorithm based on polar coordinates, utilizing bilinear interpolation in Cartesian coordinates for two dimensions. This approach effectively enhances the accuracy and smoothness of sector scanning imaging. To improve the real-time frame rate of sector scanning imaging, a method employing auxiliary matrix querying is introduced. Experimental results demonstrate a 3.8-fold increase in the average frame rate of imaging.

Aiming at the problems of false detection and missing detection caused by dense vehicle targets, mutual occlusion, and too small targets in automatic driving scenarios, an improved YOLOv7 target detection algorithm is proposed. In the detection task, the SENet module is introduced to learn the feature weights according to the loss function, which can make the effective feature maps have more weight, and the invalid or ineffective feature maps have less weight, so that the training model can achieve better results. The normalization-based NAM attention module is introduced, which reduces the weights of less significant features, making them computationally more efficient while maintaining the same performance. To verify the validity of the experimental results, the experimental results of the improved YOLOv7 were compared with the experimental results of YOLOv5 and the original YOLOv7 model.

Due to complex landscape features and illumination variation in actual complex environments, the accuracy of existing object detection methods is unstable and limited, especially under a few-shot condition, which seriously affects their application in practical complex environments. Therefore, an object detection model based on the Attentive Single Shot Multibox Detector (ASSD) network is proposed combined with corner feature fusion and Soft-non-maximum suppression (Soft-NMS) in this study. First, the local and global information, extracted from the image convolutional layer features, and the structural and detailed information, provided by the corner features, are fused together to enrich the model’s understanding ability for spatial information, as well as the diverse feature representation capability. Second, the Soft- NMS is introduced into the traditional NMS model to reduce the false negative rate under the condition of object occlusion. Finally, the PASCAL VOC2007 and COCO datasets are used to validate the proposed model, and the experimental results show that the mAP is 83.35% on PASCAL VOC2007, which is 1.29% higher than the baseline ASSD model, and the mAP improvements of 0.57%, 1.80%, 4.53%, 9.29%, and 24.14% within the NMS threshold of 0.5-0.9, respectively.

The mangrove wetland ecosystem holds immense value for the entire coastal zone. While, it has suffered significant damage because of various factors. Accurate detection and identification of species in mangrove growth areas are crucial for studying the dynamic changes occurring in this ecosystem. In this research, we propose CASWIN Unet, a remote sensing image semantic segmentation model based on Swin transformer and CBAM module with a U-shaped structure, to address the challenge of obtaining comprehensive global information in existing deep learning-based high-resolution mangrove image semantic segmentation. This challenge often leads to blurred prediction edges sparsity issues and incomplete extraction of flooded areas. To enhance performance, we introduce a remote sensing prior knowledge VDVI layer into the encoder an utilize an unmanned aerial vehicle dataset capturing mangroves in Xiamen area. The experimental results demonstrate that CASWIN Unet significantly improves the extraction accuracy of sparse and flooded areas within mangroves compared to both the original model (with an overall accuracy boost of 1.22%) and FCN (Fully Convolutional Network) (with an overall accuracy improvement of 5.8%). All in all, the CASWIN Unet outperforms other comparative models.

How to solve the problem of excessive execution time of current image processing algorithms? In this research, the CUDA parallel architecture is used as the basis to optimize and accelerate the image processing algorithm. By rationally designing parallel algorithms, optimizing memory access and parallel task allocation, and using shared memory and parallel computing capabilities, the efficiency of image processing algorithms can be maximized. The experimental results show that the execution time of the method in this paper requires a maximum of 6s and a minimum of 1s. The optimization and acceleration of the image processing algorithm based on the CUDA parallel architecture have achieved remarkable results. Compared with traditional serial algorithms, the speed of image processing is greatly improved through parallel computing and the optimization of shared memory. At the same time, through reasonable thread block and grid design, the memory access mode and parallel task allocation are optimized, and the performance and efficiency of the algorithm are further improved.

Target detection technology based on optical remote sensing images has been widely used in military and civil fields, but there is still a problem of low detection rate due to blurred targets that are too small. An improved algorithm based on YOLOv7 is proposed. First, the Efficient Decoupled Head with privacy learning mechanism is improved to improve the detection accuracy. Secondly, the Normalized Wasserstein Distance loss function is used to solve the problem of targets that are too small and difficult to detect. Finally, the PConv convolution in FasterNet is used to redesign the four ELEN modules in the BackBone structure. Experimental results show that the mAP (0.5) of the improved algorithm increased from 92.8% before improvement to 98.2%, an increase of 5.4%, and can be effectively used for the ship detection in optical remote sensing images.

Aiming at the problems of high damage rate of cultural relics, low processing efficiency and fuzzy detail processing when producing handmade rubbing, a method of generating digital rubbing of Han Dynasty Stone Relief based on image processing is proposed. The method first uses grayscale and mean value filter to eliminate the difference in shadows generated by image acquisition, and then generates adaptive thresholds for local pixel blocks for binarization to remove the noise caused by relief erosion and obtain the preprocessed image. In order to highlight the differences in texture details of the Han Dynasty Stone Relief, the foreground template with blurred edges is generated using contrast stretching and Ostu’s method to obtain an enhanced image. Finally, the preprocessed image and the enhanced image are image fused to obtain a relief digital rubbing with noise elimination and enhanced details. Testing the effect of digital rubbing generation of Han Dynasty Stone Relief by MATLAB program. The experimental results show that, compared with the manual rubbing, the similarity reaches 80.26%, and the average processing time is 0.44s, which solves the problems of low conversion rate and fuzzy texture details of the traditional scheme successfully, and it plays a vital role in the protection and inheritance of cultural relics.

With the development of unmanned aerial vehicle (UAV) technology, object detection technology has been widely used in UAVs, realizing object detection and recognition of ground scenes from the perspective of aerial photography. To further enhance the performance of object detection from an aerial photography perspective, an improved YOLOv7 network is introduced in this paper, termed as YOLOv7-Vis. Firstly, data augmentation methods are used to improve the imbalance of samples in the UAV data set. Secondly, an additional prediction head is added to detect small-scale targets. Thirdly, the SimAM attention module is inserted into YOLOv7 by discussing the performance of different attention combinations. Finally, the experimental results show that the method proposed in this paper improves the mAP@0.5 by 5.9% and the mAP@0.5:0.95 by 3.6% over the original YOLOv7 on the VisDrone dataset.

Many-objective optimization problems are challenging for the majority of extant multi-objective evolutionary algorithms to solve because they are unable to strike a balance between diversity and convergence in the high-dimensional objective space. Therefore, for many-objective evolutionary optimization, this paper proposes a vector angle-based and I_SDE⁺ indicator selection strategies. Using these two strategies, the worst individual are repeatedly removed from the population through environmental selection. More specially, the first one is aimed at finding any two individuals that have the smallest vector angle, which indicates that they are looking in the same research direction. The latter considers the individuals with poor convergence and diversity, which are eliminated according to the value of I_SDE⁺ indicator one by one. Experimental results revel that the proposed algorithm obtains competitive advantages on various benchmark problems with up to 10 objectives, when the proposed algorithm is compared with several state-of-the-art many-objective algorithms.

This paper investigates the three-dimensional reconstruction technique of unmanned aerial vehicle (UAV) images under multiscale scenarios based on neural radiance fields. The approach combines the concept of residual networks and the convergence method of unbounded region functions, and its feasibility is demonstrated through theoretical analysis. Experimental studies are conducted on UAV image datasets at various scales, with a particular focus on unbounded regions with diverse camera poses. The proposed model achieves satisfactory results in handling the challenges posed by multiscale UAV image data.

Infrared imaging technology has unique advantages in the field of aerial motion target detection, but usually the information captured by the target is limited. Because the small target is easy to be submerged in the background, the existing method is difficult to detect the small target quickly and accurately under various complex environments. This paper proposes a fast detection method by using the morphological features of targets in infrared images. This method increases the morphological differences between target and non-target regions in the image through high pass filtering, and then uses the SUSAN edge detection operator to scan the processed image and detect the target interest region. In MATLAB simulation environment, the effectiveness of the algorithm was verified through infrared image sequences captured in specific actual sky background. Through experimental data obtained under specific conditions, it has been proven that this method can detect targets and overcome the influence of complex interference when the sample data is limited.

