Improving the Long-tailed Remote Sensing Target Detection via Target-level Comparision

Gang Yan, Yingbing Liu, Wei Hu, Fan Zhang

Abstract: With the continuous development of earth observation technology, high-resolution satellite images with more detailed ground object information have become increasingly prevalent. As a result, multi-classification target detection has become one of the hot spots in this field. However, in reality, remote sensing images face the problem of limited samples under a long-tailed distribution, which hinders the model's ability to extract discriminative features for the target. To address this issue, we designed the TCLM module based on contrastive learning, aiming to address the difficulty of feature extraction for tail class targets. Meanwhile, there exists another issue: complex background noises also interfere with target feature extraction. To overcome this problem of background noise interference in complex remote sensing images, we propose the PARM module, which can help the model to focus on the target itself, thereby decreasing its attention on the background. Experimental results on the MSAR and DOTA datasets show that our method significantly improves detection performance.