DS-DETR: Dynamic Sparse Detection Transformer for Small Object Detection in High-Resolution Remote Sensing Images

Remote sensing images have been widely applied in both military and civilian fields. However, targets in high-resolution remote sensing images are typically extremely small, with features easily intertwined with complex backgrounds, leading to feature aliasing and poor discriminability. Additionally, the high computational cost of traditional attention mechanisms limits the practical deployment of Transformer-based detectors. To address these challenges, we propose DS-DETR, a Detection Transformer that integrates dynamic feature enhancement and sparse attention mechanisms, developed based on RT-DETR. It incorporates three key innovations: first, the dynamic feature enhancement network reduces the parameters of the baseline while enhancing the features of weak and small targets; second, the adaptive sparse attention mechanism focuses on potential target regions and suppresses redundant background features, while reducing computational complexity; finally, the detail refinement module refines the local textures of targets. The experimental results on the DOTA v1.5 dataset show that the proposed model improves mAP by 2.5, reduces parameters by 5.6M, and increases FPS from 94 to 112 compared to the baseline. It also exhibits stable performance gains on the NWPU VHR-10 dataset. These results indicate that DS-DETR provides a reliable solution for real-time and high-precision remote sensing object detection. The code will be available at https://github.com/yfchenbruce/DS-DETR.