基于深度学习的点云去噪方法综述

Point clouds, as an important carrier of 3D data, are widely applied in autonomous driving, embodied intelligence, virtual reality, and other fields. However, raw point clouds are often corrupted by sensor errors, environmental interference, and reconstruction algorithm defects, resulting in noise, geometric structure distortion, and degraded performance in downstream tasks. To address these issues, numerous deep learning-based point cloud denoising methods have been proposed. This paper systematically reviews developments in this field over the past eight years. First, a fivecategory taxonomy for deep learning-based point cloud denoising is established, encompassing PointNet-based backbones, convolutional neural network(CNN), Transformer networks, nontraditional networks, and unsupervised learning strategies. The core mechanisms and evolutionary trajectories of these algorithms are analyzed in depth. Second, the performance of mainstream algorithms is evaluated and compared on both synthetic and real-world datasets. Finally, current challenges, such as limited generalization ability and data scarcity, are summarized, and future research directions, including the development of all-in-one models and adaptive noise modeling, are discussed. The analysis presented in this paper provides a reference for the taxonomy, performance evaluation, and future improvement of deep learningbased point cloud denoising methods.