Research on segmentation technology of transmission lines based on aerial images

With the continuous expansion of power grids worldwide, transmission lines have become critical infrastructure for long-distance energy transmission, and their operational status directly affects the safety and stability of the grid. Unmanned aerial vehicles (UAVs) inspection, known for its high efficiency and flexibility, has been widely adopted in power systems. Accurate segmentation of transmission lines and towers from aerial images has thus become an important prerequisite for achieving autonomous inspection, path planning, and obstacle avoidance. However, existing methods still face three major challenges in real-world aerial inspection scenarios: limited annotated samples leading to insufficient model generalization, complex backgrounds that easily cause mis-segmentation, and the slender structure with blurred boundaries of transmission lines that results in poor structural continuity. To address these issues, this paper proposes a transmission line segmentation method that integrates self-supervised pretraining with an improved UNet architecture. First, a self-supervised pretraining strategy is employed to learn general feature representations from unlabeled aerial images, thereby alleviating performance degradation under small-sample training conditions. Second, a Split-Attention mechanism is introduced in the encoding stage to enhance the model’s capability of extracting fine-grained features in complex background environments. Finally, a multi-scale FocusFeature fusion module is designed in the decoding stage to improve continuity restoration and boundary segmentation accuracy for slender transmission lines. Experiments conducted on the public TTPLA Dataset dataset demonstrate that the proposed method achieves 86.00 \(\%\) , 92.33 \(\%\) , and 91.00 \(\%\) in terms of mIoU, mPrecision, and mRecall, respectively, representing improvements of 9.92 \(\%\) , 3.60 \(\%\) , and 7.08 \(\%\) over the original U-Net model. Moreover, compared with methods such as UNet-ViT, UNet-HRNet, UNet-MobileNetV3, and UNet-ResNet34, the proposed model demonstrates superior overall performance in complex-background and fine-grained target segmentation tasks. In addition, the proposed method still relies on high-quality aerial imagery data, and its robustness in extreme weather conditions, severe occlusion scenarios, and cross-domain transfer tasks requires further improvement. Future research may extend this approach to power facility defect detection, autonomous UAV inspection, and other slender-object segmentation tasks, such as road crack detection and remote sensing linear target extraction.