A deep learning small target recognition positioning method for a patrol inspection robot
By optimizing feature extraction through multi-scale dilated convolution and attention mechanisms, and combining bidirectional cross-layer fusion and adaptive anchor box generation, the accuracy and precision issues of small target recognition and localization in inspection robots are solved, achieving efficient background suppression and pixel-level localization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANENG SHANTOU HAIMEN POWER GENERATION CO LTD
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-17
AI Technical Summary
Existing deep learning algorithms suffer from insufficient feature extraction, inadequate multi-scale feature fusion, and weak background interference suppression when identifying and locating small targets in inspection robots. This results in low recognition and positioning accuracy, as well as high false negative and false positive rates.
We employ multi-scale dilated convolution combined with channel attention and spatial attention mechanisms, and optimize feature extraction and localization regression through a bidirectional cross-layer fusion strategy and adaptive weighting coefficients, combined with adaptive anchor box generation and dynamic weighted loss function.
It improves the feature extraction capability of small targets in inspection, reduces redundancy in multi-scale feature fusion, effectively suppresses background interference, achieves pixel-level positioning accuracy improvement, and significantly reduces the false detection rate and missed detection rate.
Smart Images
Figure CN122412904A_ABST