Learning Multi-view Multi-class Anomaly Detection

TL;DR

This work tackles unsupervised multi-view, multi-class anomaly detection in industrial settings by unifying across views and classes. It introduces an encoder–decoder framework with a semi-Frozen Encoder, Anomaly Amplification Module, and Cross-Feature Loss to leverage cross-view correlations and reveal anomalies at multiple semantic levels. Empirical results on Real-IAD demonstrate state-of-the-art performance for both image-level detection and pixel-level localization, with notable gains over prior methods. The proposed approach enables robust, scalable anomaly detection across diverse industrial objects, offering practical benefits for surveillance and quality control, while highlighting areas for efficiency improvements in future work.

Abstract

The latest trend in anomaly detection is to train a unified model instead of training a separate model for each category. However, existing multi-class anomaly detection (MCAD) models perform poorly in multi-view scenarios because they often fail to effectively model the relationships and complementary information among different views. In this paper, we introduce a Multi-View Multi-Class Anomaly Detection model (MVMCAD), which integrates information from multiple views to accurately identify anomalies. Specifically, we propose a semi-frozen encoder, where a pre-encoder prior enhancement mechanism is added before the frozen encoder, enabling stable cross-view feature modeling and efficient adaptation for improved anomaly detection. Furthermore, we propose an Anomaly Amplification Module (AAM) that models global token interactions and suppresses normal regions to enhance anomaly signals, leading to improved detection performance in multi-view settings. Finally, we propose a Cross-Feature Loss that aligns shallow encoder features with deep decoder features and vice versa, enhancing the model's sensitivity to anomalies at different semantic levels under multi-view scenarios. Extensive experiments on the Real-IAD dataset for multi-view multi-class anomaly detection validate the effectiveness of our approach, achieving state-of-the-art performance of 91.0/88.6/82.1 and 99.1/43.9/48.2/95.2 for image-level and the pixel-level, respectively.

Figures (4)

• Figure 1: Task settings. (a) Task setting of Class-separate, Single-view Unsupervised Anomaly Detection. (b) Task setting of Class-separate, Multi-view Unsupervised Anomaly Detection. (c) Task setting of Multi-class, Multi-view Unsupervised Anomaly Detection.
• Figure 2: Challenge of Multi-View Scenarios. (A) In the phone battery category, the scratch is exclusively visible in (A5) and cannot be observed in views (A1-A4). (B) In the Audiojack category, the contamination in is difficult to detect in view B1, yet it becomes clearly visible in view (B5), highlighting the necessity for the model to exploit inter-view correlations to accurately localize the anomaly in (B1). (C) The third row illustrates that edge-located anomalies in the u-block category tend to be challenging to identify.
• Figure 3: Framework of our method that contains three parts: (1) semi-frozen encoder SFE; (2) anomaly amplification module AAM; (3) cross-feature loss CFL. During the training step, the input $x_0$ is put into SFE and AAM to get intermediate features $f_m$. Then the feature $f_m$ is passed through the decoder, and its deep and shallow features are exchanged to compute CFL with features from SFE. During the testing step, $x_0$ is put into the same network to compute the anomaly score.
• Figure 4: Visualization of Anomaly Map on Real-IAD. All samples are randomly chosen. The anomaly map shows that our method is capable of precisely localizing anomalies across diverse object types, even when the defects are small or subtle.