关键词: 三维点云, 语义分割, 高压输电线路, 特征增强, 类别不均衡

Abstract: Due to the phenomenon of sample imbalance among target categories in high-voltage transmission scenarios,the intelligent inspection system for high-voltage transmission lines usually has low recognition accuracy for few-sample targets.To address this issue,this paper proposes a 3D point cloud semantic segmentation method that integrates category awareness.Firstly,an adaptive dynamic sampling strategy is introduced,which optimizes the distribution of point cloud data through density-aware region division to improve data balance.Secondly,a class-aware contextual feature enhancement module is designed,which dynamically fuses points features using class embedding information to strengthen the model's discriminative capability.Finally,a weighted loss function is constructed to alleviate learning bias caused by long-tailed data distributions.Experimental results on real-world point cloud datasets of high-voltage transmission lines demonstrate that the proposed method not only improves overall segmentation accuracy but also achieves superior recognition performance for minority classes.This study provides effective technical support for intelligent recognition of complex structural targets in power line inspection and shows promising potential for practical engineering applications.

Key words: 3D point cloud, semantic segmentation, high-voltage transmission line, feature enhancement, class imbalance