A Point Cloud-based Multi-scale Fusion Segmentation Model for Surface Pore Structures in 3D Printed Concrete (2026-01)¶

The pore structure of 3D printed concrete (3DPC) significantly affects its appearance and mechanical properties. Therefore, developing an efficient and accurate method to identify these micro-pores is essential for studying how pore structure influences 3DPC performance. However, this task is challenging because 3DPC pore structures are often irregular, exist in complex backgrounds, and are highly dense, making them difficult to acquire and identify clearly. To tackle these challenges, this paper first reviews existing research to address the current lack of studies on defect identification in 3DPC.Then, this paper creates a high-quality information-enhanced 3D point cloud dataset of 3DPC pore structures using a 3D concrete printer and high-precision scanning equipment.This paper also proposes a strategy to redefine data features based on information enhancement to fully represent pore structure information, filling the gap in available point cloud datasets for pores. Furthermore, this paper proposes a new multi-scale fusion model for segmenting surface pore structures from point clouds, called PointOSamba. This model aims to improve the accuracy of identifying pore defects in 3DPC. For the first time in this field, this paper introduces both point cloud segmentation and the Mamba architecture. Our contributions include: (1) a multi-scale feature-enhanced sampling strategy to reduce the loss of local details during sampling; (2) a spatial weight-fused ordering method to overcome the limitation that point clouds’ unordered nature imposes on network learning.(3)While Transformer-based methods have advanced point cloud analysis, they require substantial computation. In contrast, the Mamba backbone network in our model offers global modeling with optimized complexity. (4) this paper also proposes a weighted pooling operation to better integrate the performance of each module. As a result, the model achieves an optimal balance between the number of parameters and computation time under comparable conditions. Evaluation shows that our model achieves higher segmentation accuracy for micro-pore defects and is more effective at handling class-imbalanced data than other leading methods. Ablation studies confirm the effectiveness of each proposed module.