Galvanized Sheet Surface Defect Segmentation Algorithm

Textured image segmentation is the connection of junior and senior vision,and one of the most important aspects of the majority of vision systems,its result related to the ultimate output of the senior vision.As a kind of complex texture images,the defect segmentation of galvanized sheet is the most important part of the quality inspection system,whether the high efficiency and precision will work on the following classification and identification,so as to the whole system's performance.In this paper, based on image processing technology,we introduced two efficient approaches for galvanized sheet defect segmentation.An improved spectrum residual based method has been applied to the surface defect detection of galvanized sheet,explored its ability in textured image segmentation. By removing the averaged component from the log spectrum,we can remove the periodic,repetitive patterns of input images,only the unexpected informative signals can be delivered to later stages of processing.A new approach for galvanized sheet defect segmentation based on 1D Gabor filter has been proposed.Gabor filters have been well recognized as a combined spatial, spatial-frequency representation for analyzing textured images,and have optimal combined localization in both the spatial and the spatial-frequency domains.We first use a 1D ring-projection transformation to compress a 2D grey-level image into a 1D pattern,and then employ a 1D Gabor filter to detect defects embedded in a homogeneous texture.This method is very efficient and effective for detecting defects in structural textures such as wooden surfaces,and for statistical textures such as galvanized sheet,sandpaper,leather and so on.Carried out a detailed analysis of the performance of optimized FIR filters,and a detailed comparison with the algorithms introduced in this paper from accuracy, robustness and computational complexity finally summed up their characteristics:■The spectrum residual based method can locate or segment the defect accurately, and insensitive to environmental changes such as lighting variation;this method is simple and fast,is a totally non-supervision defect segmentation method. ■The result of single channel optimized FIR filters depend heavily on the defect textures used in the training process,this method is poor in robustness.The performance of the multi-channel optimized FIR filters is not changed with the defects size,direction and background roughness.It is shown that multi channel optimal filter has strong abilities,but how to fuse each channel's information to reach the best segmentation quality most of time depends on experience.■The 1D Gabor filter segmentation method has a good local characteristics,the experiments on wooden and galvanized sheet have show that it's segmentation results batter than or close to the result of multi-channel optimized FIR filters,and this method does not need any defect texture information,have a strong robustness.