基于深度学习和视图的三维CAD模型分类技术研究

发布时间：2018-05-31 17:21

本文选题：深度学习 + 卷积神经网络　；参考：《北方民族大学》2017年硕士论文

【摘要】：三维模型的数量在最近10年间呈现出几何级增长的态势。如何对数量庞大的三维模型进行处理、分析和理解,已经成为数字几何领域研究的焦点。而其中的基础问题则是三维形状的分类。传统的分类方式把人工设计的三维特征用于分类,其方法的优劣完全取决于专家对三维模型及其分类目标的理解和把握。存在主观性强、分类精度低的问题。不同于传统的方法,深度学习算法能够让机器自动学习特征及分类,近年来在图像领域有着优异的表现。视图作为模型的直观信息,符合人类的视觉系统,可以作为深度学习的输入信息。本文拟结合深度学习和视图对三维CAD模型的分类问题进行研究:首先进行三维模型视图提取;之后结合深度学习理论,给出深层神经网络的构建方法;在应用环节,给出分类过程与结果。主要研究工作如下:⑴三维模型的视图生成视图是三维模型的描述信息,本文以视图作为深度学习模型的输入。视图的数量以及获取视图的角度都会对最终分类结果有影响。光场描述符提取的视图存在大量的冗余,而最简单的三视图却会存在丢失模型空间信息的问题,我们需要研究合适的视图获取方法。本文利用两种视图提取技术得到混合视图,输入到深度学习模型。用以提高分类的精确度。⑵构建深度卷积神经网络由于卷积神经网络在图像处理方面具有极高的处理能力,所以本文使用了深度学习里的深度卷积神经网络。构建的深层网络分类器由多层组成:输入层、若干隐藏层和输出层。将提取出来的视图作为输入,由若干隐层来提取和合成更加抽象的概念特征,输出层用来输出模型所属的类别。⑶输出层分类器的选择卷积神经网络的最后一层是输出层。输出层分类器一般选择Logistic回归或SoftMax回归。Logistic回归用来解决二分类问题,而CAD模型数据库不止两个类,SoftMax回归模型是在多分类问题上的应用,故而本文采用了SoftMax回归模型。
[Abstract]:The number of 3D models has shown a geometric growth trend in the last 10 years. How to deal with, analyze and understand a large number of 3D models has become the focus of research in the field of digital geometry. The basic problem is the classification of three-dimensional shapes. The traditional classification method uses the artificially designed 3D feature to classify, and the advantages and disadvantages of the method depend on the experts' understanding and grasp of the 3D model and its classification target. There is the problem of strong subjectivity and low classification accuracy. Unlike the traditional methods, the depth learning algorithm can make the machine learn features and classification automatically, and it has excellent performance in the field of image in recent years. As visual information of model, view accords with human visual system and can be used as input information for deep learning. In this paper, the classification of 3D CAD model is studied in combination with depth learning and view: firstly, 3D model view extraction is carried out; then, combined with depth learning theory, the construction method of deep neural network is given. The classification process and results are given. The main research work is as follows: the view generation view of the 1: 1 3D model is the description information of the 3D model. In this paper, the view is used as the input of the depth learning model. The number of views and the angle at which they are obtained will have an impact on the final classification results. The view extracted by the light field descriptor has a lot of redundancy while the simplest three views have the problem of losing model space information. We need to study the appropriate view acquisition method. In this paper, we use two view extraction techniques to get the mixed view and input it into the depth learning model. In order to improve the accuracy of classification, the deep convolution neural network is constructed. Because the convolution neural network has very high processing ability in image processing, this paper uses the deep convolution neural network in depth learning. The constructed deep network classifier consists of several layers: input layer, hidden layer and output layer. The extracted view is used as input to extract and synthesize more abstract conceptual features by a number of hidden layers. The output layer is used to output the selected convolutional neural network of the class .3 output layer classifier to which the model belongs. The last layer of the output layer is the output layer. Output level classifiers generally choose Logistic regression or SoftMax regression. Logistic regression to solve the two-classification problem, and CAD model database more than two classes of SoftMax regression model is used in multi-classification problems, so this paper uses SoftMax regression model.
【学位授予单位】：北方民族大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.72

【参考文献】