一种基于噪声标签筛选的图像分类方法和系统
By employing adaptive prompt word learning and heterogeneous collaborative screening strategies, the accuracy and reliability issues of image classification under noisy labels are addressed, improving the robustness of the model and data diversity, and achieving efficient noisy label screening and image classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU OPEN UNIVERSITY (THE CITY VOCATIONAL COLLEGE OF JIANGSU)
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies for image classification with noise labels suffer from problems such as inaccurate noise label recognition and degraded model performance. In particular, the accuracy of selection is insufficient under high noise rates, and the reliance on manual labor is costly. Furthermore, existing methods lack data diversity and scalability.
We employ an adaptive cue word learning and heterogeneous collaborative screening strategy. By constructing fixed and adaptive cue words, we combine visual language models and DNN models of different sizes for semi-supervised training to gradually screen out clean samples with high confidence, thereby alleviating distribution bias and improving data diversity.
It significantly improves the robustness and classification performance of image classification models in complex noisy environments, achieving more efficient, robust and reliable noise label recognition, and improving the accuracy of image classification.
Smart Images
Figure CN122416104A_ABST