基于类别感知边界优化的知识蒸馏方法、设备及存储介质
By optimizing the knowledge distillation method for category-aware boundaries, the classification and discrimination capabilities of student networks are improved, solving the problem of insufficient category boundary differentiation in existing technologies and achieving higher fine-grained image classification accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES)
- Filing Date
- 2026-05-21
- Publication Date
- 2026-07-17
AI Technical Summary
Existing decoupled knowledge distillation methods do not optimize category boundaries, making it difficult for student networks to effectively distinguish the boundaries of different categories during the knowledge distillation process. This results in poor classification ability, especially in fine-grained image classification tasks where it is difficult to obtain accurate image classification results.
A knowledge distillation network is constructed. By calculating the feature cosine similarity and distribution difference coefficient of the teacher and student networks, the category difficulty coefficient and the teacher-student distribution difference coefficient are introduced. Combined with fully connected layers and batch normalization layers, the category-aware boundary loss is optimized, and the total network loss of the training set and validation set is optimized to improve the classification and discrimination ability of the student network.
By optimizing the category-aware boundary, the student network model demonstrates stronger classification and discrimination capabilities in fine-grained image classification tasks, enabling it to more accurately identify similar categories and improve the accuracy of image classification.
Smart Images
Figure CN122242645B_ABST