To improve the recognition accuracy during power line inspection, we propose a power line image segmentation network called LTSeg-Net, which is based on multi-scale context and full-stage feature fusion. First, a multi-scale context extraction (MSCE) module is introduced to capture the structure of power lines in complex backgrounds using information from different receptive fields. Second, a full-stage feature fusion (FSFF) structure is designed to enhance the extraction of key information and suppress background interference through multi-scale feature aggregation and an efficient additive self-attention (EASA) mechanism. Additionally, a power line segmentation dataset (Surveillance-based PowerLine Segmentation, SPLS) covering various weather and lighting conditions was constructed, and segmentation tests were conducted on this dataset. Compared to the state-of-the-art RCFSNet method with superior segmentation performance, the IoU, Pre, and F1 metrics were improved by 0.89%, 0.97%, and 0.99%, respectively. Experimental results demonstrate that the LTSeg-Net method can more accurately extract power line structures in complex environments.