A space-frequency cooperative multi-scale feature fusion method and device, a defect detection network and equipment
By employing a multi-scale feature fusion method that combines spatial and frequency domains, this approach selects features in the frequency domain using Fourier transform and inverse transform, and combines spatial domain feature weights. This solves the problem of neglecting frequency domain features in existing technologies, and achieves more efficient multi-scale feature fusion and defect detection.
CN122156882APending Publication Date: 2026-06-05江西省通讯终端产业技术研究院有限公司 +1
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 江西省通讯终端产业技术研究院有限公司
- Filing Date
- 2026-04-16
- Publication Date
- 2026-06-05
AI Technical Summary
Technical Problem
Existing technologies neglect video domain features in image target detection, resulting in indiscriminate fusion of features at different levels and poor fusion results.
Method used
A multi-scale feature fusion method based on spatial-frequency coordination is adopted. Features are selected in the frequency domain through Fourier transform and inverse transform, and feature fusion is performed by combining spatial domain feature weights to optimize multi-scale feature fusion.
Benefits of technology
It improves the effect of multi-scale feature fusion and enhances the detection accuracy of defects of different sizes.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN122156882A_ABST
Abstract
The application discloses a space-frequency cooperative multi-scale feature fusion method and device, a defect detection network and equipment. First, the feature maps of adjacent levels are fused, and then spatial branch feature maps and weight branch feature maps are obtained by cutting along the channel dimension. Then, Fourier transform is performed on the spatial branch feature maps, and then layer-by-layer feature extraction is performed, and then inverse Fourier transform is performed to obtain spatial domain feature maps; meanwhile, feature weight extraction is performed on the weight branch feature maps to obtain weight feature maps. The spatial domain feature maps and the weight feature maps are multiplied point by point, and then local feature extraction is performed, and then the result is added to the fusion map to output the result. The application adopts a space domain and frequency domain linkage mode, selects corresponding information fusion in the frequency domain with the space important information, and effectively improves the multi-scale feature fusion effect of the target detection method based on deep learning through the frequency domain feature optimization and the adaptive feature selection mechanism in the space domain.
Need to check novelty before this filing date? Find Prior Art