Enhanced Reverse Distillation Guided Segmentation Network for Anomaly Detection

Image anomaly detection holds significant potential across various domains. The scarcity of anomaly samples in real-world scenarios makes unsupervised anomaly detection more practical, as acquiring a substantial number of anomaly samples is often challenging. Despite the progress made by image reconstruction-based methods in this task, their reliance on fuzzy reconstruction images and the complexity of post-processing have posed challenges to the overall performance of anomaly detection. To address this issue, this paper proposes a novel Enhanced Reverse Distillation Guided Segmentation Network(ERDS-Net). Unlike conventional approaches, this network consolidates the image reconstruction and anomaly detection into a unified process, leveraging feature maps from the intermediate layers of both the encoder and decoder for segmentation. In addition, our model fully leverages a batch-wise attention mechanism, focusing on more challenging samples. This method aims to enhance overall performance by simplifying the process and mitigating the adverse effects of blurriness and complex post-processing. Moreover, this paper utilizes a self-supervised approach for training in order to extract additional information from unlabeled data. Experimental results demonstrate the outstanding effectiveness of the proposed Enhanced Reverse Distillation Guided Segmentation Network on the MVTec dataset. In comparison to traditional image reconstruction-based methods, the new approach exhibits superior performance in anomaly detection and localization tasks, demonstrating enhanced robustness with various metrics.