Detection of Power Transmission Lines Using an Improved RT-DETR Model

The stability of power energy systems is paramount for both societal productivity and the well-being of individuals. Despite being the prevailing real-time object detection frame-work, the YOLO series detectors have demonstrated respectable performance in such applications. However, due to the existence of post-processing Non-Maximum Suppression (NMS), they fall short of achieving true end-to-end detection. Similarly, DETR-type detectors are rarely utilized for real-time object detection due to their prohibitive computational costs until the introduction of RT-DETR. This paper introduces multiple plug-and-play modules aimed at addressing three key aspects (spatial redundancy, channel information loss, and insufficient multiscale feature fusion) to refine the Neck layer of the original RT-DETR model, thereby significantly enhancing its performance. Moreover, accurate electrical system detection requires heightened precision, particularly under high Intersection Over Union (IOU) thresholds, leading to the adoption of different matching cost computation methods at various training stages to shift the model's attention from local to global. Finally, on our in-house dataset demonstrates an approximately 2% improvement in mean average precision (MAP) compared to the original RT-DETR, consistently surpassing prevalent real-time object detectors and affirming the effectiveness of the proposed enhancement strategy in electrical system detection. Despite the considerable drop in Frames Per Second (FPS) due to the introduction of multiple modules, the inference speed is still comparable to that of some real-time object detectors, which we find acceptable.