基于掩膜自编码器的隧洞点云去噪无监督网络OA

Existing supervised learning based point cloud denoising networks rely on pairwise valid point clouds and noisy point clouds as training data.However,in most cases,only noisy point cloud data is available,and the corresponding valid point cloud is unavailable.In addition,existing point cloud denoising methods have shortcomings in feature extraction,and their performance is poor in complex environments such as tunnels.To address the above issues,this paper proposes a novel unsupervised Point Cloud Denoising network,Masked Autoencoders for Point Cloud Denoising(MAE-Denoised).The network consists of three parts:a feature extraction module,a noise prediction module,and an iteration module.In the feature extraction process,the feature pre-training parameters of Masked Autoencoders are transferred,and the dynamic Adapter module is combined to improve the extraction ability of local features.In the downstream task,local and non-local features are encoded by a multi-layer Perceptron(MLP)to predict the displacement of each point.Finally,the iterative module was used to perform multiple noise predictions to complete point cloud denoising.In the unsupervised training process,the following strategy is used to learn the input noisy point cloud:the loss function is constructed by expanding the backbone network into three branches for cycling and crossover.A large number of experiments show that the proposed method achieves excellent denoising results on both synthetic and real noise data sets,and its performance is superior to that of mainstream deep learning algorithms.