A Topology-Constrained SAM2-UNet with Linear Structure Awareness for Power Line Segmentation

To address the challenges of extreme aspect ratios in automatic power line segmentation for UAV inspection scenarios, this paper proposes a linear structure-aware optimization framework based on an improved SAM2-UNet. Adopting the mechanism of SAM2-UNet, the framework preserves the general representation capability of SAM2-Hiera while introducing three key enhancements tailored for accurate power line segmentation. First, a Linear Structure Enhancement Block (LSEB) integrating deformable convolutions with direction-decoupled strip convolutions is designed to adaptively capture geometric characteristics. Second, a Direction-Aware Attention (DAA) mechanism that utilizes gradient direction priors for feature re-weighting is introduced to enhance spatial coherence. Finally, a boundary-orientation consistency loss function is formulated by transforming topological connectivity constraints into differentiable optimization objectives. Experiments on the TTPLA (Transmission Tower and Power Line Aerial) dataset demonstrate that the proposed method achieves accurate power line segmentation. Compared with the standard SAM2-UNet, the proposed method yields improvements of 3.09 and 2.64 percentage points in IoU and Dice, respectively.