ComNet: A Complementary Prototypes-Guided Reconstruction Framework for Multi-Class Anomaly Detection

Reconstruction-based methods, while widely used for anomaly detection (AD), often suffer from the "identity shortcut" (IS) problem, in which anomalous regions are "perfectly" reconstructed, severely compromising detection reliability. Moreover, most methods remain vulnerable to contamination when deriving prototypes directly from input defective images. To address these limitations, we propose ComNet, a new complementary prototypes–guided reconstruction framework for multi-class AD. ComNet introduces a dual-path prototype guidance mechanism consisting of: 1) a Difference-driven Prototype Filter that adaptively identifies and excludes anomalous regions from input images to ensure clean and aligned prototypes extraction; and 2) a Class-Related Memory Bank that stores normal features and applies them to generate prototypes for restoring semantic information lost due to large defects. Furthermore, we design a distance-aware attention that modulates attention scores based on spatial proximity to alleviate IS. ComNet achieves state-of-the-art performance, yielding improvements of +16.1% p-AP on MVTec-AD, +13.1% p-AP on VisA, and +10.9% p-AP on Real-IAD, demonstrating ComNet’s strong effectiveness in multi-class AD.