RT-DETR-E: A Real-Time Model for Foreign Object Detection in UAV Inspection Power Lines

With the development of science and technology, drone technology is gradually used by the power system inspection, especially in the inspection of power lines to detect foreign objects. Since UAVs can't fly for long time imitating lines due to the influence of hardware equipment, and the photos of foreign objects after shooting are usually small and medium targets, it is difficult to recognize them using the traditional YOLO network. To address this problem, this paper proposes an improved model RT-DETR-E based on RT-DETR; for the problem of background noise interference in the dataset, Concat_CAA and BFBE are designed to enhance the feature extraction capability and lightweight design of the model; for the problem of low inference speed of real-time detection, the real-time detection effect is achieved by introducing DySample. The experiments show that RT-DETR-E is optimal in mAP50 and params compared with some classical models of YOLO and RT-DETR series in the public dataset. Compared with the baseline model, the detection accuracy is improved by 3.4% and the number of parameters is reduced by 4.2%, which demonstrates the excellent performance and potential of RT-DETR-E in real-time detection. Moreover, the RT-DETR-E proposed in this paper reaches 140.8 (frames per second) in terms of inference speed, which meets the demand of real-time UAV detection.