基于八叉树结构的三维体素模型检索

随着VR/AR技术发展以及三维模型的广泛应用,实现三维检索具有越来越重要的现实意义.基于模型的检索较好地保留了模型的空间信息和几何特征,其不仅包含模型的表面信息而且还包含模型的内部属性.但是,基于模型的检索往往存在着高存储、高计算的问题.为了解决该问题,本文研究了三维模型预处理及三维模型表示的方法,提出了一种基于八叉树结构的三维体素模型检索方法,即将模型进行体素化处理后提取模型的粗粒度特征和细粒度特征,将两种特征进行融合用八叉树形式表达特征,输入到卷积神经网络中进行训练,最终通过特征的欧氏距离度量实现模型的检索.运用八叉树特征表示法,可以有效地节省体素化存储过程的空间占用量,而且也能保留原始三维网格模型的细节信息.同时考虑到计算性能,本文还在模型体素化的过程中做出一定的改进,通过仅对模型外表面进行体素化,实现了对体素化过程以及数据存储和卷积神经网络训练的优化,大大降低了时间开销.实验中将三维体素模型特征存储在八叉树结构中作为卷积神经网络的输入,结合SOFTMAX代价函数,通过大量的模型训练数据,对该卷积神经网络模型进行训练.与其他同类算法对比,证明了该算法在三维模型检索中的优越性.

3D Voxel Model Retrieval Based on Octree Structure

With the development of VR/AR technology and the wider 3 D applications,it is realized that 3 D model retrieval is becoming more and more important.Model-based retrieval preserves the spatial and geometric features,which includes not only the surface information but also the internal properties of the model.However,there are concerns in relate to its high storage and high computation.Deep learning has demonstrated successful breakthroughs in the fields of speech recognition,graphic image classification and natural language processing etc.In this paper,after studying the 3 D model pre-processing and 3 D model representation,a method combined with 3 D voxel model and octree structure is proposed for 3 D model retrieval.First of all,the coarse-grained features and fine-grained features of the voxelization model are extracted.After the fusion,the features which expressed in the form of octree are input into the convolutional neural network for training,and the Euclidean distance is the metric for evaluating and retrieving the model in the end.In order to form an octree for storing the 3 D model,eight equal cubic meshes are able to be divided after the 3 D model is scaled and aligned with a standard unit 3 D boundary cubic volume.Such a mesh process will continue for each cubic volume,which include the 3 D model,until the mesh quality reaches the requirement.By using the octree feature representation,not only the storage consumption is effective reduced due to the process of voxelization,the details of the original 3 D mesh model are also preserved.The presented algorithm uses the improved Octree structure as the basic data structure of the model voxelization which is applied to the convolutional neural network for model classification.By designing a novel spatial octree,a 3 D model is represented by which surface information was stored into the leaf nodes of the octree.The leaf nodes are able to be trained as initial data and evaluated through the improved octree neural network structure on GPU.IO-CNN is able to support various CNN structures with different 3 D representations to extract and classify the 3 D model for 3 D model retrieval.With careful analyzing of 3 D model,it is found that it is unnecessary to process interior part of the 3 D model for voxelization if it is a closed 3 D geometry.The voxelization of the interior part of the 3 D model will never affect the representation of geometric features.In considering the computational performance,such modifications were made during the process of model voxelization.After voxelization,the normalized 3 D model is represented by octree for obtaining the spatial information.An iterate process is carried out which 1 is set to the region which includes the 3 D model and 0 is set to the region without the model respectively.By only voxelizing the outer surface of the model,the computational overhead is greatly reduced due to the optimized data storage and convolutional neural network training.The experiment has shown,with applying the SOFTMAX cost function,after a large amount of training data through convolutional neural network,the presented algorithm has more advance in 3 D model retrieval than other similar algorithms.