PIDNet-Lite: A Lightweight Semantic Segmentation Model for Power Line Recognition

To address the challenges of real-time performance requirements and class imbalance in power line semantic segmentation tasks, this paper proposes a lightweight segmentation model named PIDNet-lite. Based on PIDNet, the proposed model integrates the Universal Inverted Bottleneck (UIB) to significantly reduce computational complexity through dynamic convolution and parameter redistribution. Meanwhile, a Soft-Hard Labels Align for Focal Distillation (SHD) is designed to improve adaptability to class imbalance scenes. Experimental results show that PIDNet-lite achieves higher inference efficiency while maintaining accuracy, making it suitable for edge deployment in power inspection applications.