A Study Of Computer-Vision-Based Fabric Defect Detection System

Fabric defect detection is an important part in the process of textile production.Fabric defect is an important factor affecting fabric quality.At present,fabric defects only rely on simple tools to complete,this traditional detection method completely depends on human experience,recognition efficiency and accuracy has been gradually unable to meet the needs of enterprise automation and intelligent production.With the continuous progress of computer technology,machine vision and image processing technology,in order to improve the labor production environment,improve production efficiency and reduce production costs,it is an inevitable trend to adopt automatic defect identification technology in the defect detection stage of fabric industry production.From the perspective of texture analysis,this paper uses the difference between the surface defect texture and the background texture to detect fabric defects.Firstly,this paper describes and analyzes the common defects in fabrics,and divides them into four categories.This paper analyzes and compares several commonly used image preprocessing methods,such as spatial filtering,bilateral filtering,guided filtering,L0-norm filtering and gray value morphology.Through the peak signal-to-noise ratio(PSNR)and running time as its performance indicators,the appropriate preprocessing methods are quantitatively evaluated,Experiments show that this method can suppress the interference information of the fabric surface and retain the key texture information at the same time.Then,the texture characteristics of the fabric surface are analyzed,and the sub images are divided according to the size of the texture cycle unit.The texture features of the sub images are extracted by Gray level co-occurrence matrix(GLCM)and Contour transform(CT).Based on the above two texture feature extraction methods,a new feature extraction method based on Redundancy gray level co-occurrence matrix(R-GLCM)is proposed.In this method,images are decomposed at different scales to form image redundancy,and then features of gray cooccurrence matrix are extracted from redundant images.The following experiments show that the three methods are very effective for the feature expression of image texture,but the method proposed in this paper has more feature quantity and better division.Finally,the comparison between linear discriminant analysis(LDA)and backpropagation(BP)neural network is made.The LDA Algorithm is used to reduce the dimension of features extracted by different methods and get the recognition rate,so as to find the optimal feature combination under each feature extraction method;For BP neural network,the network structure is determined through several comparative experiments,and a multi-layer perceptron(MLP)with excellent classification performance is constructed.The features extracted by different methods are used to train and test the network off-line respectively.The experimental results show that the recognition rate of fabric defect detection method proposed in this paper reaches97.3%?...