In complex detection scenarios in machine vision, identification of the pixel object label (comprising the detection region and detection pattern information) and the pixel part label (comprising the detection part type and location information) are often essential. Subsequent analyses of the detection region and pattern information are also carried out in such cases. In this context, a method to reuse the feature map has been studied in this work. First, the utility in reusing the feature map in the fast execution of semantic segmentation at the object-level and part-level for images with complex backgrounds is discussed, which operates by striking a balance between the degree of feature reuse and corresponding segmentation accuracy. Secondly, a semantic segmentation network, based on two independent and identical encoderdecoder pairs in parallel, is proposed. In this parallel network, feature maps with identical hidden layers are merged together for reuse in order to reduce the computational complexity of associated processes, and thereby optimize segmentation time. Finally, simulation experiments are conducted to evaluate the proposed technique. The results demonstrate that the method exhibits 100% detection accuracy, and reduces the detection time by 20.3% — from 602 ms to 480 ms — when applied to chassis assembly detection. Further, the effect of segmentation time optimization on the detection process is evaluated.
|