Cross-domain semantic segmentation method based on multi-element knowledge distillation
By constructing a joint optimization framework that guides the student network through a teacher network, and utilizing feature enhancement modules and multi-level distillation strategies, the problem of performance degradation in high-definition domain testing during cross-domain semantic segmentation was solved, achieving efficient cross-domain knowledge transfer and accurate segmentation of high-definition images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-20
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies in cross-domain semantic segmentation suffer from problems such as decreased segmentation performance in high-definition domain testing and poor segmentation of edge regions and small targets. Traditional methods cannot effectively utilize the knowledge trained on compressed images for high-definition domain transfer, and methods based on super-resolution preprocessing and adversarial learning have problems with high computational overhead and stability.
A joint optimization framework is constructed to guide the student network through a teacher network. By combining feature enhancement modules, feature distillation, logits distillation, and structural distillation with a curriculum-based compressed training strategy and a quality-aware adaptive distillation strategy, multi-level knowledge transfer is achieved, thereby improving the semantic segmentation accuracy and generalization ability of the student model on high-definition images.
It effectively improves the semantic segmentation accuracy and boundary quality of the model on high-definition raw images, reduces the dependence on high-definition labeled data, alleviates cross-domain distribution shift, and enhances the model's adaptability and segmentation performance.
Smart Images
Figure CN122416033A_ABST