Fabric Defect Detection with Cartoon–Texture Decomposition

Abstract

Automatic fabric defect detection plays an important role in textile industry. Most existing works utilize machine leaning methods to classify the fabric images with defects, however, because fabric defects are generally diverse and obscure. It is difficult to precisely identify the defects by direct image classifications. Aiming to tackle this problem, in this paper, we propose a two-stage method for automatic fabric defect detection. First, we utilize cartoon–texture decomposition to extract the features of textile structures from fabric images. Second, based on the features of cartoon textures, we build up a classifier with Deep Convolutional Neural Networks (DCNN) to distinguish the image regions containing defects, i.e., the regions of abnormal feature representation. Experimental results validate that the proposed method can precisely recognize the fabric defects and achieve good performances on the fabric images of various kinds of textiles.