Enhanced multi-branch learning for long-tailed image recognition

Due to the severe class imbalance between head classes and tail classes of long-tailed data, deep learning algorithms face significant challenges when dealing with long-tailed data distribution. The class rebalancing methods are generally considered to address class imbalance, however they disrupt the feature distribution in the feature space while improving the performance of tail classes. In this paper, Enhanced Multi-Branch Learning (EMBL), a novel visual recognition model, is designed for long-tailed data. EMBL not only effectively addresses the issue of class imbalance but also avoids the damage of feature distribution, and reduces training overhead. In EMBL, the data augmentation method called Oversampling-Based Hybrid CutMix and Mixup (OHCM) is designed to generate an image with rich semantic information to expand tail classes. In addition, a Dynamic Supervised Contrastive Learning (DSCL) is proposed. In DSCL, the temperature coefficient \(\tau\) $\tau$ is dynamically varied to allow for the adaptive learning of feature representation based on the training epoch and sample similarity. Finally, an information supplementary branch is introduced in addition to a class rebalancing branch and a conventional learning branch to construct a multi-branch learning framework. A linear decay fusion strategy is employed to perform weighted fusion for those branches. EMBL is validated on four datasets consisting of the CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT and Places-LT. Specially, EMBL achieves state-of-the-art accuracy on multiple datasets.