Active Spatio-Fine Enhancement Network for Semantic Segmentation of Large-Scale Point Clouds

Point cloud neighborhood creation constitutes a pivotal element in point cloud semantic segmentation, facilitating the understanding of 3-D scenes. However, prevailing models’ capacity to comprehend scenes is limited due to their reliance on a singular neighborhood construction technique for extracting neighborhood attributes. Moreover, although deep learning has effectively utilized the attention mechanism, it is constrained by assigning a single task to attention weights, thereby lacking flexibility in expressing feature correlations among adjacent points. This article addresses these issues by introducing the active spatio-fine enhancement network (ASFE-Net), which amalgamates an innovative local spatial structure encoder (SSE) module and a sophisticated attention fusion (SAF) module into the recognition and processing of point cloud data, thereby significantly enhancing the recognition of crucial local information. Furthermore, the adaptive feature scaling (AFS) module improves the ability to perceive complicated spatial relationships and captures details of global features. Tests using several datasets, including Stanford large-scale 3-D indoor space (S3DIS) and Toronto_3D, confirm that ASFE-Net is the best option for point cloud semantic segmentation tasks. In addition, pertinent ablation experiments were carried out to demonstrate the efficacy of the different modules within the ASFE-Net.