DETR-RSI:An Improved RT-DETR-Based Object Detection Framework for High-Resolution Remote Sensing Imagery

Remote sensing object detection is an important application of computer vision in remote sensing data. However, remote sensing images often exhibit complex backgrounds, densely distributed small objects, and significant variations in object scales, which pose great challenges to detection algorithms. This paper proposes a high-resolution remote sensing image object detection algorithm, DETR-RSI, based on an improved RT-DETR, to enhance the detection accuracy and efficiency of the model on remote sensing images. A lightweight feature extraction backbone network, RPKINet, is designed, a smallobject detection layer is added, and a cross-scale feature fusion module, GD-CCFF, is introduced to address the problems of object occlusion and information fusion. Additionally, CIoU loss and Smooth L1 loss are incorporated to optimize the loss function, significantly improving the performance of small object detection. Ablation experiments conducted on the DOTA v1.0 dataset demonstrate the effectiveness of each proposed module. Meanwhile, comparative experiments on the DIOR dataset show that DETR-RSI achieves an average precision (AP) of 57.3%. Compared with various classical object detection algorithms, DETR-RSI also exhibits significant advantages in detection accuracy, especially excelling in the detection of medium and large objects. The source code of DETR-RSI will be made publicly available at https://github.com/Fangkui00/DETR-RSI upon acceptance for reproducibility and further research.