紧凑原型与异常抑制的工业异常检测方法、设备及介质
By combining differential complementary adapters and sparse mask decoders with cross-attention mechanisms, the problems of feature redundancy and anomaly leakage in industrial anomaly detection are solved, achieving efficient and accurate anomaly detection and localization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- EAST CHINA JIAOTONG UNIVERSITY
- Filing Date
- 2026-06-18
- Publication Date
- 2026-07-17
AI Technical Summary
Existing industrial anomaly detection technologies suffer from problems such as redundant and isolated pre-trained features, huge memory consumption and difficulty in alignment of memory-based methods, and easy leakage of anomaly information during network reconstruction.
We employ a compact prototype and anomaly suppression approach, using a differential complementary adapter for feature enhancement and attention aggregation, leveraging a cross-attention mechanism to extract sparse normal prototype features, and using a sparse mask decoder to suppress the transmission of anomalous information. By combining feature consistency, label diversity, and channel diversity losses, we achieve efficient anomaly detection.
It achieves cleaner feature extraction, reduces memory consumption, improves the accuracy of anomaly detection and pixel-level localization capability, and significantly enhances detection performance in various and complex industrial scenarios.
Smart Images
Figure CN122413263A_ABST