Abstract
Recently,intensive attempts have been made to design robust models for fine-grained visual recognition,most notably are the impressive gains for training with noisy labels by incorporating a reweighting strategy into a meta-learning framework.However,it is limited to up or downweighting the contribution of an instance for label reweighting approaches in the learning process.To solve this issue,a novel noise-tolerant method with auxiliary web data is proposed.Specifically,first,the associations made from embeddings of well-labeled data with those of web data and back at the same class are measured.Next,its association probability is employed as a weighting fusion strategy into angular margin-based loss,which makes the trained model robust to noisy datasets.To reduce the influence of the gap between the well-labeled and noisy web data,a bridge schema is proposed via the corresponding loss that encourages the learned embeddings to be coherent.Lastly,the formulation is encapsulated into the meta-learning framework,which can reduce the overfitting of models and learn the network parameters to be noise-tolerant.Extensive experiments are performed on benchmark datasets,and the results clearly show the superiority of the proposed method over existing state-of-the-art approaches.
基金项目
National Natural Science Fund for Distinguished Young Scholars of China(62125203)
National Natural Science Foundation of China(62302233)
National Natural Science Foundation of China(62276143)
Key Program of the National Natural Science Foundation of China(61932013)
Natural Science Foundation of Jiangsu Province of China(BK20180470)
Natural Science Foundation of Jiangsu Province of China(BK20200739)
Postdoctoral Science Foundation of Jiangsu Province of China(2021K172B)
China Postdoctoral Science Foundation(2021M691655)
NETL
NSTL
万方数据