PLRefiner: Discrete Diffusion Model for Power Line Segmentation Refinement

Refining segmentation of power lines is critical to ensuring the flight safety of unmanned aerial vehicles. However, due to their extremely slender linear structures and high confusion with complex backgrounds, traditional segmentation methods are prone to breakage, blurring, or false detection. These challenges render power line segmentation a particularly difficult task in the field of computer vision. To address these challenges, this study proposes PLRefiner, a discrete diffusion-based power line refinement model that progressively eliminates background false positives and repairs fragmented regions through an iterative denoising strategy. The core innovations of PLRefiner include: (1) Global context modeling-integrating a Vision Transformer module to capture long-range dependencies of power lines via a self-attention mechanism; (2) Structural continuity constraints-introducing a Hough space alignment loss that employs an angular segmentation penalty to guide the model in restoring defects along the power line direction; and (3) Lightweight post-processing-leveraging dynamic thresholding and connected-component analysis to remove discrete false detections. Experimental results on the TTPLA and PLD datasets demonstrate that PLRefiner significantly outperforms existing methods. After refining the coarse masks produced by multi-class segmentation models, the proposed approach achieves an average IoU improvement of 6.82% and an mBA increase of 12.54%, which is 4.18% and 1.41% higher than other comparative refinement methods. Ablation studies and field validation further confirm its effectiveness in long-range restoration, edge optimization, and noise suppression, establishing a high-precision and robust segmentation foundation for power grid inspection applications.