Knowledge Distillation for Unsupervised Defect Detection of Yarn-Dyed Fabric Using the System DAERD: Dual Attention Embedded Reconstruction Distillation

Detecting defects of yarn-dyed fabrics automatically in industrial scenarios can improve economic efficiency, but the scarcity of defect samples makes the task more challenging in the customised and small-batch production scenario. At present, most reconstruction-based methods have high requirements on the effect of reconstructing the defect area into the normal area, and the reconstruction performance often determines the final defect detection result. To solve this problem, this article proposes an unsupervised learning framework of dual attention embedded reconstruction distillation. We try to use this novel distillation scheme to provide some contribution to the defect detection field. Firstly, different from the encoder-encoder structure of traditional distillation, the teacher-student network in this article adopts the encoder-decoder structure. The purpose of the student network is to restore the normal feature representation of the pre-trained teacher network. Secondly, this article proposes a dual attention residual module, which can effectively remove redundant information and defective feature information from the teacher network through the double feature weight allocation mechanism. This helps the student network to recover the normal feature information output by the teacher network. Finally, the multi-level training deployment at the feature level in this article aims to make the model obtain accurate defect detection results. The proposed method has been extensively tested on the published fabric dataset YDFID-1, ZJU-Leaper dataset and the anomaly detection dataset MVTec. The results show that this method not only has good performance in fabric defect detection and location but also has universal applicability.