Fabric Defect Detection Method Based On GAN And Semantic Segmentation

Fabric defect detection is a crucial aspect of ensuring quality control in textile production.It enables timely identification of defects,helps to control the quality of fabric production from the outset,and effectively minimizes losses for enterprises.Due to the wide variety and diverse shapes of fabric defects,traditional detection methods are gradually inadequate to meet the current needs of textile product inspection.In recent years,deep learning algorithms have become the research trend and the primary choice in fabric defect detection due to their advantages in speed,accuracy,and robustness.However,due to the complex and varied background texture of fabric images and the diverse types of fabric defects,traditional convolutional neural networks are often unable to balance global information and defect feature information resulting in the poor defect detection performance.When meeting the small sample size,it is difficult for network models to effectively learn the features of defects.In this paper,related research and experiments have been carried out to solve these problems,and the main related research work includes the following aspects:(1)The proposed method in this study is a new fabric defect detection approach based on an improved U-Net network.However,due to the poor segmentation performance of the traditional U-Net network when detecting fabric images with complex backgrounds,we propose improved U-Net network to achieve better fabric defect detection.Firstly,the MAC attention module is introduced to enhance the effective feature extraction of the model and suppress the influence of the background on feature extraction.Secondly,to cope with the problems of gradient explosion and network degradation due to the increase in network layers and parameters,a residual learning module is introduced in the process of up-sampling and under-sampling to improve the residual mapping.Experimental results show that on a simple fabric image texture dataset,the proposed method achieves an average accuracy of 94.2%,an average recall of 93.9%,and an average precision of 91.1%;for complex fabric image texture datasets,the average accuracy is 84.9%,the average recall is 78.4%,and the average precision is 73.1%.The proposed method achieves good fabric defect detection results.(2)A fabric defect detection method based on improved conditional GAN is proposed.In view of the fact that the original condition GAN may not be able to pay attention to the defect characteristics of the fabric image and control the global information of the fabric image,the network is improved.Firstly,the MFE feature extraction module is introduced into conditional GAN to allow the network to fully focus on fabric defect features while utilizing the global information of the fabric image to reduce computational complexity.Secondly,in terms of the loss function,a mean square error-based loss is combined during training to allow the generator G to better generate images.Finally,a comparative experiment is carried out on the proposed method.The experimental results demonstrate that the proposed method achieves good results in both SSIM and FID evaluation indicators and is capable of generating fabric images well.When the produced images are added to the training set,it can enhance the model’s ability to detect fabric defects.