ReMaskNet: Regenerate Mask Network for Industrial Anomaly Detection and Segmentation

We propose ReMaskNet, an innovative unsupervised anomaly detection network designed to address the challenges of defect detection and localization. ReMaskNet leverages the strengths of three major methods: reconstruction-based, synthesis-based, and embedding-based methods. Specifically, it first employs a feature extractor to extract multi-scale features from input images, which are then adapted to the target domain through feature adapters. Following this, a dual-level anomaly synthesis strategy, incorporating both image-level and feature-level anomaly synthesis, is applied to enhance the model’s ability to learn from synthesized abnormal samples. Subsequently, a feature reconstruction network is employed to reconstruct the features. By computing the difference between the reconstructed and original features, an initial anomaly segmentation map is generated. To refine this segmentation, ReMaskNet introduces a feature inpainting module, where the generated anomaly segmentation map serves as a mask. Anomaly regions are masked out and subsequently restored using an inpainting network. Additionally, an iterative op-timization mechanism is employed to progressively enhance segmentation accuracy. Experimental results demonstrate that ReMaskNet outperforms state-of-the-art (SOTA) methods on the MVTec AD dataset, validating its effectiveness in handling complex backgrounds and subtle defect detection. Code is available at: https://github.com/gwb55/ReMaskNet.