An Improved RT-DETR Model for UAV-Based Power Line Inspection Under Various Weather Conditions

With the rapid advancement of deep learning, convolutional neural network-based target detection models have been effectively applied in UAV power inspection. However, the success of convolutional neural networks (CNNs) heavily relies on image augmentation and post-processing, both of which consume significant amounts of time in real-time object detection. This paper presents an end-to-end improved RT-DETR model that eliminates the post-processing module and introduces a reweighted L1 Loss to enhance the model's ability to learn small objects. Additionally, to improve the model's multi-scale feature fusion capability, we incorporate the GELAN module, combining CSPNet and ELAN models, allowing the network to more effectively propagate and aggregate feature information from different levels. Furthermore, to solve the shortage of power inspection datasets, we propose a Multi-Weather UAV Power Inspection dataset (MWUPI). Experimental results demonstrate that the improved RT-DETR model achieves high accuracy on this dataset while maintaining the fast detection speed of the original RT-DETR, meeting the performance requirements of real-time object detection. The codes for this study are publicly available at https://github.com/HantaoDing/Improved-RT-DETR-for-Power-Inspection.