Multi-Scale Feature Optimization Point Cloud Completion Network Integrating SoftPool

作     者：Zhang, Wanpeng Zhou, Ziwei 

作者机构：School of Computer Science and Software Engineering University of Science and Technology LiaoNing Anshan114051 China School of Computer Science and Software Engineering University of Science and Technology LiaoNing Anshan114051 China 

出 版 物：《IAENG International Journal of Computer Science》 (IAENG Int. J. Comput. Sci.)

年 卷 期：2025年第52卷第1期

页      面：233-243页

核心收录：

基　　金：This work was supported by the Natural Science Foundation of China(Grant No. 61575090)  the Natural Science Foundation of China Youth Fund(No. 61803189)  Natural Science Foundation of Liaoning Province(2019-ZD-0031 and 2020FWDF13) 

主　　题：Ability testing 

摘      要：Point cloud completion is crucial in point cloud processing, as it can repair and refine incomplete 3D data, ensuring more accurate models. However, current point cloud completion methods commonly face a challenge: they fail to fully utilize multi-scale information from local features, leading to limitations in accuracy and detail preservation. To address this issue, this paper proposes a multi-scale feature optimization algorithm for point cloud completion that integrates SoftPool. Based on DGCNN, the method combines dilated convolution and bottleneck attention mechanisms to extract features at different scales, enhancing the ability to capture detailed information in point clouds. The bottleneck attention mechanism is used to optimize important detail features. The extracted local features are concatenated with their corresponding positional information to form point proxies, enhancing the effective extraction of local geometric features, resulting in more refined completed point cloud shapes. A Transformer architecture is employed to model these features. Finally, SoftPool is introduced for fine-grained feature downsampling, improving the network s ability to recover point cloud details. FoldingNet is used to reconstruct missing structures and output the completed point cloud. To validate the model s completion performance, training and testing are conducted on the PCN and ShapeNet55 datasets. Experimental results demonstrate that the model has better feature detail retention and more accurate completion results. On the PCN dataset, the average CD value is reduced by 6.5% compared to the best-performing model among the comparison methods. On the ShapeNet55 dataset, the average CD value across three difficulty levels is reduced by 6.9% compared to the best-performing model among the comparison methods. Additionally, the model also achieved a 2.1% improvement in F-score. © (2025), (International Association of Engineers). All rights reserved.