|
Thermal infrared images have been widely employed to detect defections. However, it is challenging to identify the defects from thermal infrared images contaminated by shadows, noise, etc. Those factors do be recorded in the thermal infrared images such that the recorded pixels not only contain the surface temperatures but also have the impacts from the factors. Those factors illustrated in the recorded pixels can be called intensity inhomogeneity. Several researchers have reported that the multiplicative way is feasible to approximate intensity inhomogeneity. A gaussian function is introduced to make sure that the intensity inhomogeneity works on not only a particular pixel but also its neighborhoods contribute. Usually, the fixed window size illustrated in the Gaussian function is used for the whole image for simplification. Intensity inhomogeneity is not spread uniformly over the given image. For those areas with high-intensity inhomogeneity, the Gaussian function with larger window sizes is introduced; otherwise, the Gaussian function with smaller window sizes is supposed to apply to the low-intensity inhomogeneity areas. Adaptive window sizes were proposed such that each pixel will have the Gaussian function with the specified window sizes according to the amount of intensity inhomogeneity. However, the adaptive approach needs a lot of computation, so defect detection will be slowed down. This study proposes the algorithm to modify the algorithm of adaptive window sizes provide an efficient way for defect detection. The proposed approach is based on image entropy; the image entropy will have a bigger value while the intensity inhomogeneity is larger. The image entropy was classified such that limited window sizes were introduced. In doing so, intensity inhomogeneity can be approximated, and the results of image segmentation can identify the defect.
|
