Efficient and Lightweight Semantic Segmentation Network for Land Cover Point Cloud With Local-Global Feature Fusion

• Published in: IEEE transactions on geoscience and remote sensing Vol. 63; pp. 1 - 13
• Main Authors: Hu, Hong, Cai, Lantian, Kang, Ruihong, Wu, Yanlan, Wang, Chunlin
• Format: Journal Article
• Language: English
• Published: New York IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Cognitive ability; Computer applications; Computing costs; Convolution; Data mining; Datasets; Deep learning; Feature extraction; Feature fusion; Floating point arithmetic; Image segmentation; Inference; Land cover; land cover semantic segmentation; Land surface; point cloud; Point cloud compression; Semantic segmentation; Semantics; Three-dimensional displays; Training; Transformers; Urban planning; VEF-Net
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2025.3559590

Utilizing deep learning techniques to extract high-precision features from point clouds is essential for accurately capturing land cover information, which is instrumental in urban planning and environmental conservation. Despite delivering high-accuracy outcomes in the semantic segmentation of extensive terrestrial point clouds, prevalent methodologies encounter considerable hurdles, particularly in training and inference duration, as well as the associated hardware expenses. To solve these issues, this article introduces an efficient and lightweight deep learning network called Uniform Voxelization Geometric Enhancement and Local-Global Feature Fusion Network (VEF-Net). VEF-Net is designed to improve the training and inference efficiency of large-scale point cloud semantic segmentation while maintaining accuracy. Uniform voxel downsampling is employed to discretize point clouds, resulting in a substantial enhancement in computational and memory performance. To counteract potential information loss due to voxel downsampling, VEF-Net integrates a mechanism unit to enhance local geometric features, enriching point cloud data. Furthermore, it incorporates a local-global feature fusion module, adeptly capturing the global contextual relationships within the point cloud. Experimental results show that VEF-Net achieved excellent performance on both the proprietary Bengbu dataset and the publicly available Toronto-3D dataset. VEF-Net maintained comparable segmentation accuracy to mainstream models while operating at a lower computational cost, and it demonstrated significant advantages in certain key tasks. On the Toronto-3D dataset, VEF-Net achieved an intersection over union (IoU) of 20.4% in the road marking category. Additionally, VEF-Net demonstrated lower model parameters and computational complexity [floating point operations per seconds (FLOPs)], achieving training speeds 12 times faster and inference speeds 16.8 times faster than point transformer (PT).