Enhancing industrial anomaly detection with Mamba-inspired feature fusion

Abstract

Image anomaly detection is crucial in industrial applications, with significant research value and practical application potential. Despite recent advancements using image segmentation techniques, challenges remain in global feature extraction, computational complexity, and pixel-level anomaly localization. A scheme is designed to address the issues above. First, the Mamba concept is introduced to enhance global feature extraction while reducing computational complexity. This dual benefit optimizes performance in both aspects. Second, an effective feature fusion module is designed to integrate low-level information into high-level features, improving segmentation accuracy by enabling more precise decoding. The proposed model was evaluated on three datasets, including MVTec AD, BTAD, and AeBAD, demonstrating superior performance across different types of anomalies. Specifically, on the MVTec AD dataset, our method achieved an average AUROC of 99.1% for image-level anomalies and 98.1% for pixel-level anomalies, including a state-of-the-art (SOTA) result of 100% AUROC in the texture anomaly category. These results demonstrate the effectiveness of our method as a valuable reference for industrial image anomaly detection.