This study proposes a blind super-resolution (SR) method with degradation estimation, combining the DASR (Degradation-Aware SR) and CLIP models. Firstly, we employ unsupervised degradation representation learning to acquire abstract representations that can distinguish various types of degradation in the representation space, rather than explicitly estimating in the pixel space. Next, the DASR network is introduced, which utilizes the learned representations to adapt flexibly to different degradation scenarios, thereby improving the performance of SR. Furthermore, to further enhance the quality and accuracy of SR, the CLIP model is integrated into DASR. The CLIP model is primarily designed for pretraining tasks between images and text, but it can be utilized to compare and match images by utilizing image embedding vectors in the CLIP model, enabling a better understanding of image content. By integrating the CLIP model, the SR reconstruction process can be better guided, ensuring visually more realistic generated images. Experimental results demonstrate that our method accurately recovers high-resolution details and generates clearer, more realistic images, both in synthetic and real images.

Vector graphics are very common in the field of graphic design, and their characteristics allow them to be fully utilized in scenarios that require flexibility and clarity. In order to meet the need for vectorization of such grayscale graphs, this paper proposes a method based on contour tracing algorithm and describes it. Firstly, the pixels in different intervals of the image are binarized, then the edge contours are traversed to get multiple closed polygons, the corners that meet specific requirements are smoothed, and then the images in each interval are accumulated and merged to get the final vectorized graph. Finally, the object-oriented operation software is designed under the MFC framework by using VS, and at the same time, the improved method is used to vectorize the image with more details and textures to get the vector graph and compare it with the original graph, and the accuracy of the vectorized graph and the original graph is obtained by calculating the RMSE as a measurement index, which verifies the accuracy of the scheme. The article concludes with an analysis of the SVG file format, and the output of the vectorization results of this paper is saved to a file in SVG format.

In this paper, a multistage rain removal optimization algorithm based on residual networks is proposed. The relationship between pixels and the intensity of corresponding background pixels is deduced from the physical process of the image, and the intrinsic features of rainy day images are utilized to form an architecture using a multilevel residual ring to construct a network model for function training, so as to achieve the purpose of rain removal. Through the deep extraction of feature information, combined with the combination of densely connected networks, the rain morphology in the image is captured separately through joint training, so as to obtain rain removal images with different visual effects. The algorithm proposed in this paper is experimentally compared with existing rain removal algorithms, and the subjective and objective evaluation indexes prove that the method proposed in this paper is superior to the existing rain removal methods, which proves that the method proposed in this paper improves the performance of rain removal and the practicality of the existing rain removal task.

Slurm (Simple Linux Utility for Resource Management) is a popular open-source cluster job scheduling system. In scenarios involving multiple queues and a large number of scientific computing tasks, the Slurm scheduling system faces challenges such as improper queue depth settings and uneven job loads in queues. This research aims to optimize the scheduling performance of the Slurm scheduling system in scenarios where resources are shared among multiple queues. By forecasting future CPU load increases based on the historical CPU utilization of queues, a window period is identified. Using the forecasted results, a dynamic adjustment method for queue priorities is introduced. This method aims to elevate job priorities in specific queues when a significant number of jobs are queued, ensuring prompt execution. Additionally, this study involves a dynamic adjustment strategy for queue depth to address issues arising from inappropriate queue depth settings, where resources in the queue are idle, but queued jobs face delays in scheduling. Experimental results demonstrate that this approach enhances the scheduling performance of the Slurm system in multi-queue scenarios, improving cluster resource utilization and better meeting the diverse demands of users.

To obtain fast and accurate image recovery result, an image recovery algorithm based on the improved smoothed L0 norm is proposed in the article. Firstly, the negative exponential function called smoothed function can be used to approach L0 norm, the optimal value of the smoothed function is solved by constructing the sequence of control parameters. Secondly, the gradient projection algorithm can be used to solve the optimal value which recovery the sparse signal. Finally, a comparison correction step is added which can search according to the steepest descent direction to ensure the precise recovery of sparse signal can be obtained. Experimental results express the performance of the proposed algorithm is superior to traditional methods.

Traffic signs, as critical assets on highways, play a crucial role in road safety, traffic flow optimization, and intelligent transportation systems. In urban road environments, the influence of factors such as urban greenery and buildings leads to the occlusion of traffic signs and the complexity of the background environment, making the detection process more intricate and challenging. Against the backdrop of the diversity and variability of urban environments, this paper proposes a method for the rapid collection of traffic signs that integrates binocular vision and computer vision. Firstly, data collection and preprocessing are accomplished using a binocular camera mounted on a mobile measurement collection vehicle. Subsequently, our proposed fast detection and localization method are applied to detect and track objects in the left image, followed by matching detections in the right image. The detection results from both images are then matched, and the disparity is calculated. Finally, combining the results with the calibration information from the binocular camera and feedback from a programmable controller, the position information of traffic signs is obtained. Experimental validation demonstrates that our proposed fast detection and localization method achieves a detection accuracy and recall rate of 95% and 97%, respectively, showing robustness against occlusion and complex backgrounds.

Aerial-ground unmanned systems are the mainstream of research in recent years, target localization is a key step in rescue or strike missions, and the use of aerial-ground unmanned systems for target localization has great research value. However, aerial-ground systems have large differences in maneuverability and large sensor parallax. In this paper, an aerial-ground cooperative target localization method based on trajectory matching is proposed. First, the algorithm obtains the traveling trajectory of the unmanned vehicle in the UAV pixels. Then, the pixel trajectory and the UAV local coordinate trajectory are utilized to solve the transformation matrix from the pixel coordinate system to the geographic coordinate system. Finally, the geographic coordinates of other targets using the transformation matrix. The algorithm uses very little sensor data and a small number of platforms, and the experimental results show that the algorithm has an average localization accuracy of 1.55m at a UAV flight altitude of 100m, which is better than other target localization algorithms.

To improve the quality of alloys, experts need to intelligently analyze the metallographic images, where semantic segmentation serves as a vital means of achieving this analysis. However, traditional fully supervised semantic segmentation methods often suffer from high annotation costs. Weakly supervised semantic segmentation methods require only simple image-level annotations to achieve high-precision semantic segmentation. This paper constructs a dataset of weakly supervised microstructure images of aluminum alloys, greatly reducing manpower and time costs. Furthermore, addressing the sparsity and incompleteness of Class Activation Maps (CAMs) in traditional weakly supervised methods, a weakly supervised semantic segmentation method for aluminum alloy microstructure images is proposed based on consistency constraints. The method first introduces consistency constraints in the feature extraction process to improve the quality of CAMs. It then uses the improved CAMs as a target cue to generate affinity labels and employs the AffinityNet network and random walk strategy to mine target pixels in non-discriminatory areas, obtaining completer and more detailed pixel-level pseudo labels. Finally, semantic segmentation of the obtained pixel-level pseudo labels is performed using the U-Net network to obtain the final segmentation result. Extensive experimental results demonstrate that our method effectively improves the quality of CAMs, thereby enhancing the performance of the segmentation network.

Breast cancer is recognized as one of the most prevalent cancers among women worldwide, and early detection plays a crucial role in improving patient survival rates. Computer-aided diagnostic systems (CAD) can provide additional information to assist doctors in diagnosis. However, there is a scarcity of deep learning models for breast cancer diagnosis based on CEUS videos, and obtaining datasets is challenging. Given the widespread use of CEUS due to its non-invasive and cost-effective nature, there is an urgent need to develop a diagnostic support system specifically for breast cancer CEUS images to aid doctors in diagnosis. In this study, we constructed a dual-mode dataset consisting of contrast-enhanced ultrasound and B-mode ultrasound. We proposed a Bimodal Feature Fusion Three-Branch Network, which achieves crossmodal information fusion through feature fusion and decision fusion collaborative reasoning. We validated our model using a breast cancer video dataset comprising 332 cases. Ultimately, our model achieved an accuracy of 88.24% while maintaining high sensitivity and specificity. These results indicate a reduction in breast cancer misdiagnosis and missed diagnosis rates compared to existing models.

