Detection of defect on woven fabric at multiple magnification levels by mitigating data unbalancing problem

Computer vision based defect detection system for woven fabric has become an essential tool in textile industries, due to its reliability and ability to produce defect free fabric of high quality. This article presents a novel approach to detect wide variety of defects on woven fabrics of different classes. To improve the performance of defect detection system, some saliencies have been introduced. First, an adaptive image enhancement technique is proposed to adjust the brightness and contrast of the image. Second, U-Net-based classifier, improvised to study the fabric image at multiple magnification level is used for pixel level detection of fabric defects. Finally to mitigate the problem of data unbalancing during training of U-Net based classifier inter class exchange of fabric defect is done by using class specific conditional generative adversarial network (cGAN) to generate desired fabric defects on desired fabric classes. This method is validated on 800 defective and 200 non defective fabric images from the synthesized dataset derived from TILDA, AITEX, and Fabrics datasets. Detection success rate of 96.1%, recall value of 95.75%, precision of 99.35%, and F-1 measure of 97.52% have been achieved during detection of defects by this method.