An Integrated Image Generation and Collaborative Training Framework for Transmission-Line Foreign-Object Detection

Abstract

Foreign objects on power transmission lines pose a serious threat to the safe and stable operation of smart grids. Currently, deep learning-based detection models have become the mainstream method for detecting foreign objects on transmission lines. However, all foreign objects are treated as one detection class, whose large intra-class diversity in shape, material, scale, and attachment configuration is difficult to cover with limited real training data. One promising solution is to use generative models to synthesize supplementary training data for transmission-line foreign objects. However, most existing generative models are designed for a general domain. How to generate high-quality synthetic images and effectively utilize them for model training remains an open problem. In this paper, we propose an integrated image generation framework specifically designed for transmission-line foreign-object detection. This framework edits background images into foreign-object images, automatically annotates the generated images, and uses the annotated images to train detection models. Firstly, we propose a decoupled generation method, which decouples the generation of the foreign object itself from the embedding of the foreign object in the transmission-line scene, reducing the complexity of synthesis and encouraging diverse appearances and visually plausible attachment relationships. Secondly, we design an automatic annotation module for synthetic samples. By utilizing the dense feature differences between generated images and original images and combining zero-shot segmentation methods, we achieve automatic annotation of the generated foreign objects. Finally, we propose a collaborative training strategy based on gradient consistency, dynamically evaluating and giving higher training weights to synthetic samples that are beneficial for real-domain optimization, thereby effectively suppressing the interference of conflicting supervision signals. Experimental results on real transmission-line datasets show that the proposed method not only generates images with higher visual quality and naturalness, but also improves the foreign-object detection performance of YOLO11s and RT-DETR under complex conditions.