Research On Fabric Defect Detection Based On Deep Learning

In the process of fabric production,problems such as broken warp,stains and drags can not be avoided due to machine equipment failure and human factors,so it is necessary to detect the fabric.Because the current fabric defect detection algorithm still has the phenomenon of high false detection rate and low detection rate,it is necessary to design a set of fabric defect detection system with high detection rate,low error detection rate and high real-time performance.This paper takes normal types of fabric samples and common fabric defect types as research objects.In this paper,the SRC-CNN fabric defect detection algorithm based on deep feature fusion and the deep convolution neural network defect detection algorithm based on Fisher criterion are proposed.The main research work and achievements are summarized as follows:(1)SRC-CNN fabric defect detection algorithm based on deep feature fusion.In the case of limited number of labeled fabric defects,first of all,the AlexNet model is improved by transfer learning theory,and a 13-layer convolution neural network structure was designed.Secondly,the deep features of fabric defects are extracted from the convolution neural network model based on transfer learning,and the deep multi-feature fusion is carried out.Finally,the sparse representation classification framework was applied to the deep multi-feature fusion for classification.The experimental results show that the classification rate was 95.34% and the false detection rate was 4.95% on the limited black fabric database.The classification rate on the limited pink grid fabric database was 94.56% and the false detection rate was 5.76%.Compared with the popular network model and the traditional algorithm,the classification rate is improved and the false detection rate is decreased.(2)Deep convolution neural network defect detection algorithm based on Fisher criterion.In order to solve the problem of high complexity and long running time of the convolution neural network model,the depth-wise separable convolution was introduced to improve the Vgg16 model,and a small deep convolution neural network(DCNN)was designed.In addition,in order to further improve the fabric defect classification accuracy,the constrained Fisher criterion was added to the softmax layer of the DCNN network model,which reduces the intra-class distance and increases the inter-class distance,and then updates the whole network parameters continuously.The experimental results show that the network parameters,running time andclassification accuracy were 1.8 million,1375 seconds and 97.89% on the TILDA database,respectively.On the pink plaid fabric database,the network parameters,running time and classification accuracy were 1.8 million,1737 seconds and 98.16%,respectively.The algorithm can greatly reduce the network parameters,reduce the running time and improve the classification accuracy.Based on the theory of deep learning algorithms,this paper studies the method of fabric defect detection.On the one hand,it helps to improve the quality of fabric,on the other hand,it strengthens the competitiveness of textile export,which is of great significance to society and economy.