Magnetic resonance imaging (MRI) has high resolution for soft tissues and significant advantages in the examination of the bone and joint systems and central nervous system. However, due to cost and usage limitations, MRI cannot be applied in various scenarios. Computer tomography (CT), as a fast medical examination technique that does not require special preparation, has a wider range of applications, but its accuracy in organ imaging is lower than MRI. At present, image conversion methods suffer from low resolution and low signal-to-noise ratio in synthesizing MR images. In addition, these methods have limited capabilities and cannot accurately synthesize MR images with diverse anatomical and functional differences. Therefore, this article proposes a multimodal medical image conversion method based on Generative Adversarial Networks (GAN) to convert CT images into MR images. This method uses two parallel generative adversarial networks to generate two MR images to compensate for the missing generation of a single generative adversarial network. The experimental results demonstrate that this method is effective and can generate MR images close to the reference MR image.

Computed tomography (CT) is a commonly used medical imaging modality for diagnostic purposes. However, the high radiation dose associated with CT scans can pose risks to the human body. Furthermore, reducing the radiation dose can significantly impact the quality of CT images, leading to a decline in image quality. To enhance the denoising performance of Low-Dose CT (LDCT) and improve the quality of CT images, a novel multi-head self-attention mechanism and multiscale feature fusion network (MAMSNet) is proposed on basis of encoder-decoder network architecture. In new MAMSNet, there are threefold modifications. Firstly, a multi-head self-attention(MHSA) block is introduced to improve the global extraction feature and achieve better denoising effect. Secondly, to enhance the receptive field and obtain information at multiple scales, a block called multi-scale feature fusion (MSFF) is utilized. Moreover, a Structure-aware loss function is added to maintain the texture structure information and elevate the image fidelity of CT scans. The improved method can be effectively used to remove the noise and artifacts of LDCT images. Compared with other methods, more texture features, structural details and higher objective metrics (such as PSNR and SSIM) can be preserved.

Face image translation is the transformation of an input face-labeled image from the source domain to the target domain. However, most of the current methods are for low-resolution image translation, while the high-resolution models are less effective, and problems such as loss of detail information and blurring distortion occur. Thus, a mask-guided highresolution face image translation method is proposed. First, for the problem of irrelevant background information interfering with salient region features in face images, generator introduces parameter-free attention module to obtain global information. Second, to enhance the network's ability to extract fine-grained semantic features for high-resolution faces, the backpropagation of gradients is facilitated by utilizing the improved ConvNeXt block. Finally, discriminator is upgraded to multiscale discriminator that guides generator to produce globally consistent images and finer details. Experimental results show that the proposed method outperforms mainstream models in both similarity and detail generation of image translation results on the CelebAMask-HQ dataset with a resolution of 512×512.

In order to enhance the level of power grid services, understand the individual needs of different user groups, and optimize energy planning, a method for constructing and applying power user portraits based on multi-feature integration is proposed in this paper.By combining user social attributes, daily electricity consumption attributes, and quarterly electricity consumption attributes, features are extracted from multiple dimensions to construct a multi-dimensional user behavior label database.The k-means clustering algorithm is employed to construct power user portraits, and the CNN-LSTM model is applied to predict the total energy consumption of user groups under different feature label selection rules. This is done to assess the application potential of the proposed power user portrait construction method.The experimental results indicate that the proposed method provides accurate characterization of user group features, significantly improves energy consumption prediction accuracy, and effectively enhances the level of power grid services, assisting in energy dispatch decision-making.

To address the challenge of poor image quality in industrial surveillance videos under low-light conditions, hindering the normal execution of target recognition and detection tasks, this paper proposes a low-light surveillance image enhancement model based on Multi-Scale Retinex (MSR). Firstly, the MSR algorithm is employed to enhance the brightness of the images. Due to the loss of fine details in the brightness-enhanced part, the Laplacian sharpening algorithm is subsequently applied to enhance the details, allowing the brightness-enhanced images to retain more detailed features. Experimental results indicate that compared to the MSR algorithm and histogram equalization algorithm, the proposed algorithm in this paper performs better in terms of signal-to-noise ratio, grayscale values, and standard deviation. It effectively improves image quality while achieving a balance between brightness, details, and real-time performance.

In order to infer the behavioral intention of airborne multifunctional radar, a working mode transformation model was studied, and an intention inference algorithm was designed based on LSTM. A pattern transformation simulation dataset was constructed, and the intention inference practice of airborne multifunctional radar was carried out. The research results indicate that time length and batch processing rate have a significant impact on the final prediction results. When the timing length is set to 60 and the batch processing rate is set to 20, the prediction accuracy can mostly be maintained above 85%, which lays an important foundation for determining combat actions and countermeasures in actual combat.

To address the lack of effective grading and processing methods in the current large-scale development of lemon cultivation in China, this study introduces a Class-guide feature extraction module on top of the existing YOLOv5s model. This further enhances the network's feature extraction capabilities and integrates the correlation between lemon detection and grading tasks. The study designs a multi-task learning algorithm that combines image-level and box-level fusion and incorporates defect detection and quality grading into a single task model through joint training. Experimental results show that the improved model achieves a 5% increase in mAP and mAP50-95 values compared to the original YOLOv5s model in detection, and an accuracy improvement of over 2.2% in classification compared to commonly used algorithms. The lemon quality grading detection method based on improved YOLOv5s achieves improved performance in both detection and grading tasks, meeting the practical application needs of the market.

To address issues such as low accuracy in vehicle and pedestrian detection and slow inference speed, an improved version of the FCOS object detection algorithm is proposed. The algorithm replaces the original backbone with ResNet18 and incorporates deformable convolution to enhance its ability to capture shape features of the targets, thereby improving feature extraction and model inference speed. A coarse localization box mechanism is added to the algorithm's localization detection head branch to improve the model's classification accuracy and detection box recall. A new background prediction branch is introduced to enhance the model's detection capability for positive samples by predicting the probability of differences with the environment. Additionally, an ATSS (Adaptive Training Sample Selection) label assignment scheme and GFocal Loss function are employed to replace the original label assignment scheme and loss function, mitigating the impact of imbalances between positive and negative samples in the dataset. The experimental results show that the improved algorithm achieves 94.7% on the KITTI traffic target data set, which is 22.2% higher than the original FCOS model, improves the inference speed to 74 pieces per second, and has better effect compared with some mainstream target detection algorithms.

In view of the shortcomings of small object feature information, low detection rate, and high error and missed detection rates in object detection tasks. This paper proposes an improved Single Shot Multibox Detector (SSD) algorithm—SSDRes50- AF for small object detection. First, the ResNet50 residual network is used as the backbone network of the SSD algorithm to enhance the feature extraction capability of the SSD algorithm. Secondly, a lightweight and efficient feature fusion module is added before the prediction feature layer to improve the multi-scale feature fusion capability of the SSD algorithm. Localize small objects more accurately. Finally, an attention mechanism is added and applied to the medium-scale feature map of the network to enhance the effective information in the feature map. Experiments on the PASCAL VOC 2007 data set show that this algorithm effectively improves the accuracy of small object recognition, improves the situation of missed detection and greatly improves the detection accuracy and robustness of the algorithm.

The field of multi-modal medical image fusion is gaining prominence in healthcare systems. How to obtain clear structure, rich details and complete contextual fusion information in the fusion results is the key point of research. In the current works, too much attention has been paid to the extraction of local detail information while the essential overall information has been neglected. To extract local and global context-related information simultaneously, a multi-scale network based on coordinate attention and Swin Transformer, called CASTNet, is proposed in this article, which combines the advantages of convolutional neural network (CNN) and Swin Transformer. Specifically, in the CASTNet framework, a multi-scale coordinate attention embedding module is designed, to extract more detailed local texture features and long-range dependencies. Next, Swin Transformer is introduced to further capture overall contextual information. Finally, in feature reconstruction, dynamic convolution decomposition instead of normal convolution is used to enhance the reconstruction performance. The test results on a mainstream database demonstrate that the results obtained by the proposed method are superior to those of the comparison methods (such as multi-scale adaptive transformer), both in terms of subjective visualization and objective quantitative evaluation.

