Improved deformable DETR for transmission line hardware detection

Transmission lines are crucial for maintaining safe and reliable power grid operations, necessitating regular and efficient inspections. Considering the limited computing resources of drones and other equipment used in inspections. This study introduces a lightweight detection model based on Deformable DETR enhanced by two novel mechanisms: a Scoring Selector (SS) network and a lightweight backbone. The Scoring Selector network effectively discriminates foreground from background regions by integrating positional and semantic category information, significantly reducing computational complexity through precise token selection. Meanwhile, the lightweight backbone helps to accelerate the inference speed, and can be deployable on edge devices with limited computational resources, enabling wide-scale adoption in practical monitoring systems. Experiments conducted on a custom-constructed transmission-line hardware dataset demonstrate that the proposed approach significantly reduces computational costs (measured in GFLOPs and FPS).