A global-local frequency domain double-flow cooperation-based zero sample industrial anomaly detection method and system

By employing a global-local frequency domain dual-stream collaborative method, combining global frequency domain auxiliary branches and local feature enhancers, the problems of complex background interference and insufficient detection of local subtle anomalies in existing industrial anomaly detection technologies are solved, achieving high-precision zero-sample industrial anomaly detection.

CN122415519APending Publication Date: 2026-07-17CHONGQING UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHONGQING UNIV
Filing Date
2026-04-22
Publication Date
2026-07-17

AI Technical Summary

Technical Problem

Existing CLIP-based zero-shot industrial anomaly detection methods are susceptible to interference in complex industrial scenarios and struggle to effectively detect fine-grained texture disturbances and minute structural anomalies, especially in terms of insufficient pixel-level positioning accuracy and inadequate integration of frequency domain modeling with zero-shot vision-language frameworks.

Method used

A global-local frequency domain dual-stream collaborative approach is adopted. The global frequency domain auxiliary branch suppresses background noise, and the local feature enhancer amplifies local anomalies. By combining frequency domain features with a visual-language model, cross-modal feature fusion is achieved.

Benefits of technology

It improves the robustness and stability of the model in complex backgrounds, enhances the ability to perceive subtle local distortions, and achieves high-precision identification and pixel-level localization of unknown defects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122415519A_ABST
    Figure CN122415519A_ABST
Patent Text Reader

Abstract

The present application relates to a kind of global-local frequency domain double flow cooperation-based zero sample industrial anomaly detection method and system, belong to computer vision and industrial intelligent manufacturing technical field.The present application embeds global frequency domain auxiliary branch and block-level local feature enhancer in visual feature extraction in view of the problems that existing zero sample method excessively relies on spatial domain macroscopic semantics, is easily interfered by complex background and is insufficient for subtle local defect perception.Global branch suppresses periodic background noise by frequency transformation and multi-scale modulation, and constructs pure global feature reference;Local branch uses the frequency band energy statistics prior of image block, and targets amplifies small high-frequency anomaly by frequency gate gating and variance residual.The two cooperates to realize global background suppression and local anomaly amplification, effectively decouples normal texture and abnormal defect.The method significantly improves the detection robustness of unknown defects under zero sample condition without changing the weight of pre-trained model.
Need to check novelty before this filing date? Find Prior Art