Jump connection generative adversarial network for fabric defect detection

Abstract Among the fabric defect detection methods, unsupervised methods are based on the principle of training a network to restore a fabric image with defects to a flawless image with a consistent background and no visible defects, and to obtain specific information about the defects by comparing the two images for defect detection. However, most of the restored images cannot remove the defective part completely, and the more obvious defects are still visible. To solve the above problem, this paper proposes the jump connection generative adversarial network for fabric defect detection (JCGAN). JCGAN uses a jump connection structure to improve the ability of the network to extract detail information by introducing the detail information from downsampling into the upsampling process. It introduces a low-dimensional loss function to control the training of the network to improve the quality of the generated images. It uses two detection algorithms (SSIM computational detection algorithm, multi-channel defect detection algorithm) to compute the disparity grayscale information of the defected and recovered images separately, and finally fuses the information from both sides to obtain more detailed detection results. Compared with the six commonly used methods, the f-value of JCGAN is improved by 13.40\% on average compared with other methods.