A machine vision-based industrial product surface defect detection method and system
By extracting frequency-space joint features through multi-scale Gaussian-Laplace pyramid decomposition and dual-tree complex wavelet transform, and combining them with graph attention networks, the problems of high false detection rate and resource-constrained equipment deployment in defect detection under complex texture backgrounds are solved, achieving efficient and accurate surface defect detection of industrial products.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN POLYTECHNIC
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-17
AI Technical Summary
Existing machine vision-based industrial product surface defect detection technologies struggle to effectively distinguish defect features from background textures, resulting in high false detection and false negative rates. Furthermore, large-scale models are difficult to deploy in real time on resource-constrained edge devices.
Multi-scale Gaussian-Laplace pyramid decomposition and dual-tree complex wavelet transform are used to extract frequency-space joint features. Morphological component analysis and graph attention network are used to construct a semantic topology graph of defects, so as to achieve accurate extraction and detection of defect candidate regions.
It improves the robustness and accuracy of defect detection, reduces computational complexity, is suitable for resource-constrained edge devices, and meets real-time detection requirements.
Smart Images
Figure CN122415544A_ABST