Cumulative dual-branch network framework for long-tailed multi-class classification

Abstract

The long-tailed data distribution problem (i.e., a few classes account for majority data, while most classes account for minority data) is widespread in large-scale and real-world datasets, and it poses a huge challenge to the computer vision field. Existing methods of long-tailed classification mainly focus on re-sampling, re-weighting, and transfer learning. Although class imbalance learning can yield better long-tailed classification performance, the feature representative ability of the feature extraction network is damaged to a certain extent. To deal with these issues, the present work proposes a novel cumulative dual-branch network framework (CDBNF), which takes into account the class imbalance learning and feature representation learning at the same time by the dual-branch network architecture. In CDBNF, the class imbalance learning branch greatly improves the classification performance of tail classes, while the few-shot learning branch enhances the feature representative ability. Furthermore, a cumulative learning strategy (CLS) is proposed in CDBNF to make it pay more attention to the tail classes gradually in the training process. The effectiveness and practicability of the proposed CDBNF are verified by the four benchmark datasets. Experimental results show that the classification performance of CDBNF is superior to other state-of-the-art methods, while the intra-class feature variance is smaller than other state-of-the-art methods.