Texture is a key and basic visual cue for various image processing applications. To capture rich and discriminative local texture information, this paper develops a completed region contrast binary pattern (CRCBP), which introduces a novel pixel-region contrast measure scheme to extract contrast difference information between the pixel and its local background region. The proposed CRCBP consists of three components: local difference sign, local difference magnitude, and local pixel-region contrast measure sign. According to the local neighborhood encoding mechanism, we can get three sub-patterns: sign binary pattern, magnitude binary pattern, and local region contrast binary pattern. To achieve completed texture feature representation, these three sub-patterns are jointly combined to obtain the completed region contrast binary pattern. The comparative evaluations on benchmark database shows that the proposed CRCBP descriptor yields the state-of-the-art classification performance.

The millimeter-wave radiometer has the ability to work all day and all-weather conditions, it also has good concealment. In order to solve the problem of detection and identification of sea surface target, this paper proposes to use microwave radiometers to accurately detect short-range sea surface target, especially metal target. Based on the radiation characteristics of target in the sea surface background and the millimeter-wave radiation measurement principle, the observation model of millimeter-wave radiometer at any height in the sea surface background is constructed. Firstly, the radiometer antenna temperature of the metal target under different parameters is simulated, and then the effective situation of the 3mm band and 8mm band radiometer for the detection of sea surface target is analyzed, so as to provide some theoretical support for the passive detection of sea surface target.

Passive detection has a wide range of applications in the positioning, angle measurement or frequency measurement of targets with interference or radiation sources. Previous research on passive detection has mainly focused on ground targets, so this paper proposes to detect aerial metal targets with a 3mm band radiometer. Based on the radiometer detection principle and the radiation characteristics of aerial targets, the radiation temperature of the sky background and the radiation temperature of metal targets are analyzed, and the antenna temperature simulation is carried out for metal targets. The simulation results show that the detection results of the 3mm-band radiometer on the aerial metal target are relatively satisfactory, thus providing certain theoretical support for the millimeter wave passive detection of aerial targets.

Target detection, a crucial component of computer vision, has undergone rapid development in recent years. While existing target detection networks have demonstrated significant improvements in accuracy and speed compared to traditional counterparts, they often come with a drawback: a large number of parameters and high hardware requirements, making them less practical for use on mobile devices. The design of lightweight networks aims to address this challenge by further reducing the number and complexity of model parameters while preserving accuracy. To tackle this issue, we propose a channel pruning method based on the comprehensive evaluation technique of Topsis to trim down the model parameters. During the model fine-tuning stage, we employ a dichotomous pruning fine-tuning strategy to minimize the effort required to restore the model's accuracy.

Image denoising has always been a hot research topic in the field of computer vision, playing an important role in improving image quality and accuracy. In this paper, a method for image denoising based on residual network is proposed, which uses deep learning technology to improve the effect of image processing. This method can automatically learn useful features in the image and effectively remove noise in the image, achieving automatic removal of image noise. Finally, experiments were conducted on standard image datasets for evaluation. The experimental results show that the image denoising method based on residual network has better performance compared to traditional methods. It can effectively remove various types of noise, improve image quality and clarity, and has broad application prospects in image processing.

Image style transfer has been widely concerned in the field of computer vision, and has important application value in cultural heritage, social media and public security case detection. At present, the mainstream style transfer algorithm is based on deep learning, which is fast and can achieve real-time effect. However, the generated stylized image has some problems such as texture deformation and edge blurring. In this paper, the arbitrary style transfer algorithm is studied. The arbitrary style transfer method is not limited by style types, and it does not need to retrain the model when transferring new styles. However, the stylized images generated by the existing algorithm are prone to fuzzy texture and inconsistent color distribution with semantic information. To solve these problems, a framework for semantic perceptual style transfer based on identity loss is proposed. Firstly, the adaptive attention normalization model is adopted to comprehensively consider shallow features and deep features to carry out stylistic attention weighting. Secondly, an image is input into the network as both style image and content image to measure the loss. At the same time, the semantic local mapping and global statistics between content features and style features are considered. Finally, by learning the same features between content and style, richer and more accurate semantic and style expressions can be obtained. The experimental results show that the texture features of the stylized image obtained by this method are clearer, the color is more vivid, and the effect is better.

Multiple robots are more effective in mapping unknown environments than a single robot. Time reduction resulting from parallelization is crucial for surveying complex areas. Multi-robot mapping, there are commonly two solutions. In the first solution, robots exchange raw data from sensors. The second method assumes that each robot independently creates a local map and exchanges and integrates it with other robots. In this paper, we propose a 3D map merge framework and a fine registration method using purely geometric maximal cliques that enhances efficiency through the utilization of overlapping regions in feature-based alignment processes. This algorithm does not require any initial guessing about the conversion between local maps. However, for successful integration, the map needs to have a common area. We demonstrate the effectiveness of the implemented methods in various environments.

The storage conditions for paper cultural relics entail strict requisites. Any alterations to the storage conditions can lead to the proliferation of diseases, with mould being particularly prevalent and causing discoloured splotches on the surface of the relics. The resultant physical and biological damage is severe. In this study, we examine the differences in spectral characteristics between moldy and healthy areas in paper cultural relics samples. We used hyperspectral imaging technology, known for its non-destructive and fast features, to set up a discrimination model for phytophotocyanobacteria based on KNN (K-nearest neighbor), SVM(Support Vector Machine), 2D-CNN and VVG16, and evaluated the performance of the model by comparing the mold identification rate between different models to screen the optimal modelling scheme. The primary contributions of this paper are the amalgamation of hyperspectral imaging technology with mold spot detection of cultural relics on paper.

To address the current issues of slow detection speed and poor accuracy for small targets in traffic sign detection, an optimized model based on YOLOv8 is proposed. This model integrates a small target detection layer, merging deep and shallow semantics to reduce semantic loss due to scale inconsistency. Additionally, it employs the lightweight EfficientNetV2 as the backbone network to decrease model parameters and enhance performance. The Coordinate Attention Mechanism is also incorporated, improving feature localization. Testing on the TT100K dataset, this model surpasses YOLOv8, with an accuracy increase of 7.1%, mAP@0.5 increased by 5.8%, a weight file of 5.92MB, and a detection rate of 279FPS. The model shows improved accuracy and performance while maintaining a simpler structure.

The utilization of unmanned aerial vehicles (UAVs) serves as a pivotal role in the inspection of transmission lines, specifically in the detection of anomalies. This study aims to enhance the efficiency of anomaly detection by proposing an automated method using UAV inspection to identify flaws in components. Through the utilization of an end-to-end coordinate attention mechanism and a bidirectional feature pyramid network (BC-YOLO), defects in both dampers and insulators can be discerned. To overcome the challenge of background interference, researchers have integrated a coordinate attention (CA) module and implemented a bidirectional feature pyramid network (BiFPN) for effective multi-scale feature fusion. The experimental results obtained from the dataset authenticate the effectiveness of the suggested approach.

Segmentation plays a crucial role in medical image analysis, enabling the identification and localization of diseases, monitoring morphological changes, and extracting discriminative features for further diagnosis. Skin cancer, a widespread global malignancy, emphasizes the importance of early diagnosis in completely eliminating malignant tumors from the body. We propose a skin lesion segmentation network based on a hybrid reverse attention mechanism to address the issues of significant individual differences, blurred edges, and inconsistent scales between diseased and normal skin in dermoscopic images. The network combines spatial attention, which combines multidimensional feature maps and adaptively adjusts the target region of the network, with channel attention that follows the self-attention mechanism to obtain clear segmentation boundaries. By utilizing cascaded reverse attention modules and employing an erasing strategy, we guide the model to actively focus on important regions. When compared with seven existing segmentation networks on three open-source datasets, our approach achieves the best performance in three out of five evaluation metrics.

