Self-Supervised Surface Defect Localization via Joint De-Anomaly Reconstruction and Saliency-Guided Segmentation

Abstract

Anomaly localization plays one of significant roles in practical industrial applications, but still faces some unique challenges due to the rarity and diversity of anomalies.To address this issue, we propose a self-supervised surface defect localization method via Joint De-anomaly Reconstruction and Saliency-guided Segmentation, also named by JDRSS. Considering the severe lack of anomalous images, an approach for synthesizing anomalous samples is proposed to simulate different types of anomalies, which results in a variety of realistic and diverse anomalous images. To promote the reconstruction quality to the normal reference, we develop a novel auto-encoder based reconstruction network by applying the bundled de-anomaly constraint to refrain both the embedding space and the reconstruction space from the interference of abnormality. Furthermore, in order to accurately localize the surface defect, an anomalous saliency map guided segmentation network is proposed, in which the residuals from the reconstructed image and input are dexterously injected into each segmentation layer as spatial attention, thus enhancing the sensitivity to anomalies. We have conducted extensive experiments on the MVTec anomaly detection dataset and the proposed model achieves considerable performance for different classes of anomaly localization.