Transmission line icing is one of the major natural hazards threatening the safe and stable operation of power grids. Real-time and accurate icing monitoring plays a crucial role in preventing accidents such as line galloping and tower collapse. Traditional manual inspection and image processing methods suffer from low efficiency and high cost, while deep learning-based semantic segmentation methods often fail to achieve fine-grained segmentation due to complex backgrounds and adverse weather conditions. To address challenges including complex background interference and blurred edge features in transmission line icing segmentation, this study proposes a fine-grained segmentation algorithm for transmission line icing under complex background conditions. Based on the lightweight SegFormer-B2 framework, a multi-scale feature enhancement U-shaped network (MSFE-USN) is introduced between the encoder and decoder. The network first performs feature fusion through a U-shaped network (USN), and then employs an edge-guided multi-scale attention module (EMSAM) to enhance the discrimination of icing contours and background boundaries. Subsequently, a convolutional block attention module (CBAM) is utilized to suppress interference from complex backgrounds such as rain, snow, fog, and water surfaces. Experimental results on a self-constructed transmission line icing semantic segmentation dataset (TLI-SSD) show that the proposed model achieves a mean intersection over union (mIoU) of 94.07%, outperforming other comparative and ablation models across all evaluation metrics. Visual comparison results further demonstrate the superior fine-grained segmentation capability of the proposed method under extreme weather and complex background conditions. This study provides an effective solution for the automatic and high-precision transmission line icing segmentation and demonstrates significant engineering value for intelligent power grid operation and maintenance.