Integrating Convexity and DeepLabv3+ for SemanticSegmentation of Power Lines

To address the challenges of segmenting slender, low‑contrast power lines in complex aerial inspection imagery, this paper proposes PLE‑DeepLabv3+, an enhanced deep learning model based on DeepLabv3+. The method employs a channel reconfigured ConvNeXtV2 backbone to maintain high resolution spatial details while reducing parameters, incorporates a direction selective coordinate attention module (CA\_DS) to enhance the perception of linear structures, and replaces standard convolutions in the atrous spatial pyramid pooling (ASPP) block with depthwise separable convolutions for efficient multi scale fusion. Additionally, an adaptive edge detection module (AEDM) is integrated to reinforce weak boundaries, and a category aware focal loss (FL\_AL) is designed to alleviate extreme class imbalance. Evaluated on a composite power line dataset, the proposed model achieves 85.2\% mIoU and 92.7\% mPA, with only 2.3 M parameters and 78.7 GFLOPs, surpassing both general and task specialized segmentation networks. Cross dataset tests confirm its strong generalization across varied scenes. The lightweight and accurate architecture provides a reliable visual solution for automated inspection in power grid monitoring systems.