Research On Surface Defect Detection Based On Image Processing

Surface quality of industrial products is an important component of its quality,and also an important performance of its commercial value.With the advent of the digital age,the hardware of machine vision has developed rapidly,and the advantages of visual detection based on image processing in surface defect detection are becoming more and more obvious.At present,the machine vision detection is widely used,and the image processing technology has become the focus of research.Therefore,the key technologies of surface defect detection based on image processing are systematically studied in this paper.Based on the main process of surface defect detection,this paper applies Super-Resolution Reconstruction,Convolution Neural Network,Sparse Coding,Low-Rank Decomposition,Image Measurement,Reliability Estimation and other related fields to solve five scientific problems,which includes pre-processing of defective image,defective image classification,defective object extraction,defect feature measurement and application of defect detection.The corresponding problems are further studied in this paper,and reflects a good originality and flexibility.The main research works and innovation include the following five aspects.?1?Aiming at the problem that the original defect image is not high resolution because of the influence of acquisition equipment,optical aberration,diffraction blur and noise,a new method of super-resolution reconstruction of defect image based on multi-frame blind deblurring is proposed,which provides a new way to improve the resolution of defect image.The main innovations of this method include:an improved forward MFSR model is proposed to overcome the contradiction between the rotation matrix(2₆)and the high-resolution image,but the low-resolution image is usually used to estimate it,which can reduces the inaccuracy;the improved forward MFSR model is transformed into a MFBD model,which is easier to solve than the former one;The MFBD problem is transformed into two related sub-problems,solving high-resolution image and solving fuzzy kernel,which are solved by ADMM algorithm.Experiments on simulated and real images demonstrate the superiority and robustness of the proposed algorithm.?2?Aiming at various types of surface defects,a new method of defect image classification based on convolution neural network and sparse coding is proposed,and the classification of defects are realized.The main innovations of this method mainly includes:a method of extracting deep CNN features of defect images by using convolution layer in CNN structure is proposed,and the CNN trained on ImageNet is migrated to feature extraction of defective image,and it can get more abundant semantic information and overcome traditional feature extraction methods,such as geometry,texture and gray;the sparse CNN features are obtained combining the deep CNN features with sparse coding,the sparse representation of high-dimensional features is realized by dictionary learning,sparse coefficient coding and maximum pooling;SVM is used to classify the sparse CNN features of defect images.The accuracy of the proposed algorithm is verified on the NEU surface defect image database.?3?Aiming at the defect extraction problems such as complex background texture,uneven gray distribution and small defect,a new defect extraction algorithm based on texture features and low rank decomposition is proposed to segment the weak defect target in complex background.The main innovations of this method include:a prior information extraction method for defective image is proposed,which constructs a texture prior image by calculating the texture prior of the input image,and the higher the pixels in the prior image,the higher the probability of becoming defective part;a weighted low rank decomposition W-LRR model is proposed,and the W-LRR model is constructed by texture prior map;the adaptive parameter updating method is used to solve the W-LRR model,and finally the defect target detection is realized.Experiments on simulated defect images and real textile defect images verify the effectiveness of the proposed method.?4?In order to solve the problem of calculating the crack length feature,an image measurement algorithm based on tree structure model is proposed.The main innovations of the algorithm include:a tree structure model is designed to model the crack skeleton,and the pixels on the crack skeleton are classified by analyzing the neighborhood relationship of each pixel,and the pixels are divided into three categories:trunk point,branch point and endpoint;the trunk of the crack skeleton is extracted by a proposed pruning algorithm,and the crack length is calculated by traversing.The feasibility of the proposed algorithm is verified by comparing the results of the image measurement algorithm with the manual measurement results.?5?Aiming at the practical application of surface defect detection,a reliability evaluation method of surface defect detection based on image processing is proposed.The innovation of this method mainly includes:establish a reliability evaluation framework based on image detection method by combining image measurement technology with reliability evaluation method organically;using Wiener process to model the degradation of image measurement results;using Markov chain to estimate the degradation model parameters,and realize reliability evaluation of industrial products.The fatigue test of a group of plastic plate samples verifies the validity and feasibility of the method.