Since the general representations produced by pre-trained feature extractors are often insensitive to intra-class variations, existing anomaly detection methods that store them directly in memory banks are constrained in performance. In addition, the distribution discrepancy among different modalities may cause cross-modal interference, further weakening the ability to discriminate anomalies. To address these issues, we propose an Adaptive Prototype Guidance Network (APG-Net) for multi-sensor anomaly detection. First, to avoid cross-modal feature interference, we construct independent anomaly detection branches for multi-sensor data including RGB images, point clouds and infrared images. Then, we introduce a non-parametric feature space reshaping paradigm for each modality. This paradigm adds no additional trainable parameters, ensuring efficiency and ease of deployment. It first learns guiding prototypes from normal samples and matches them to the general representations produced by pretrained extractors. Subsequently, a two-stage adaptive prototype guidance strategy is applied to reshape the feature distributions. This strategy enlarges the separation between normal and anomalous features in the feature space. Finally, we perform decision-level fusion to integrate the anomaly detection strengths from all sensor data. Extensive experiments demonstrate that our method achieves an object-AUROC of 97.4% on the MulSen-AD multi-sensor anomaly detection benchmark, surpassing previous state-of-the-art approaches.

Anomaly detection is a key technology in quality control for automated production lines. Currently, 2D-based anomaly detection methods fail to identify geometric structure anomalies in products. To address this limitation, this paper proposes a multimodal anomaly detection model using 3D point clouds and RGB images. To ensure the single-domain inference capability of each modality, we design an attention-enhanced dual memory bank to separately store local point cloud features and RGB features. The attention mechanism enhances the informativeness and discriminability of the feature descriptors, significantly improving the data quality in the memory bank. During the inference phase, the local point cloud features in the dual memory bank guide the RGB features in calculating anomaly scores in the 2D modality. This memory-guided approach strengthens the correlation between information across different modalities. Moreover, to improve the overall segmentation precision of the model, we propose an anomaly scoring scheme based on a weight map of signed distance values. The final anomaly detection results are obtained by integrating the advantages of point cloud data in geometric structure anomaly detection and RGB data in color anomaly detection. Extensive experiments demonstrate that the proposed method achieves superior segmentation precision compared to other advanced methods on the MVTec 3D-AD and Eyecandies datasets.