PNLNet: Partial Supervision-Driven Nonlocal-Local Network for Weakly Supervised Semantic Segmentation of Substation Point Clouds

Abstract

Existing semantic segmentation algorithms for substation point clouds typically rely on fully supervised learning, which requires a large amount of manually labelled point cloud data. This makes the segmentation process both time-consuming and costly. To address this issue, this paper proposes PNLNet, a pseudo-label-based weakly supervised learning algorithm. The network comprises two core components: a semi-supervised pseudo-label generation module and a point cloud segmentation main network. The pseudo-label generation module employs a semi-supervised strategy that combines a local-nonlocal relational graph with a graph convolutional network. This strategy performs a semi-supervised graph node classification task to generate point-level pseudo-labels for the point cloud data. The main network adopts an encoder-decoder architecture, incorporating a Convolutional Block Attention Module (CBAM) to enhance feature representation across different channels and extract critical information from diverse locations. Additionally, the network uses an inverted residual multi-layer perceptron (InvResMLP) to mitigate overfitting and gradient vanishing issues. Furthermore, a substation point cloud dataset was constructed for this study and the PNLNet algorithm was evaluated through ablation studies and comparative experiments with state-of-the-art fully supervised and weakly supervised learning methods. Experimental results demonstrate that PNLNet achieves segmentation performance comparable to the best fully supervised algorithms while reducing annotation time by 90%. The model significantly lowers the time and cost associated with processing substation point cloud data, maintaining high segmentation accuracy.