Research On Woven Fabric Defect Detection Based On Generative Adversarial Network

Fabric defect detection is an important part in the process of textile production quality monitoring.The traditional defect detection method is human visual inspection,which has the disadvantages of low accuracy and high labor intensity.With the development of computer vision technology,the research of using image processing technology to detect fabric defects has emerged.However,the traditional image processing technology relies on artificially designed feature extractor,which is difficult to adapt to the complex and changeable fabric weave and defect types.In recent years,the rapid development of deep learning technology has brought a new way for fabric defect detection.However,general deep learning relies on a large number of data for training,especially the labeled data,which is lacking in the field of fabric defect detection.Therefore,the method that does not rely on the defected fabric image for training has attracted the attention of researchers.This kind of method usually obtains a neural network model that learns the expression of defect-free fabric texture characteristics through deep learning training,and reconstructs the defected fabric image into a defect-free fabric image through this model,and determines the location of defects by the difference between the two images.However,the existing methods generally have the problem of insufficient ability to distinguish the texture features and defect features in the fabric image,which affects the effect of defect detection.In order to solve these problems,this paper proposes a new generative adversarial network model for fabric image reconstruction,which adds MLP layers in the network to reduce the rank of the features extracted by the network,so as to extract the texture features of the fabric image more accurately,and then improves the defect region segmentation algorithm based on the reconstructed image to achieve better detection effect of fabric defects.This paper mainly does the following work:(1)Research on a new generative adversarial network model.Based on the architecture of convolutional generative adversarial network,a new network model is proposed,in which a MLP layer is inserted into the convolutional selfencoder network,and the self-encoder structure network is used as the generator of the generative adversarial network and trained together with the discriminator.The trained generator can reconstruct the defect-free fabric image as it is,and can reconstruct the defected fabric image into the defect-free fabric image with the same texture.In the experiment of AITEX dataset,it can achieve good reconstruction results for defect-free fabric images,and the PSNR of three kinds of fabrics are 28.016,26.385 and 29.626,and the SSIM are 0.814,0.743 and 0.858.Both of which meet the detection requirements and are better than previous studies.(2)Research on the algorithm of defect region segmentation.On the basis of reconstructing the defected fabric image into the defect-free image,the segmentation algorithm is improved.The pixel-level segmentation of the defect region in the fabric image is realized by a three-step algorithm,which contains gray adjustment of the reconstructed image,saliency map generation and threshold segmentation.In the AITEX dataset,the defect-free fabric images have a lower false detection rate,and the F1 scores of the segmentation results of defective fabrics compared with the labels are 0.76,0.77 and 0.75,respectively,which are better than the results of previous studies.Experiments on multiple datasets show that the proposed method is effective in more types of fabric weave and defect images.(3)Analysis of relevant influencing factors of defect detection effectThrough the analysis of the experimental results,the factors related to the image that affect the detection effect are found.The influence of the gray intensity and size of the defect area in the image on the detection is experimentally verified and theoretically analyzed,and it is found that the reason for the negative influence on the reconstruction when the difference between the gray intensity of the defect area and the gray level of the fabric texture is too large is limited by the pixel-by-pixel error.Through the comparison of different illumination conditions when shooting images,it is concluded that the illumination brightness should be increased as much as possible on the premise of ensuring the integrity of the texture in the image.