Research On Detecting Method For Image Of Surface Defect Of Steel Sheet Based On Visual Saliency And Sparse Representation

The research background of this paper is online detection technique based on machine vision for the surface defect of steel sheet,the foundation of research is visual saliency and sparse representation,some key detection methods of screening,segmentation and classification of the image of surface defect of steel sheet are proposed to solve some shortcomings of traditional detection methods in speedability,accuracy and robustness.Meanwhile,an experimental platform of online detection of surface defect of steel sheet is constructed to further verify the proposed methods.The paper mainly contains the following four parts:Firstly,the gradient projection based screening method for the surface defect image is introduced.In order to reduce the image data and enhance the real-time level of detection system,it is necessary to screening a small part of surface defect images from mass surface images before segmentation and classification.To solve some shortcomings of traditional screening methods in poor screening result and weak robustness,the method based on gradient projection is proposed to screening the surface defect image.In the first place,two histograms of horizontal and vertical projection are calculated separatedly from the gradient map of surface defect image.Then,the means and standard deviations of two histograms are calculated.Next,the proper screening threshold is confirmed by the difference of means and standard deviations between surface defect images and normal surface images.In the end,the surface defect image is determined by the screening threshold.Secondly,the visual saliency based segmentation method for the surface defect image is introduced.As the surface defect is regarded as an object of visual salient foreground,the surface defect image can be decomposed into a linear combination of the salient foreground and the non-salient background.To solve some shortcomings of traditional segmentation methods in low accuracy and low generality,the method based on visual saliency is proposed to segment the surface defect image.In the first place,the feature matrix of surface defect image is constructed by low-level features based on color,edge and texture,mid-level features based on superpixels and high-level features based on prior knowledge.Next,the surface defect image is decomposed into the image of surface defect and the image of background by the exploited method of double low rank and sparse matrix decomposition,therefore,the salient map of surface defect is obtained.After that,the integrity and uniformity of the surface defect area is further improved by post-processing of the salient map.Finally,segmentation of the surface defect image is finished by threshold processing of the high-quality salient map.Thirdly,the discriminative sparse representation based classification method for the surface defect image is introduced.For the influence of steel sheet rolling and external environment,there is a great of differences in surface defect of the same category,and a lot of relevances in surface defect of the different categories.To solve some shortcomings of traditional classification methods in extraction of feature vector and low precision,the method based on discriminative sparse representation is proposed to classify the surface defect image.According to "difference in intra-class" and "similarity in inter-class" of surface defect images,the method of discriminative dictionary learning based on the class-specific sub-dictionary and the shared sub-dictionary is exploited.The exclusive features of surface defect images of the same category is characterized by the class-specific sub-dictionary and the common feature of surface defect images of the different categories by the shared sub-dictionary through the dictionary learning constrained by structured incoherence and Fisher discriminative rule.The new dictionary that constructed by all the class-specific sub-dictionarys and the shared sub-dictionary not only has the compact structure,but also has the excellent ability of reconstruction and discrimination.Therefore,there is the better discrimination in sparse encoding vector of the surface defect image.The classification of surface defect image is carried out by the reconstructive error between dictionary and sparse encoding vector.Fourthly,the experimental platform of online detection of surface defect of steel sheet is built by some appropriate hardware equipments.The effectiveness and efficiency of the proposed methods of screening,segmentation and classification of surface defect images is verified by the experimental platform.The experimental results further indicates that the proposed methods not only can improve the detection rate and recognition rate,reduce the missing rate and false alarm rate,but also can keep better robustness for the diversity,randomness and complexity of surface defect of steel sheet.