Feature Extraction And Retrieval Of 3D Model

The number of 3D models has gained an increase at geometric rate in recently ten years, withal the great development and widely popularization of Web convenience the exchange of multimedia information, therefore it is imperative to develop effective 3D model retrieval systems. One of the kernel problems of 3D model retrieval is"feature extraction". While compared with image retrieval, the spatial structure of 3D models is more complicated, which makes it difficult to find suited descriptor to represent them. As a result, the known features at present have difficulties in satisfying the need of practical retrieval and there aren't also very ripe 3D model retrieval system provided.Features of 3D model can be divided into two groups: rotation-variant ones and rotation- invariant ones according to the criterion whether they need rotation normalization. From the current researches, the former overperforms the latter on the whole. While for some irregular model, rotation normalization is easy to fail, the latter holds the predominance instead on this occasion.Under the support of The National Natural Science Foundation of China: On the Technology of 3D Model Retrieval Based on Spherical Wavelet(60573146), this paper deeply investigates and probes into both types of features above. Firstly this paper analyzes the two pre-processes(rotation normalization and isotropization) for 3D model retrieval before feature extraction, and presents two rotation-variant features: depth images and region entropies, based on rotation normalization. These two features are respectively applied on different kinds of 3D models and enhance the performance of retrieval. Moreover, to overcome the shortcoming of their serious dependence on rotation normalization, a spherical images based rotation-invariant feature is presented. According to above researches, a 3D model retrieval system is developed with preparatory application value. The main research fruits and innovations of this research can be summarized as following:1.Improve the robustness of rotation normalization and greatly simplify the process of isotropization. This paper firstly emphasizes to investigate the way to improve the robustness of PCA-the tool for rotation normalization. According to the basic principle of PCA, this paper summarizes three possible unsuccessful cases and in allusion to each one, presents our corresponding solution to reduce the probability of loss and eliminate the negative effect of loss as possible as we can. Since isotropization can help to improve the performance, this paper demonstrates the essential relation of isotropization with PCA, hereby presents the idea of isotropization using the PCA matrix, and adopts an iterative strategy to realize isotropy.2.Present a depth images based algorithm for feature extraction, which gains a balance between performance and time consumption. Firstly the 3D model is syncopated along with its three PCA coordinate planes, and then each part is projected onto the corresponding plane of the model's cubic bounding box. Two measures: Moment Invariants and Polar Radius Fourier Transform are unitedly used to extract the feature of depth images. To get a better comparison, the six images projected respectively from X-Y, Y-Z, Z-X coordinate planes are endowed with different weights according to the eigenvalues of its covariance matrix. Experiments show that this algorithm can obtain nearly the same performance as Lightfield does while costing far less time. For those comparatively regular models, this descriptor will exhibit preferable results.3.Represent the 3D models which are composed of scattered trigonal facets with regular signal using voxelization as a solution, and in this way many classical tools for signal analyse can be applied. For binary voxelization, a fast 3D triangle-box overlap testing algorithm based on bit operation is used, which greatly quickens the speed. To approach the 3D model in a more accurate manner, this paper presents the conception of"voxelization with grey"with its baic idea being to map the exterior area contained in a voxel into its grey. Also this paper analyzes its applications in visualization. Based on voxelization with grey, an algorithm for feature extraction using regional entropy is presented. Firstly the whole voxelized 3D model is divided into several regions and then the entropy of each region is calculated, finally the feature is constituted according to these entropied with some manner. Experiments show that this algorithm has strong power of differentiation and clustering, which can be applied to retieve comparatively complicated 3D models.4.Present an effective algorithm for feature extraction based on spherical images. As a rotation-invariant descriptor, it needn't PCA in advance, so it is suited to retrive those 3D models which may get unsuccessful rotation normalization. Firstly an icosahedron is mapped onto the sphere to obtain a spherical basic mesh, and then the mesh is 1-4 divided several times to form the format of spherical images. Whereafter the spheres are intersected with the voxelized 3D model and a series of homocentric spherical images in different radius are got. Afterward these images are decomposed using butterfly wavelet and the resultant 20 coefficients are regarded as the feature of each image. Finally, the coefficients of all spherical images are united to form the feature vector of the original 3D model. Experiments show that this algorithm also obtains prominent improvement compared with Shape Histogram and Shape Distributions. To solve the problem of slow-retrieval mainly caused by high dimension, a super-sphere based structure of the library of feature vectors is adopted to eliminate those obviously dissimilar 3D models at online retrieval process, which greatly quickens the speed of retrieval.3D model retrieval is a hotspot direction in the field of Computer Graphics, this paper mainly carries out four aspects of research around feature extraction: deeply analyzing the two pre-processes and presenting three kinds of feature descriptors to satisfy different kinds of retrieval. The contents are described in respective chapter while maintaining relation with each other in a more and more in-depth manner. Based on the above theoretical fruits, a 3D model retrieval system is developed and gets preferable results. Finally this paper gives the conclusions and prensets the assumption of further research.