With the rapid development and wide application of speech technology, the use of spoofing speech has become a serious problem. Spoofing speech can be maliciously used to fake information, identity fraud and other activities, which brings serious risks to society. Therefore, effective spoofing speech attack detection methods become essential. In this paper, we propose a spoofing speech attack detection method based on EfficientNet. The feature maps obtained by speech conversion are input into the designed EfficientNet, and a classifier is used to determine the spoofing speech. The experimental results show that the EficientNets-B5 detection model is more than 90% accurate in dealing with various types of spoofing attacks, and the equal error rate is the lowest of 5.82%.

In intersection scenarios, agents exhibit diverse intent choices, making trajectory prediction problems fraught with significant uncertainty. This study proposes a cross-intersection trajectory prediction method that considers the probability of agent intent. It integrates a vehicle speed model, an intent predictor based on agent kinematics, and a trajectory prediction method based on graph neural networks to enhance the accuracy of vehicle agent trajectory predictions by precisely capturing the intent of vehicle agents at intersections. Through training and validation on a large dataset of real-world driving data, experiments have demonstrated the method's capability to predict the behavior of traffic agents in intersection scenarios accurately. Specific experimental results on the nuScenes dataset show MinADE_5, MinADE_10, MissRate_5,2, and MissRate_10,2 values of 1.70, 1.45, 0.63, and 0.48, respectively.

To address the increasing diversity in the forms and complexity of Webshells in recent years, as well as the challenge of high-dimensional feature sets, this paper proposes a Webshell detection algorithm based on the Improved Forest Optimization Algorithm (IFSFOA). This algorithm employs a process of random tree reduction and utilizes an adaptive parameter selection strategy to perform feature selection on the combined features of text characteristics and opcode at the compilation layer. It effectively filters out low-dimensional Webshell features, thereby enhancing the representational capability in feature dimensions. Experimental results demonstrate that the Webshell detection method based on the Improved Forest Optimization feature selection algorithm outperforms commonly used Webshell feature handling methods in terms of both detection rate and recall rate in recent years.

In this paper, a multi-objective optimization model with shelf stability, goods access efficiency, goods relevance, first-in-first- out, and equipment load balancing as optimization objectives is established for the optimization problem of storage space in a multi-storey shuttle truck three-dimensional warehouse. According to this model, an improved multi-objective genetic algorithm (NSGA-II) is proposed to address the shortcomings of the traditional multi-objective genetic algorithm, introduce the initial population generated by the optimization of the set of good points, change the characteristics of the initial population generated randomly, and solve the problem of uneven distribution of the population in space. (1) By adopting the parent generation selection based on linear ranking, the main idea is that the individuals of the population are ranked according to the adaptation value from small to large, and the added index term is gradually adjusted with the number of generations. In NSGA-II, the rank is first sorted in descending order, and then the crowding distance is sorted in ascending order. This method ensures the search ability in the early stage of the algorithm to expand the search space to avoid falling into the local optimum, and increases the pressure on the selection of the parent generation in the later stage of the algorithm to ensure that the algorithm's convergence speed can be better approximated to the real value. (2) Change the model evaluation method, in order to assess the uniformity and breadth of the distribution of the solution set, this paper adopts the spatial evaluation method to assess the distribution of the algorithm. Experiments under different order sizes show that the improved NSGA-II improves the performance of the algorithm by more than 45% compared with the traditional NSGA-II algorithm, and the overall performance is more significant and more suitable for dealing with the cargo space optimization problem, and it is an effective method for solving the optimization problem of the storage space in the multi-storey shuttle truck three-dimensional warehouse.

The PETSc (Portable, Extensible Toolkit for Scientific Computation) library is one of the fundamental general-purpose numerical libraries in high-performance computing environments. It is widely employed for solving problems related to partial differential equations, sparse linear algebra, and other related issues. PETSc plays a crucial role in assisting developers in rapidly creating parallel programs, thereby enhancing the computational efficiency of high-performance computing. This paper initially ports the PETSc program onto the SW26010-pro processor. Following that, we choose a representative solver from the KSP module of PETSc. Addressing four hotspot functions invoked by this solver, we implement many-core optimization for its execution under the Sunway heterogeneous architecture. The experimental results show that the maximum speedup on a single node can reach 62.3 after many-core optimizations of core hotspot functions. This indicates that the many-core optimization of hotspot functions for PETSc linear solvers demonstrates excellent parallel efficiency on the Sunway supercomputer.

Convolutional neural networks (CNNs) are a type of artificial neural networks (ANNs) that use the method of convoluting images with filter-kernels for object recognition and classification. More than 80% of the computation in CNNs is convolution which is rather difficult for traditional serial electronic computing architectures. To tackle this issue, photonic computation has harvested great attention due to its wide bandwidth and intrinsic parallelism. Here we propose an integrated photonic chip in order to accelerate convolution operations. The photonic chip consists of 144 microring resonators which can be programmed to perform arbitrary 9×16 matrix. The 9×16 matrix can be divided to 16 3×3 convolution kernels. Thus, the proposed chip is capable of processing 16 different 3×3 convolution operations simultaneously. The theory simulation results of the device design are demonstrated.

To address the issues of defect leakage and mis-segmentation in the surface defect segmentation of solid oxide fuel cells, we propose a cascaded mixed pyramid pooling segmentation method. In the encoding part, attention-based depth-wise separable convolution is implemented to replace the traditional convolution. This reduces the computational complexity and achieves hierarchical feature extraction. Secondly, cascaded sampling is applied to the last three layers of the encoder to promote multi-level feature fusion, and mixed pyramid pooling consisting of atrous convolution and strip pooling is adopted to achieve feature extraction for long-range band defects. Lastly, an adaptive loss function is constructed to supervise the training process, balancing the relative importance among the losses and improving the learning of defect features by the network. The experiments demonstrate that the proposed method improves segmentation accuracy, as evidenced by better results in mean absolute error, F1 score, intersection over union, and Pratt's figure of merit.

This signal distortion measuring device can be used to analyze total harmonic distortion (THD) of periodic signals. This device is mainly composed of a signal conditioning circuit, a signal processing MCU, a display screen and an mobile phone APP for display measurement results. The signal processing MCU uses TI's C2000 series TMS320F28379D dual-core MCU LaunchPad. This system uses the periodic signal (with harmonic components) output by the signal source as the input signal. After the amplitude amplification and level raising of the signal conditioning circuit, it becomes a signal that matches the MCU analog-to-digital conversion voltage range; The MCU's on-chip ADC is used for quantized sampling, and then the MCU's FFT operation is performed. Then calculate the distortion degree of the signal and the normalized amplitude of each harmonic. The MCU transmits the measured signal waveform through UART1 to the serial port screen for display, and transmits the waveform through UART2 through the serial port to Bluetooth module to the mobile phone APP for display.

In order to investigate the failure mode of fuze in the high temperature and high humidity environment, an accelerated aging test method of main components of fuze is established based on the humidity diffusion theory in humid-thermal coupled stress environment. Firstly, the three-dimensional model of a certain kind of fuze and the humid-thermal coupled scene are constructed by utilizing ANSYS workbench simulation platform. Meanwhile, the critical components of fuze with stress concentration are identified under different environmental conditions. Moreover, the shape variable range and stress concentration distribution of the critical components are acquired in the humid-thermal coupled environment. Finally, the simulation results show that the temperature transfer inside the fuze basically reaches saturation and the humidity distribution gradient is obvious with the loading conditions of the humid-thermal coupled environmental parameters. And, the stress concentration and shape variable generated are maximum in the turbocharger shell arrived to 0.0204mm and the stress concentration of about 10.7MPa is generated. All above results provide the data support for reliability analysis in practical storage scenarios.

