Density Prior-Guided Unsupervised Deep Point Cloud Denoising Algorithm

In order to improve the performance of the unsupervised deep point cloud denoising algorithm, based on the existing network framework, firstly, the density prior is designed, which uses the distribution of the noisy point clouds to describe the probability of each point being on the real underlying surface. Then, the denoised point cloud is obtained by the deep network which uses upsampling, downsampling and other strategies to overcome the influence of high-noise points. Finally, the density prior is used to optimize the conditional probability distribution of clean points, and the unsupervised loss function is designed which optimizes the network to obtain the final algorithm. In addition, based on the density prior, a low-noise points screening method and filtering algorithm are proposed. The algorithm is implemented on PyTorch, testing on the real scan dataset PERL and the simulation dataset constructed based on ModelNet-40, and taking the Chamfer distance and the point-to-surface distance as evaluation metrics. Compared with other algorithms such as DMR, the Chamfer distance is reduced by 0.35~1.34, and the point-to-surface distance is reduced by 0.68~1.94 on average. The experimental results show that the proposed algorithm outperforms state-of-the-art denoising methods with robustness, universality, and generalization ability.