Open-Set Fabric Defect Detection With Defect Generation and Transfer

Fabric defect detection is indispensable but challenging due to the diversity of fabric texture and defect types in textile mills, and a variety of deep learning-based supervised methods have been introduced to defect detection and segmentation. However, these methods suffer from the scarcity of defective samples and are ineffective in dealing with unknown types of defects. In this study, a novel open-set fabric defect detection method is proposed. To alleviate the problem of scarcity of defective samples, a defect generation and transfer (DGT) module is first designed to generate a large number of defective samples with different defect shapes and types. Then, a defect detection module is constructed, which learns the discriminant representations from real and generated defective samples to identify and segment defects. Finally, an open-set defect classification (DFC) module is introduced, which fits the key features of each known defect type with the Weibull distribution and realizes the classification of known defect types as well as the rejection of unknown defect types. Extensive experiments conducted on five fabric datasets show that the proposed method outperforms other state-of-the-art methods, achieving an average detection performance improvement of 3.0% and 1.4% for image- and pixel-level area under the receiver operating characteristic curve (AUROC), respectively, and an average classification improvement of 8.2% and 17.2% for closed-set and open-set DFC, respectively.