一种图像多标签分类方法和装置
By optimizing the class embedding matrix using the divergence regularization function of label relations in federated learning, the slow convergence speed and privacy leakage issues of multi-label datasets are solved, achieving more efficient multi-label classification and secure model training.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JD DIGITS HAIYI INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2022-07-27
- Publication Date
- 2026-07-17
AI Technical Summary
Existing federated learning algorithms suffer from slow convergence, low model accuracy, and privacy risks when processing multi-label datasets. In particular, the class embedding matrix collapse problem is severe in federated learning with only positive labels, which affects model performance.
By optimizing the class embedding matrix using a divergence regularization function with label relationships on the server side, the relationships between labels are mined. A fixed class embedding matrix is pre-learned, and the divergence regularization function with label relationships is applied during training to optimize the class embedding matrix, reducing communication volume and improving model adaptability and safety.
It improves the convergence speed and model accuracy of multi-label datasets, reduces the risk of privacy leaks, reduces communication costs, and enhances the security and efficiency of federated learning.
Smart Images
Figure CN115546530B_ABST