一种基于知识蒸馏的轻量级遥感影像变化检测方法
By combining prototype comparative distillation and channel-space normalized distillation, the problem of identifying complex and irregular regions in remote sensing change detection is solved, and lightweight models are used for efficient and accurate identification of changes in remote sensing images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2022-11-08
- Publication Date
- 2026-07-17
AI Technical Summary
Existing remote sensing change detection methods struggle to accurately identify complex and irregular change regions in remote sensing images, and existing knowledge distillation methods fail to achieve ideal performance in remote sensing change detection, especially due to the ambiguity in edge pixel classification caused by intraclass diversity resulting from changes in imaging conditions at different times and the complex shapes of change regions.
A novel knowledge distillation method combining prototype-contrastive distillation and channel-space normalized distillation is adopted. By guiding the student model to imitate the discriminative feature distribution of the teacher model, the misclassification of confused objects is reduced. Furthermore, by focusing on spatial probability distribution and category probability distribution, the knowledge output by the teacher model is comprehensively learned.
The accuracy of the lightweight student model in remote sensing change detection has been improved, making its performance comparable to that of the large teacher model. It is better able to identify complex and irregularly shaped change areas and generate more accurate remote sensing image change detection results.
Smart Images
Figure CN115546196B_ABST