UMLLA-AD: Mamba-Driven Adaptive Feature Selection for Industrial Anomaly Detection
Tingting Fang, Junjie Wang, Ming Ye, Yuefei Huang · 2025
In modern industrial environments, anomaly detection is a critical technology for ensuring production quality and safety. However, existing methods face the challenge of information redundancy when handling high-dimensional data, and traditional autoencoders struggle to effectively capture subtle anomaly features. To address these limitations, we propose UMLLA-AD, an efficient network for detecting and localizing anomalies. UMLLA-AD attains performance enhancement via two core innovations. Firstly, we have designed an Adaptive Feature Channel Selection (AFCS) mechanism that significantly reduces noise and information redundancy by dynamically screening feature channels containing critical anomaly information. Secondly, we have developed the UNet-based Mamba-like linear attention mechanism (UMLLA) reconstruction network, which combines the global modeling capability of the Mamba architecture with a multi-head linear attention mechanism. This integration enhances the sensitivity to subtle anomaly features and improves reconstruction accuracy within the UNet framework. We conducted extensive experiments on several industrial datasets, including MVTec-AD, BTAD, MPDD, and VisA. The UMLLA-AD framework showed significant competitive advantages in all datasets tested. Particularly on the MVTec-AD dataset, our approach achieved an Image AUROC of 99.7% and a Pixel AUROC of 99.1%. These results clearly demonstrate that this research provides an efficient and robust solution for industrial anomaly detection.