3D CAD Model Classification Based On Deep Learning And Views

The number of 3D models has gained an increase at geometric rate in recently ten years.How to process,analyze and understand the 3D models has become a focus of the research in the field of digital geometry.The key problem is to obtain a good representation for shapes.The previous methods use hand-crafted features in classification problem.The representation quality depends on people's understanding.Therefore,these previous methods are subjective and difficult to generate accurate representations for complex 3D shapes classification.Different from most of the methods mentioned above,the deep learning methods learn the representation and classification from data with a variant of CNN,which are widely used and have achieved superior performance in the field of images.This thesis focuses on 3D CAD models classification based on deep learning and views: 3D models are firstly converted to a collection of their rendered views on 2D images;then,based on these images,a deep learning method is proposed to generate representations of given 3D models;finally,the classification process and results are given.The main researches are as follows:? Views generation of 3D modelsThe views are the first descriptor that describes the model directly.The number of views and angles of view obtained will have influence on the final classification.Considering that the views obtained by LFD are redundant,which the three views are too simple to lost the spatial information of models,we need to study the appropriate method for obtaining views.In this thesis,two kinds of views extraction techniques are used to obtain mixed views and then input the mixed views into network so as to improve classification accuracy.? Construction of deep neural networkBecause convolution neural network shows high processing ability in field of computer vision,this thesis uses this kind of deep learning models in 3D CAD model classification.The deep network constructed in this thesis consists of multiple layers: one input layer,several latent layers and one output layer.The network uses the extractedfeatures as input,then the higher abstract conceptual features are extracted and synthesized through several hidden layers,and the category of the models are generated and outputted through the output layer.? Selection of classifier of output layerThe last layer of the convolution neural network is the output layer.The output layer classifier generally selects Logistic regression or SoftMax regression.However,Logistic regression is just suitable for binary classification problems,while there are many classes in CAD database.In order to solve the multiple classification problem,we introduce SoftMax regression model in this thesis to generate multiple classifications.