Research On Methods Of Hierarchical Defect Detection And Recognition For Silicon Steel Surface

Silicon steel materials are mainly used as magnetic materials in motors,transformers,electrical appliances and electrical instruments.With the development of machine vision technology,using visual detection and recognition technology to detect and identify surface defects shows more and more advantages.In order to improve the quality of products,the precision positioning and recognition of surface defects are still valued by silicon steel enterprises,and related research is invested to improve the accuracy and efficiency of detection.In this paper,image processing,pattern recognition and other related technologies are used to analyze the problems existing in the current steel surface defect detection and recognition tasks.This paper adopts the mechanism of hierarchical detection and recognition to complete the detection of silicon steel's surface from rapid initial detection,accurate detection and tiny detection to the recognition of defect types.In addition,the validity of the method is verified by experiments on data sets for different key problems.The main research content and achievements of the thesis are as follows:(1)An automatic defect detection method based on robust principal component analysis model is proposed for the problem of primary automatic defect detection on silicon steel surface.In the process of real-time detection,the current image sequence is decomposed into low-rank principal components,sparse defect targets and noise by using the robust principal component analysis model.OTSU threshold method is applied to the sparse defect target image to complete the task of surface primary defect detection.In this paper,three optimization methods are used to solve the model of robust principal component analysis and the experiments on the defect images illustrates the performance;(2)In order to solve the problem of exact defect detection at pixel level of silicon steel surface,an approach of target saliency extraction with binary structure constraint is proposed.First,the binary candidate region of the image is generated by using the binary classifier based on convolutional neural network,and then the binary information of each pixel is determined by voting based on the weight of the candidate region.Finally,the k-nearest neighbor enhancement graph is used for significant propagation,and the final accurate region is obtained.By comparing with the detection results of other methods,this method can not only accurately locate the surface defects of silicon steel,but also maintain a stable detection effect under the interference of noise.(3)The complex texture on the surface of silicon steel makes it difficult to detect the small defect target.Therefore,a pixel-level machine learning method is proposed to detect the small defect target on the surface of silicon steel.First,Graph Cut technique is used for coarse segmentation to obtain the over-segment regions,and then Fuzzy c-means clustering technique is utilized to refine the segmentation result.According to modeling the two regions using the generalized Gaussian model,the pixel is determined as the defect or background through the Bayesian probability estimation model.Through the comparison experiment,the proposed method has obtained the best result in the small defects on the surface of silicon steel.In addition,this method can achieve higher correct detection rate and lower error detection rate;(4)In order to solve the problem of introducing threshold manual setting into common local feature descriptors,one kind of local feature descriptors with manifold structure constraints is proposed to complete the task of defect recognition.This method first calculates pixel difference vector,then the iterative approach is adopted to establish the manifold structure of the vector.The maximum margin criterion is introduced to achieve the discriminant projection of the manifold learning.Finally,weighted distance measure is introduced to improve the accuracy of recognizing the characteristics in the process of matching.The open texture data set is employed to validate the method for texture feature extraction ability.According to the results of comparative methods,the recognition rate of the proposed method is higher than the compared ones.Then,the results of the method and the comparison method are verified by using silicon steel surface defect data set.