A Triplet Branch End-To-End Framework for Long-Tailed Object Recognition in Remote Sensing Images

In the real world scenarios, the classes of many remote sensing images typically exhibit an imbalanced distribution, where only some classes occupy abundant samples while the other classes have decreasing number of samples. Although recent studies on SAR target recognition have achieved substantial progress based on deep learning models, they still suffer from an obvious performance drop on the tail classes. Therefore, how to improve the long-tailed SAR target recognition performance is a promising and challenging topic. In this work, we explore an improved contrastive learning based end-to-end training framework, which consists of triplet branch to enable balanced feature representation and less biased classifier. Specifically, the recalibrated prototype is introduced to handle data imbalance and assist contrastive learning. The optimization of two sub-tasks is achieved by a loss scheduler to progressively transmit the learning priorities. Experimental results on long-tailed SAR image dataset and optical remote sensing dataset demonstrate that our approach obtains remarkable performance boosts on tail classes and also achieves the overall state-of-the-art performance.