Self-supervised and Template-Enhanced Unknown-Defect Detection

Tingting Li, Yaqiao Liao, Xu Wang, Guowen Kuang, Zhibin Chen, Jinfeng Yang

Abstract: Surface defect detection has great significance to the manufacturing industry. It is also a challenging task because of the large variation of unknown defects and limited available datasets. In this paper, we propose a self-supervised generative approach for unknow-defect detection. The training process only involves masked defect-free samples, and additional template information is used to achieve a more robust performance of GAN-based image reconstruction. We design a fusion module based on an attention mechanism, which aligns features of an image with mask or defect to template image for better image reconstruction and defect-localization. The reconstructed defect-free image with uncertainty heatmap is generated by a subsequent decoder. Our proposed method outperforms the baseline method in defect description on a bottle cap dataset from the real industrial process.