This study combines geometric features based on Point Pair Features (PPFs) and spatial features under polar coordinates to innovatively design a series of new local feature descriptors. Through an exhaustive examination and comparison across four benchmark datasets (B3R, U3OR, U3M, and QuLD) in relation to these descriptors, along with the assessment against three contemporary descriptors (SHOT, LFSH, and SDASS), the superior efficacy of the novel descriptors becomes evident. Experimental findings showcase that, in terms of overall efficacy, fr(α1α2α3) stands out as the most remarkable performer, showcasing exceptional compactness and efficiency. Upon rectifying LRF/A errors, descriptors founded on 'rea' exhibit optimal performance, underscoring their exceptional discriminative prowess. In contrast to established descriptors like SHOT, LFSH, and SDASS, the recently introduced descriptors, such as fr(α1α2α3), fre(α1α3), and fr(α1α3), manifest significant advancements in terms of compactness and efficiency. The significance of this research lies in furnishing a set of local feature descriptors that excel in three-dimensional computer vision, and it elucidates their potential applications in 3D pairwise registration through empirical studies.

A method for optimizing the layout of electronic components of PCB board based on sequence pair representation is proposed. Starting from three aspects of heat conduction, heat convection and heat radiation, the PCB board level thermal analysis is carried out, and combined with the specific action mechanism of the three, the optimization goal of PCB electronic component layout is to set a reasonable number and area of heat dissipation channels, optimize the arrangement and spacing of components, and minimize the role of heat on the PCB board. In the specific optimization stage, the relationship between the high power components and their fins or fans is taken as a sequence pair, the minimum spacing and directionality are taken as constraints, and particle swarm optimization is used to solve the optimal layout. In the test results, although the PCB board temperature also showed a steady upward trend with the increase of continuous operation time, it was always stable within 26.0 ℃, and there was no obvious high temperature.

With the rapid development of water conservancy and hydropower in the largest developing country in the East, the country's mining industry is also developing rapidly. With the continuous development of mineral resources, good waterproofing and drainage work is a necessary condition to ensure the safety and normal production of open-pit mining. In particular, underground mining is objectively restricted by conditions such as precipitation, surface runoff, and underground excavation, and these conditions are necessary conditions for the entire production process and infrastructure phase of open-pit mining. This article uses the flow field method to detect leakage around the mine pit. After energization, sensing electrodes are used to detect the measured area. Due to the existence of leakage, the resistivity of the leakage location will be lower than that of the non-leakage location. The drain position will become the main channel for the current flow. In actual experiments, it can be found that the effect is good and meets the required requirements.

During the test flight of civil aircraft, the prediction method recommended by ITU-R P.533 was used to predict the frequency of the high frequency(HF) system at the situation of the high-altitude and long-distance communication. There were many cases which the communication effect was lower than the design expectations. Through comparative analysis of test flight scenarios and scenarios recommended by ITU-R P.533, it is found that the flight altitude of the aircraft is a key factor affecting the accuracy of frequency prediction, and the communication distance will also exacerbate the impact of flight altitude. This paper establishes the test flight model of the airborne high frequency system to obtain the correction amount of flight altitude and communication distance for the frequency prediction. It proves the correctness and effectiveness of the method discussed in this paper through multi-scenario test flights.

A low ripple fast boost CMOS charge pump is designed for charging mobile phones. The charge pump has the advantages of low power consumption, low cost, and simple circuit. In the design of the charge pump, in order to improve its boosting speed, voltage boosting circuits for pre-charging the charging node and substrate automatic following circuits to overcome the substrate effect of field-effect transistors are designed separately; And the simple common drain amplifier is used to reduce the charge pump’s output ripple effectively. Based on CSMC 0.5 μm process and Dickson charge pump, a clock generation circuit, a charge pump circuit, a voltage regulation circuit, a pre-charging circuit, a substrate effect elimination circuit, and a common drain amplifier are designed. The charge pump circuits are designed and simulated by cadence. The simulation results show that the static current of the circuit is a few μA, the input voltage range is from 1V to 3V, and the output voltage is 4V. And the boosting time is below 100 μs, the ripple is less than 1mV, and when the input voltage is 1V, the static power consumption is only 0.3mW.

Aiming to address the issue of poor holographic image reconstruction quality and significant impact on observer experience resulting from the pre-appended phase mask extended holographic display viewing angle, a novel high-quality reconstruction method is proposed. The method consists of two steps: 1) Utilizing the GS iterative phase optimization algorithm to enhance pure phase holograms, which significantly improves the reconstruction quality of individual phase holograms when a random phase mask is pre-appended to expand the viewing angle. 2) Employing a time-averaging technique for holographic image reconstruction based on the visual "persistence effect" of the human eye to effectively suppress speckle noise in the holographic reconstruction results. Through the aforementioned two steps, numerical simulation experiments were conducted, and the results indicate that the combination of the GS iterative optimization algorithm and time-averaging method produces holographic reconstructions with clear details, low speckle noise, and high quality for the phase mask extended holographic display viewing angle method.

With the rapid development of integrated circuit technology, the scale of power grid matrices is growing larger. Power grid analysis consistently presents formidable computational challenges, involving the solution of matrices with scales ranging from millions to even billions. Power grid analysis has also become one of the most time-consuming steps in chiplevel analysis. Therefore, efficient power grid analysis techniques are highly sought after for their capacity to conserve computational resources and expedite iteration speed. In this paper, leveraging the characteristics of power grid structures, we propose a text-based matrix reordering and decomposition technique. Through this approach, the original matrix of the power grid system can be seamlessly decomposed into two submatrices without compromising the integrity of the original matrix information. The submatrices can still maintain symmetric positive definiteness, enabling acceleration throughout the network solving process without altering the original solution. In contrast to current power grid hierarchical schemes, power grid decomposition technology, in order to ensure its computational accuracy, requires expensive costs to determine the boundary voltages and currents of the subgrid. The proposed technique exhibits wide applicability across various solvers, and we perform comparative analyses by selecting several representative solvers for static analysis of power grids. Experimental results demonstrate that this technique achieves nearly 100% speed improvement. For solvers lacking support for parallel computation, this approach represents an efficacious means of enhancing computational efficiency. Additionally, we observe that solvers designed for parallel acceleration can also achieve performance improvements exceeding 50%.

The information data in the electromagnetic environment space is extremely large and complex, so understanding the electromagnetic space is very important and necessary. This article will conduct an study of the application of attention mechanism in electromagnetic space cognition and explore its innovations in signal recognition, spectrum sensing. For the modulation identification and classification of electromagnetic signals, an improved Transformer model integrating data slicing and gating units is proposed. The model first performs patch processing on the data, then introduces a gating mechanism to adjust the channel-wise and step-wise, and at the same time learns the optimal parameters through training. Experimental results show that at a lower signal-to-noise ratio, this model has an accuracy improvement of about 10% compared to other classic models. At the same time, by observing the confusion matrix diagram in the experiment, the model's performance in high-order modulation recognition has also been greatly improved.

Improving the uniformity of electromagnetic field and temperature in the resonant cavity is a hot issue in microwave heating. In this study, a new method for the design of slot waveguides applied to microwave heating is proposed, which is based on Taylor synthesis method and multi-physics calculations. The initial length, width, and offset between the centerline of the slot are designed by Taylor synthesis method, and the spacing between two slots, including the number of slots, is also determined by Taylor synthesis method. Subsequently, by combining electromagnetic fields, temperature fields, and specific characteristics of the heated object, multi-physics calculations are performed, and the above parameters are further optimized to design the slot waveguides. The results show that the slot waveguides designed by the proposed method produces a more uniform temperature and electromagnetic field distribution than the conventional method.

Wind turbine generators, which convert wind energy into electricity, require particularly crucial inspection of their internal blade cavities. To achieve precise inspection of the internal blade cavities, a quadcopter drone was designed with collision prevention capabilities and equipped with dual-camera stereo vision to obtain 3D point cloud reconstruction and achieve three-dimensional defect coordinate localization. The application of this advanced technology significantly enhances the efficiency of internal blade cavity inspections. During on-site testing, the drone effectively navigated a 32-meter-long blade cavity, reducing safety risks and providing precise defect location information for subsequent maintenance. This study not only keeps inspection personnel away from hazardous areas but also enables inspections to cover regions inaccessible to individuals, thereby ensuring the safe and stable operation of wind power generation.

