Unsupervised fabric defect detection based on a deep convolutional generative adversarial network

• Published in: Textile research journal Vol. 90; no. 3-4; pp. 247 - 270
• Main Authors: Hu, Guanghua, Huang, Junfeng, Wang, Qinghui, Li, Jingrong, Xu, Zhijia, Huang, Xingbiao
• Format: Journal Article
• Language: English
• Published: London, England SAGE Publications 01.02.2020; Sage Publications Ltd
• Subjects: Coders; Defects; Fabrics; Generative adversarial networks; Image processing; Image reconstruction; Image segmentation; Inspection; probability; Surface defects; textile fibers; texture; Training; generative adversarial network; fabric inspection; deep learning; defect detection
• ISSN: 0040-5175; 1746-7748; 1746-7748
• DOI: 10.1177/0040517519862880

Detecting and locating surface defects in textured materials is a crucial but challenging problem due to factors such as texture variations and lack of adequate defective samples prior to testing. In this paper we present a novel unsupervised method for automatically detecting defects in fabrics based on a deep convolutional generative adversarial network (DCGAN). The proposed method extends the standard DCGAN, which consists of a discriminator and a generator, by introducing a new encoder component. With the assistance of this encoder, our model can reconstruct a given query image such that no defects but only normal textures will be preserved in the reconstruction. Therefore, when subtracting the reconstruction from the original image, a residual map can be created to highlight potential defective regions. Besides, our model generates a likelihood map for the image under inspection where each pixel value indicates the probability of occurrence of defects at that location. The residual map and the likelihood map are then synthesized together to form an enhanced fusion map. Typically, the fusion map exhibits uniform gray levels over defect-free regions but distinct deviations over defective areas, which can be further thresholded to produce a binarized segmentation result. Our model can be unsupervisedly trained by feeding with a set of small-sized image patches picked from a few defect-free examples. The training is divided into several successively performed stages, each under an individual training strategy. The performance of the proposed method has been extensively evaluated by a variety of real fabric samples. The experimental results in comparison with other methods demonstrate its effectiveness in fabric defect detection.