Deep non-local point cloud denoising network

As an efficient representation of objects, the 3D point cloud is increasingly prevalent in various application fields. However, raw point clouds captured from scanning devices often contain noise, which significantly impairs the performance of downstream tasks such as surface reconstruction and object recognition. Consequently, point cloud denoising has emerged as a crucial task in geometry modeling and processing. Although deep learning has proven effective in this domain, existing learning-based methods predominantly focus on local information and tend to neglect the non-local features inherent in 3D point clouds. In this paper, we propose a deep non-local point cloud denoising network, DnPCD-Net, to address this issue. DnPCD-Net consists of three key components: 1) a feature extraction module that extracts local features for each point; 2) a densely-connected Transformer module that captures long-range dependencies across the input point set and feature channels; and 3) a feature fusion module that adaptively combines local and non-local features. Extensive experiments on both synthetic and real-scanned datasets demonstrate that DnPCD-Net achieves superior denoising performance, with statistically significant improvements in Chamfer Distance and Earth Mover's Distance, as well as better visual quality, confirming its effectiveness and robustness in practical applications.