Electronic Medical Records (EMRs) is the original record of the whole process of hospital diagnosis and treatment, and there exists a large amount of medi-structured and unstructured text, which is inconvenient to be used for subsequent auxiliary diagnosis and disease prediction tasks. Therefore, structured processing of pediatric EMRs is necessary in this paper. The structured processing process includes four parts: structured analysis, data preprocessing, joint extraction and data integration. Among them, joint extraction is the key step of the processing process, and this paper adopts the BARTbased joint extraction model JREwBART, which completes the task of joint extraction of pediatric EMRs after training using the constructed pediatric dataset, and the F1 value of JREwBART is improved by 1.5% compared with other models. The final extracted information is integrated at text level to achieve the structured processing of electronic medical records.

UnifyFS, a user-level file system specifically designed for burst buffers, aims to integrate the existing node-local storage layers in parallel computing systems, thereby providing a unified namespace for parallel computing applications. This study details the successful integration of UnifyFS with the Slurm resource manager, and emphasizes the synergistic effects of this integration in enhancing the performance of parallel computing systems. Detailed evaluations of this integration on leading parallel computing systems show that the integrated system exhibits significant advantages in scalability of data writing operations, and achieves notable performance improvements in the read-write bandwidth of applications, bringing higher efficiency and improvements to parallel computing applications. The paper thoroughly analyzes the effective integration mechanisms of UnifyFS and Slurm, and explores its potential in optimizing the collaborative operation of distributed storage and task scheduling, offering a viable solution to the broad parallel computing community. Future research will focus on further optimizing this integration framework, adding more advanced features, to achieve broader applications and technological innovations.

Because of the complex geometry and changeable materials, it is difficult to control the noise in the interior of special-shaped buildings. For this reason, a new method is proposed to analyze and control the indoor noise characteristics of special-shaped buildings considering the structure-acoustic sensitivity. This paper analyzes the indoor noise characteristics of special-shaped buildings, calculates the indoor structural noise of special-shaped buildings, analyzes the structure-acoustic sensitivity, and obtains the normal velocity amplitude sensitivity of structural response by using finite element method. Sound waves are soundproofed by using the different vibration characteristics of materials to control the indoor noise of special-shaped buildings. The experimental results show that the acoustic sensitivity of the proposed method is higher than 70dB, the longest control time is 29.88s, the peak value of indoor vibration basically disappears, and the abnormal noise completely disappears.

Against the backdrop of growing demands for efficient healthcare systems, balancing patient privacy with the accuracy of hospital readmission rate predictions has emerged as a prominent research issue. This study aims to address this challenge by proposing a differential privacy method that combines substitution cipher technique with noise addition. Initially, this approach anonymises medical data through substitution cipher, effectively blurring sensitive information while ensuring data integrity. Furthermore, Gaussian noise addition mechanism is utilized to maintain data privacy without excessively compromising predictive accuracy. Rigorously tested, this method demonstrated outstanding performance with an AUC of 0.8816, marking significant progress in accurately predicting hospital readmission rates while ensuring patient privacy protection. This achievement holds substantial theoretical and practical significance in the field of medical data analysis and offers new directions and methodologies for future related research.

Currently, large language models (LLMs) are increasingly capable of processing natural language problems. However, there is still limited understanding of the internal mechanisms these models employ for such tasks^[1]. Additionally, existing works on model interpretability face challenges such as low complexity of interpretable language tasks, smaller scale of interpretable models, and lengthy attribution times. In our paper, we use the method of gradient accumulation to attribute the process by which the Vicuna-33b model handles the classical Winograd Schema Challeng (WSC) natural language task, explaining the behavior of large models from their internal components. Gradient accumulation helps us rapidly identify the intrinsic logic in natural language problem-solving within the model, and understand the importance of hidden units and their role during the forward pass. Compared to existing interpretability methods, this approach significantly improves in terms of attribution efficiency and the scale of models it can attribute.

The demand side power load in the power system is affected by various factors, and relying on experience to ensure stable power operation is complex and difficult. Only accurate prediction of power load can provide scientific and reasonable scheduling and power generation plans for power system scheduling. In response to the problem of low accuracy in power load forecasting relying on manual experience, this article uses Pearson correlation coefficient to conduct correlation analysis on the influencing factors of power load and establish a power load dataset. On this basis, a convolutional neural network and a bidirectional long short-term memory recurrent neural network are constructed to construct a power load prediction model, and attention mechanism is integrated to use the whale algorithm to search for the optimal parameters that match the network structure. Train the dataset through predictive models to obtain predictive data and achieve power load forecasting. The experimental results show that the constructed power load prediction model can predict power load data, providing scientific basis for guiding power system scheduling.

This study introduces a novel method for analyzing how the choice of a reference camera influences the super-resolution reconstruction in camera array imaging systems. Through a combination of simulations and experimental validations, it becomes evident that traditional, uniformly arranged camera systems often fall short of achieving optimal super-resolution effects, irrespective of the reference camera chosen. The introduction of a central camera into this conventional arrangement, either by adding it or substituting an existing camera, not only yields improved super-resolution outcomes but also significantly enhances the system's robustness to variations in object distance. This advancement notably elevates the functional utility of sparse camera array systems in practical applications.

In order to meet the needs of calligraphy and digital creation of ink wash painting, this paper establishes a dual-layer paper simulation model to achieve the effect of irregular diffusion of ink on rice paper during the ink creation process. The duallayer model divides rice paper into a contact layer and a diffusion layer. By extracting the range of the user's input strokes and using the relevant information as the basis for initializing ink in the contact layer, the ink stored in the contact layer will proportionally flow to the diffusion layer. The diffusion process of ink in the diffusion layer is simulated using an improved coefficient diffusion equation. Experimental results demonstrate that the model can realistically simulate the ink diffusion effect in the ink creation process in a real environment, providing convenience for artists in digital platform creation.

Pillar porcelain insulator is an important component of power plants and substations, a large number of use in the power system, play a role in supporting the conductor, circuit breaker and high-voltage switching, however, porcelain insulators charged with the detection technology research field there are still gaps. For this reason, this paper for the detection technology in the information acquisition circuit system to carry out the design, which is mainly divided into three parts of the information acquisition and processing module, respectively, for the following circuit, subtraction and amplification filter circuit and analogue circuit noise analysis. The role of the signal and processing circuit is mainly for the detector output to provide a suitable load to accurately obtain the detector output of the analogue signal, and the analogue signal modulation so that it can be analogue-to-digital conversion circuit for better quantification. Finally, the correctness and effectiveness of this part of the design is verified through field tests.

Digital twin achieves interactive mapping between physical entities and virtual models, enabling continuous real-time monitoring of process quality in virtual spaces. This paper addresses the issues of high latency and poor correlation with heat sealing process parameters in the quality inspection of plastic film heat sealing, and explores real-time analysis of heat sealing quality based on digital twin technology. The focus is on the finite element analysis of typical heat sealing process parameters combined with thermo-structural coupling, boundary condition modification and data-driven modeling. The study reveals the relationship between controllable heat sealing process parameters and film heat sealing quality, achieving real-time monitoring of the heat sealing quality of plastic sealing film

Wildfire detection is an important societal, economic, and environmental task. Accurately and timely identifying wildfires is crucial for protecting natural resources and ensuring the safety of human life and property. Currently, deep learning based algorithms for wildfire detection in the wild suffer from issues such as low detection accuracy, high false positive rates, and slow inference speeds. This paper proposes a wildfire detection method, WG-YOLOv5, based on the improved YOLOv5 architecture. The method introduces two-dimensional wavelet transform into the YOLOv5 network model to extract frequency domain information from wildfire images and integrates it with spatial features extracted by the backbone network. Additionally, the YOLOv5 neck structure is optimized based on GSConv. Experimental results on a collected dataset of wildfire images demonstrate that WG-YOLOv5 outperforms YOLOv5 and other comparison methods in terms of higher accuracy, lower false positive rates, and improved real-time performance. This makes it well-suited for real-time wildfire detection tasks in wild environments.

