Abstract: [Purposes] Object detection methods serve as the core technology in the process of detecting foreign objects on transmission lines, bringing significant convenience to line inspection. However, due to the complex and variable environment of transmission lines, existing methods still face issues such as slow inference speed, poor detection performance for occluded objects, and low precision for small object detection. To address these issues, research on the RT-DETR object detection method was conducted, leading to the proposal of LDN-DETR (lightweight denoising-DETR) for foreign object detection on transmission lines. [Methods] Firstly, a lightweight feature extraction network based on variable kernel convolution was constructed to enhance detection performance for objects of various shapes and sizes, while improving inference speed. Subsequently, a cross-scale feature fusion module based on region-sensitive attention was designed to improve detection accuracy for occluded objects. Finally, a Gaussian modulation-based IoU-aware query denoising training method was adopted to address the issue of poor detection precision for small objects. [Findings] To prove the effectiveness of the method, experiments were conducted in this paper on RailFOD23 and VisDrone 2019. The results show that LDN-DETR has a 4% improvement in mAP over RT-DETR on RailFOD23, a 3% improvement on VisDrone 2019, and an inference speed of 35 frames per second with post-processing. [Conclusions] This method outperforms various mainstream object detection methods, thereby meeting the needs for inspection of foreign objects.

Key words: transmission inspection, feature extraction, cross-scale feature fusion, query noise reduction