An Improved DCGAN for Fabric Defect Detection

Abstract

Textile defect detection is an important part of textile quality control. Due to the diversity of fabric texture and the lack of defect fabric images, detection methods based on deep learning which does not rely on defective fabric samples has been gradually applied. However, in previous methods, the ability to distinguish the image features of fabric texture and defects is insufficient. In order to solve this problem, this paper proposed an improved generative adversarial network, which introduced a self-encoder with MLP layers into the generator module. Fabric images with defects will be reconstructed into the ones without defects through the trained generator. Then some image processing operations will be carried out to compare the original defect image and the reconstructed image in order to obtain the segmentation of the defect area. By adding MLP layers to extract lower rank fabric image features, the developed model has a stronger ability to capture fabric texture features. Compared with previous studies, it can achieve a better segmentation effect of defects. Precision, recall rate and Fl-score are improved significantly in the experiments.