This paper explores deep learning methods for bridge surface defect detection, particularly an improved model based on YOLOv7. To address the detection of bridge-type defects, a Dual-Stream Attention Module (DSAM) and a Hybrid Atrous Pyramid Module (DP) were introduced, aiming to enhance the model's capability to capture key features and the efficiency of multi-scale feature extraction. Experimental results show that the improved model demonstrates higher detection accuracy on a bridge defect dataset, with the YOLOv7+DSAM+DP model achieving a mean Average Precision (mAP) of 91.3%. Additionally, the study compares the SSD and Faster R-CNN networks, confirming the superiority of the proposed model. Overall, the model presented in this paper significantly improves the precision and efficiency of bridge surface defect detection by integrating attention mechanisms and hybrid atrous pyramid modules.

Ultra-short-term wind energy prediction refers to the prediction of wind electricity in the subsequent few minutes to a few hours. It is of great significance for the operation and management of wind farms, but due to the randomness and nonlinearity of wind speed, this task is very difficult. In this paper, an ultra-short-term wind power prediction method based on CNN- -BIGRU-Attention uses convolutional neural network (CNN) and bidirectional gated recurrent unit (BIGRU) to form a deep neural network, which can effectively extract the spatio-temporal characteristics of wind speed data and enhance the prediction ability of the model by using attention mechanism (Attention).First, the data is preprocessed to normalize the data and transform the raw data through the data preprocessing module. Secondly, the deep learning module is entered, and the data set is processed and analyzed through CNN and BIGRU to Attention layer. Finally, a wind energy prediction model for CNN-BIGRU-Attention was developed. The input of the CNN-BIGRU-Attention model is multidimensional wind speed data, including historical wind speed, real-time wind speed and future wind speed, and the output is the predicted value of wind power in the future. This paper conducts experiments on two real wind farm data sets, and the results show that the proposed method outperforms other comparative methods in ultra-short-term wind power prediction, with high accuracy and stability.

To solve the problem that the data volume of the monitoring video is large, and the bandwidth requirement for the transmission of this data is high, an adaptive-rate block compressive sensing scheme is proposed in this paper. The sparsity of image blocks in the target video frame is estimated firstly by observing the initial measurements with a fixed low sampling rate, and then the number of measurements can be determined for the current block based on the estimation result, the supplementary measurement is performed to obtain the final sampling results. The simulation results show that the proposed algorithm can effectively improve the quality of the reconstructed video without significantly increasing the computational complexity.

With completing the basic exploration of shallow mineral resources in China, exploring deep mineral resources has become an important solution to the resource shortage. The traditional transient electromagnetic method mainly uses loop-source devices, which have a relatively limited detection depth. However, grounded electrical sources have a strong ability to conduct deep field surveys. Currently, research and application of the transient electromagnetic method for grounded electrical sources are still advancing. This paper calculates the transient electromagnetic response of grounded electrical sources based on the layered formation model. It also studies and analyses the plane distribution characteristics of each electromagnetic component at different moments. Additionally, the optimal observation range of line sources with different lengths is investigated. The research results are significant for exploring the grounded electrical source transient electromagnetic method.

With the accelerated process of Chinese-style modernization and the establishment of a new energy system, the requirements for the construction of a modern intelligent power distribution network are constantly improving. For the comprehensive assessment of the distribution network, the self-attention algorithm can fully take into account the correlation between various indicators in the face of the huge and complex distribution network, so as to carry out the comprehensive assessment in a more comprehensive and efficient way. Therefore, this paper proposes a fuzzy comprehensive evaluation method based on self-attention mechanism. Firstly, the indicator system is established, the attention weights of all indicators are calculated by the self-attention algorithm, and then the assessment results are derived by combining the relevant regional data through the fuzzy comprehensive evaluation method. The experimental results show that the adopted evaluation method can efficiently and comprehensively derive the evaluation results, and at the same time has good training stability and flexibility.

Ground Penetrating Radar (GPR) offers the benefits of non-destructive testing (NDT) and high efficiency, making it extensively employed for subsurface material detection. However, conventional GPR image interpretation heavily relies on researchers' experiential knowledge, which will lead to reduced detection efficiency and more recognition errors. To address these challenges, this study proposes a new technology to achieve GPR detection in high efficiency. By enhancing the network structure of YOLOv5 through incorporating attention mechanisms, data augmentation techniques, and other approaches, we aim to mitigate issues related to false detections and missed identifications during real-time GPR-based localization of directly buried underground cables.

This article establishes a multi-body dynamic model of a certain type of tracked vehicle through dynamic simulation analysis software, which can effectively simulate and analyze its driving characteristics. This article uses the high-speed tracked vehicle toolkit in Recurdyn software to establish a vehicle simulation analysis model, including driving wheels, support wheels, suspension devices, and road surface models. Preliminary simulation analysis is conducted on the maneuverability, preparing for subsequent research on parameter optimization and performance evaluation of this type of vehicle.

In the current security analysis and research of AES cryptographic algorithm, most of them use energy attacks through computer simulation, microcontroller experiments and other means. This is not only very different from the actual power consumption generated by the AES hardware circuit used in practice, but also Moreover, there are obvious differences in the power consumption caused by different bus structures that appear in the actual use of AES. This paper analyzes the security of the hardware circuit implementation of an AES algorithm under different bit input data. It uses FPGA as the hardware circuit foundation, uses the mean difference method in DPA as the security analysis method, and uses the first round of The S-box is used as the attack point to attack, and the 128-bit source key can be derived in turn. Finally, the corresponding security is discussed.

Based on the requirements of the low-carbon energy system, it is necessary to optimize the 1.2MW doubly-fed wind generator. Based on the operating characteristics of the generator, selected the cost of generator materials and the efficiency at rated load as the objective function, transforming the engineering problems into an optimization problems, and establish a relevant mathematical model for generator optimization. Due to the fact that generator design is a complex problem with multiple constraints and nonlinear, a BSO algorithm improved by BAS is proposed, which can effectively handle nonlinear optimization problems and achieve optimal design of 1.2MW doubly fed wind generator. The results show that the algorithm has good convergence, stability, and speed, while also reducing the probability of falling into local optimal solutions. The final optimized design scheme is also relatively ideal.

High-performance multi-core processors serve as the foundational hardware for intelligent computing systems. Compared to traditional multi-core CPU processors, they achieve higher processing performance and lower power consumption by densely integrating multiple general-purpose processing cores with heterogeneous specialized accelerator units on the chip. Currently, high-performance multi-core processors are widely employed in commercial sectors such as cloud data centers, edge processing nodes, and terminal devices. This paper analyzes and summarizes the research achievements of leading global institutions in high-performance processor chip development, including Intel, AMD, IBM, ARM, etc. The focus is on novel multi-core processors such as Intel's Alder Lake series with heterogeneous cores, AMD's latest Zen3 processing core, IBM's new Power10 processor, and Telum processor. The paper also outlines the key technological trends in the development of these multi-core processors, providing a foundational technological support for the future hardware platform development of intelligent computing systems.

Due to the nonlinear and uncertain characteristics of aerospace electromechanical systems, the temporal information is challenging to fully exploit during the fault diagnosis process. In order to address this issue and further enhance the accuracy of fault diagnosis models, this paper proposes a fault diagnosis approach that integrates a One-Dimensional Convolutional Neural Network (1D-CNN) with Bidirectional Gated Recurrent Unit (Bi-GRU). Initially, a sliding time window method is employed to prepare samples from the raw data as input for the 1D-CNN. Subsequently, the features extracted from the convolutional and pooling layers are fed into the Bi-GRU, and the diagnostic results are output through a fully connected layer. Finally, the algorithm model is embedded and deployed in the Xilinx Zynq 7Z100 environment. The proposed algorithm is validated, and the algorithm model undergoes optimization using the High Level Synthesis (HLS) tool to demonstrate the feasibility and necessity of model optimization.