UAV small object detection algorithm based on dynamic sampling and scale-adaptive loss

Abstract

To address core issues in small object detection for UAV aerial imagery—including irreversible loss of fine-grained features caused by fixed-stride down sampling, insufficient perception of small-object features by the original RT-DETR model, and low sensitivity of localization loss—this paper proposes a detection algorithm named DyS-RTDETR-SOD based on dynamic sampling and scale-adaptive loss. Taking RT-DETR as the baseline architecture, the algorithm first introduces the Dysample dynamic sampling module to replace fixed-stride convolutional down sampling in the backbone network, which adaptively preserves edge and texture details of small objects during feature dimensionality reduction. Second, the original position regression loss function is substituted with Normalized Wasserstein Distance (NWD) loss to strengthen the model’s sensitivity to small-object bounding boxes and boost localization accuracy. Experiments conducted on the public CODrone dataset demonstrate that the proposed DyS-RTDETR-SOD achieves an mAP50 of 37.5%, representing a 1.6 percentage point improvement over the baseline RT-DETR. Meanwhile, its computational complexity is slightly reduced. The overall performance outperforms mainstream lightweight detection algorithms, enabling real-time deployment on edge devices for UAVs.