LRS-DETR：A lightweight and efficient real-time detection algorithm for small targets of UAVs based on RT-DETR

To address the challenges of small object detection in complex UAV environments, including dense target distribution, difficulty in feature extraction, and the limited computational resources of UAV platforms. This paper proposes LRS-DETR, a lightweight and efficient small object detection algorithm for UAV applications. First, we introduce the Feature Re-parameterized Partial Convolution (FRPCBlock) to enhance the backbone network, improving feature extraction capability while maintaining lightweight characteristics, thereby increasing computational efficiency. Then, we employ the Dynamic Position-Bias Attention Feature Interaction module (DPAFI) to enhance the model's cross-scale feature modeling ability. Next, we integrate the P2 detection layer into the ASF architecture to enable dynamic feature selection and hierarchical transmission, improving feature fusion quality through scaling and sequential processing. Finally, we employ the Focaler-Powerful-IoU regression loss function to enhance the model’s small object detection capability. Experimental results show that LRS-DETR reduces the number of parameters by 41.1% and computational complexity by 13.8%. On the VisDrone-2019 dataset, mAP0.5 and mAP0.5:0.95 increase by 2.4% and 1.7%, reaching 49.9% and 31.2%, respectively, achieving both lightweight efficiency and improved accuracy.