Deep Learning Based Point Cloud Quality Improvement Method

3D data model is the most similar digital representation with the real world.With the development of science and technology,it has been applied in frontier fields such as industrial manufacturing,urban 3D modeling,medical diagnostic images,virtual augmented reality,etc.on a certain scale.There are many representations of 3D data,such as volume,depth map,mesh,and point cloud.Point cloud has expressive,homogeneous,and compact characteristics because of its simple data form.With the continuous updating of laser scanning equipment,point cloud becomes a kind of data that can be easily obtained.However,in the process of obtaining point cloud data,there are inevitable occlusions,disturbances,environmental impacts and possible human intervention that will cause the quality of the point cloud to decline,resulting in problems such as sparseness and missing,so these actual point cloud data obtained by scanning They often cannot be matched with high-quality objects in reality and cannot be directly used in model applications.The traditional point cloud upsampling method is difficult to play in the extremely dense part of the point cloud.At the same time,the point cloud completion method mostly fills small holes based on the geometric structure,and cannot repair the large missing area.In order to overcome the shortcomings of traditional methods,based on the two major problems of sparse point cloud and missing point cloud,this paper proposes a deep learning-based point cloud quality improvement method.It uses a large database as a priori information of the network,merges the global characteristics of the 3D model,and generates high quality point cloud data.For the problem of sparse point cloud,this paper uses the feature extraction method of PointNet to construct a point cloud upsampling network with autoencoder structure.The encoder is used to extract the low-dimensional latent vector of the input point cloud;the decoder uses the low-dimensional vector to predict the three-dimensional space coordinates of the point cloud.This network is different from the traditional upsampling algorithm and can play a role in the extremely sparse point cloud.For the point cloud missing problem,this paper proposes a point cloud completion algorithm based on cascading generative adversarial network.The encoder network is used to extract the features of the input point cloud data,and the training to generate the confrontation network is performed in the latent space.The dual cascade generation adversarial network generates the low-dimensional vector of the preliminary repair model mapping in the first GAN,and the low-dimensional vector of the final completion point cloud map in the second GAN,and finally relies on the decoder to output the complete point cloud model.Compared with the current high-performance deep learning methods,this network is more lightweight and has the point cloud completion effect that is not inferior to other methods.