Research On Defect Detection Of Glossy Fabric Based On Generative Adversarial Network

China is the world's largest textile exporter.In order to improve the quality and brand value of textiles,defect detection has become a very important part in the textile manufacturing process.Aiming at the problems of high gloss fabric and leather defect detection,such as high gloss,serious texture interference and difficult detection,this thesis focuses on the research of a series of new defect detection algorithms based on generative countermeasure network.The main work of this thesis includes the following aspects:1?Compared with the traditional defect detection algorithm,the fabric defect recognition algorithm based on deep learning has obvious advantages in feature extraction.But the main problem is the lack of large-scale data sets,and the recognition rate is too low in complex texture background.Therefore,how to extract fabric defect features based on small sample unpaired data set becomes a key problem.In this thesis,we first propose the idea of using GAN to enhance the data.The innovation of GAN is to generate the same pattern of defects instead of directly generating defect images,and then combine the images without defects to generate a large number of pairs of images,so as to train the convolution neural network to realize the fabric image defect feature extraction.At the same time,an attention module is added in the front of the generator to make the defect points have higher weight and make the generator pay more attention to its own tasks.Attention module is composed of feature extraction network Res Net and LSTM to prevent model collapse.Numerical experiments show that the image set generated by our method ensures the diversity of samples and improves the effectiveness of classification.2?Aiming at the problem that the specular reflection of high gloss fabric makes the camera full and affects the defect recognition,a style conversion network from specular reflection image to diffuse reflection image based on improved Cycle GAN is proposed.The network adds perceptual loss to the cyclegan network to enrich the texture details of the converted image.In the design of the generator,Dense Net plus U-Net is used to replace the original Res Net structure,so that the network parameters are relatively less to save time and cost.The research results can remove specular highlights from a new perspective,and provide a new solution to the problem of removing specular highlights in the three-dimensional measurement of objects with specular characteristics.3?Finally,the most suitable recognition algorithm for fabric defect detection is selected by comparing various mainstream target detection algorithms.Because the defect points are usually small and local,the feature map can not distinguish the background from the defect points.In this project,Faster R-CNN is used to fuse the high-dimensional and low dimensional features in the image,which helps the model to distinguish the background texture and defect points of the fabric.Experiments show that this method has some advantages in fabric defect detection.