The Study Of Point Cloud Denoising And Model Classification Based On Deep Learning

Reverse engineering is a process of technology reproduction centering on product design technology.In other words,it mainly carries out reverse analysis on target product to obtain the organizational structure,functional characteristics,technical specifications and other design elements of the product during the processing process,so as to produce similar but different.products.Reverse engineering is widely used in the fields of new product development,product modification design,product imitation,quality analysis and testing.As a model in reverse engineering,point cloud is the data that can best represent the threedimensional features of objects,and it is favored by researchers.In the study of reverse engineering,point cloud processing is a particularly important step.However,point cloud cannot be directly used as the input of CNN,which needs further processing,because of it have spatial chaos,rotation invariance and unstructured characteristics.At present,there are three mainstream methods in deep learning of point cloud: multi-view,voxel and point cloud.Based on this,this research is mainly divided into two aspects,denoising based on deep learning for point cloud and point cloud model classification.Specific research contents are as follows:(1)In terms of denoising deep learning for point cloud,this paper proposes a denoising method based on neural network classification in view of the complexity and "inadaptability" of traditional algorithms.Firstly,the characteristics of all points in the point cloud of the training set are marked: they belong to or do not belong to the original clean point cloud.Then use the Multilayer Perceptron(MLP)neural network to identify the features of the points and determine the points belonging to the point cloud.The points belonging to the point cloud are retained,and the points not belonging to the point cloud are removed to obtain the coarse denoising point cloud.Finally,the outliers in the rough denoising point cloud are removed to obtain the denoising point cloud.Experiments were carried out on the point cloud with random noise,in the mixed data of 20,000 points of point cloud and noise points,the method identified 10030 noise points and 9970 non-noise points,the results showed that it could identify 99% of noise points.(2)In the classification problem of point cloud model,in view of the effect of data set on classification efficiency not mentioned in the previous multi-view neural network algorithm,this paper uses the mapping image feature function of convolutional neural network to analyze this problem.The weights of projection features from different perspectives in the convolutional neural network are analyzed with mixed view data set,and the collection density of the mixed view data set is optimized according to their weight ratios.The final experimental results show that the classification accuracy of two-dimensional images generated from different angles is different,of which the classification accuracy of the overhead projection is the worst.The classification accuracy of the optimized data set was improved by 2.3% under the same neural network model.