An Improved RT-DETR Algorithm for Small Object Detection in Aerial Images

To address the issues of low detection accuracy in drone aerial images due to small object sizes, indistinct features, complex backgrounds, and significant variations in object scale, this paper proposes an improved small object detection algorithm for drone aerial images: RT-DETR (Real-Time Detection Transformer). First, a lightweight multi-scale feature extraction module (CF-Block) is introduced into the backbone network, where the FasterNet Block is enhanced with convolutional gated linear units to improve the extraction of small object features and reduce computational redundancy. Second, a lightweight feature fusion mechanism, SOEP (Small Object Enhance Pyramid), is designed to integrate additional shallow feature maps, enhancing small object feature information, while the Omni-Kernel Module (OKM) is employed to fully fuse multi-scale feature information. Finally, the loss function is optimized with the proposed W-ShapeIoU, which maintains scale invariance while adopting a gradient gain distribution strategy to suppress harmful gradi-ents generated by low-quality anchor boxes, thereby improving the detector's overall performance. Experiments on the public Visdrone2019 dataset show that the improved algorithm reduces the parameter count by 18.8 % compared to the baseline model, while increasing [email protected] and [email protected]:0.95 by 2.6% and 1.9 %, respectively. Compared to current mainstream algorithms, the proposed method achieves an accuracy improvement of more than 8 %, demonstrating superior detection performance.