Research On Image Defect Detection For Structured Texture Industrial Products

With increasingly fierce competition,defect detection plays an important role in the process of automatic industrial production.Defect detection usually refers to the detection of surface defects,the shortcoming of traditional manual detection are low efficiency,unreliability and other issues,so now the surface defect detection is based on advanced machine vision detection technology,to detect the spots,pits,scratch,chromatic aberration and other defects.The detection technology which based on machine vision have a lot of advantages such as more efficient,convenient,economic and security.The products of structural texture background are widely used in industrial production such as textiles,wood,semiconductor products and so on.Therefore this thesis is devoted to the study of image defects of industrial products with structural texture background,and provide reliable and efficient algorithms.(1)The feature extraction and classification algorithm is the key in defect detection algorithm of structured background textures.In view of the feature extraction,this paper improves the existing local inlier-outlier ratios(LIO)features by studying the relevant feature extraction algorithms and proposes the feature of similarity-dissimilarity ratios(SD),which has good classification accuracy.In order to classify the defect images of industrial products with structural texture background,ELM(Extreme Learning Machine)classification algorithm is proposed.Two classification methods are proposed to achieve a simple binary classification of structured texture industrial images:GLCM(Gray Level Co-occurrence Matrix)classification method,M-LBP(modified Local binary patterns)features and SD feature classification method.(2)In order to extract image defects of industrial products with structural texture background,this paper proposes a defect area extraction method based on generalized low-rank approximations of matrices(GLRAM).In this method,the background of the industrial product image is first modeled using the generalized low-rank approximation to obtain the texture background of the image of the industrial product.Then the image with defects and noises is obtained by using the background difference.Finally,the defect area can be extracted by a simple threshold segmentation algorithm.This method is not sensitive to image rotation and illumination variation,and it is suitable for the product testing of various structured texture backgrounds and is an efficient and reliable algorithm.Through the above methods for feature extraction,being testing image classification and defect extraction,defect detection of industrial product images with structured texture is realized.The final experiment proves that the classification method based on ELM can achieve a classification accuracy of 94.46%.Compared with the existing literature results,the defect extraction method based on generalized low-rank approximation can extract the defects better and the robustness is good...