关键词: 电力线分割, 轻量级网络, 深度可分离卷积, 注意力机制, 方向感知卷积

Abstract: To improve the segmentation accuracy and deployment efficiency of power line targets in power inspection images, a lightweight semantic segmentation network named LPL-Net (Light Powerline Net) was proposed. The network adopts a dual-branch feature extraction mechanism. The backbone path consists of a channel-adaptive Multi-scale Lightweight Convolution (MLC) module. This module constructs Lightweight Convolution (LightConv) modules using depthwise separable convolutions, combined with a channel attention mechanism Squeeze-and-Excitation Block (SEBlock), to efficiently extract local features while enhancing the adaptive weighting ability between channels. The auxiliary branch uses bilinear interpolation and shallow convolutions to extract multi-scale detail information, alleviating the feature loss of power lines caused by downsampling. Additionally, a (Multi-Direction-Aware Convolution (MDAC) module is introduced to improve the model's ability to model the elongated and directional structure of power lines. To validate the model's performance, a dataset named SPLS (Surveillance-based PowerLine Segmentation) was constructed, covering diverse and complex scenarios. Comparative experiments were conducted with representative lightweight semantic segmentation methods, including FR-Unet (Full-Resolution Network), CS^2-Net (Curvilinear Structure Segmentation Network), and MobileNet-V3. Experimental results show that LPL-Net, with only 0.48M parameters and 160.69 GFLOPs of floating-point operations, achieves a 0.22% and 1.12% improvement in F1-score and IoU (Intersection over Union), respectively, compared to the baseline model FRD-Net, with an approximate 2.8-fold increase in inference speed. This method balances accuracy and deployment friendliness, making it suitable for rapid power line target perception tasks in resource-constrained environments.

Key words: power line segmentation, lightweight network, depthwise separable convolution, attention mechanism, direction-aware convolution