Yarn-Dyed Fabric Defect Detection based on Convolutional Neural Network

Yarn-Dyed fabric defect detection is an important part of the textile production process, in which rapid and accurate detection is the main challenge in textile industry. However, the performance of defect detection largely depends on whether the manually designed features can properly represent the features of the defects. In this paper, a new detection algorithm for automatic fabric defect detection using the deep convolutional neural network (CNN) is put forward. Our defect detection algorithm is based on three main steps. In the first step, a preprocessing stage decomposes the fabric image into local patches and labels each local patch accordingly. In the second step, labeled fabric samples are transmitted to deep CNN for pre-training. Finally, defects are detected during image inspection that trained classifier slides over the entire fabric image and returns the category and position of each local patches to achieve defect detection. The proposed method was validated on two public and one self-made fabric databases. By comparing manually designed image processing solutions with other deep CNN networks for feature extraction methods, the experiments show that the proposed method can inspect defects at a higher accuracy compared with some existing methods.