Dual-Modal Fusion Model with Virtual Depth for Power Line Semantic Segmentation

Accurate semantic segmentation of power lines in complex environments remains challenging due to their slender structures, significant directional variations, and low contrast against backgrounds. Existing methods typically rely on expensive multi-sensor systems (e.g., LiDAR or thermal cameras) or fail to effectively address orientation sensitivity and edge preservation, particularly in multi-scale feature fusion. To overcome these limitations, this paper proposes a Dual-modal Fusion Vision Network (DFV-Net) leveraging virtual depth guidance for robust power line segmentation. First, we introduce a hardware-free depth generation strategy that produces virtual Digital Surface Models (DSMs) from RGB images using the Depth Anything v2 model, enriching spatial structural information without additional sensors. Second, we design an improved ResNet34-based feature extractor enhanced with Asymmetric Residual Blocks (ARBs), which integrate parallel heterogeneous convolutions, asymmetric convolution groups, and residual learning to optimize sensitivity toward slender, multi-directional power lines and enhance orientation-aware feature extraction. Finally, a Multi-modal Feature Fusion Module (MFFM) combines 3D convolution-based cross-modal interaction (MFFR layer) and multi-head self-attention with spatial reduction (MFSAF layer) to achieve efficient fusion of spectral and geometric features across scales, coupled with a lightweight multi-layer perceptron (MLP) decoder for pixel-wise classification. Experimental results demonstrate that DFV-Net achieves state-of-the-art performance on high-resolution imagery, with an F1-score of 85% and an overall accuracy of 98.812%, providing a cost-effective and reliable solution for power line segmentation. This advancement promotes the practical development of vision-based grid maintenance systems.