Researches On Mold-based Feature Extraction And Retrieval Technology Of 3D Models

Nowadays, a huge number of 3D models are available with the development of 3D scanning and modeling technologies. And many 3D model databases have emerged, such as NIC database. In other words, we are living in a world containing 3D models. Needless to say, it is a pressing and practical problem in finding the deserved one from the vast number of models. The related works have been developed after the three-dimensional models retrieval technology came into being. Current researches focus on content based retrieval methods. The feature extraction of 3D models is considered as the key point of content-based retrieval and many methods have been proposed. Generally, these methods can be divided into three categories: image-based methods, methods based on surface properties and shape-based methods. However, the studies show that even the method which is relatively better than others does not apply to all models. Therefore, it is still an important subject to find out more efficient feature extraction method. As most of former practices are based on attributes of the model itself, it is promising direction to utilize the categorization information in feature extraction. Research following this direction has just begun. In 2007, Shilane P. proposed to obtain local descriptors of each model by using the categorization information; however, it has relatively high computation complexity, because the extraction of local descriptors performs on each single model and it is necessarily to match the whole local descriptors of every two models. In contrast, retrieval technology with learner uses category information during the process of features matching. The method based on category information, which performs dimension reduction on content features of model, is also a subsequent process of models'feature. Theses two technological roads do not directly address the process of feature extraction.The first chapter is introduction part. The background and the sense of the research of this article is described. The introduction of the main task of this article isthe most important part. The second chapter is about the correlative content of shape descriptor .The shape descriptor is only a part of the system. One needs to know the whole system if he want to research on shape descriptor more deeply. This chapter contains 5 parts. In the first part is summarizes the methods of 3D model shape extraction; the second introduce the method for evaluate the descriptor, such as precision-recall plot, the nearest neighbour, the first tier, the second tier; the third part introduce the Princeton Shape Benchmark database; in the fourth part the two methods of representing the 3D models-the voxel and the mesh are introduced. At last, in the fifth part, the necessity of the 3D model predispose. And introduce the method and step of a common method -Principal Component Analyze.From the third part to the sixth part the article focus on realizing and improving three important 3D model descriptors and implementation.In section 3, three traditional algorithms of feature extraction are studied and implemented, including Shape Distribution, Histograms and Spherical Harmonics. Robert Osada et al proposed shape distribution algorithm which is a simple and effective algorithm to measure the similarity of between 3D models. The main idea is to measure 3D models using shape distribution functions, in another word, using the probabilities of distances between two random sampling points on objects'surfaces. The advantage of this algorithm is simplicity compared to others. Furthermore, this algorithm can measure the similarity of objects only by small amount of data obtained on surfaces. Ankerst et al get points cloud by sampling from objects, then they construct the histograms of distribution of points cloud in order to perform retrieval. Firstly, they construct the histogram of distribution of 3D model points by dividing the region around the 3D model into different sub regions and making statistics of points contained in each region. Secondly they construct histograms using the statistics to perform retrieval. In addition, the algorithm of spherical harmonics is also achieved in this paper. In this section, the math principle of the sphere harmonics and the reason why this algorithm has the rotation invariance are both explained. In this paper, the features of model are extracted by using spherical functions and spherical harmonic function, which have the advantage of rotation invariance. According to its stability and simplicity compared with former ones, this algorithm is adopted in the next step of our work as the foundation.In section 4, the solution of mold is explained in detailed. A new way for representing the contents of models is explored. According to the fact that there are many similarities and little differences among the models in the same category, we propose that the contents of models in one category can be represented by an elastic mold. The more rigid parts of the mold shows similar parts of models in the same category, while the elastic parts show the dissimilar parts. First, the concept of mold and its data structure have been proposed in this section. Secondly, the feature extraction is performed by using spherical harmonics. After the screening of the mold, the optimal molds with optimal probabilities were selected by DCG algorithm. Lastly, the deletion of molds and dimensional reduction are both discussed.In section 5, the 3D models retrieval strategy based on mold has been proposed. In this section, all the experimental results are also showed and analyzed. In this paper, we preserve the optimal mold of models in mold database according to the category information. Some molds which are not qualified are deleted from the molds database. For a query object, the similarity between its and each mold is evaluated firstly. The possible category could be determined in this step by returning l categories as selective collection. After that, fine selection is performed by using the most distinctive dimensions among k1 categories as retrieval results. According to this strategy, system will evaluate the feature vectors of query object submitted by user, and then perform matching between this one and each mold. The nearest k1 molds will be returned as the representatives of their models. These models coming from k1 categories are performed a further retrieval together with deleted models. The nearest k2 models will be returned as final retrieval results. Comparing to the traditional method, which is one-to-one, the impact of specificity of model can be avoided and performance can be improved during the retrieval process. From the experimental data and the retrieval results, there is a great improvement by using mold-based strategy, which proves this strategy a new technological road in 3D model retrieval.Section 6 contains the summaries of current work and future work.This paper has the following characteristics: First, all experiments were performed on PSB which is generally acknowledged. All the evaluation indicators used in experiments are universally adopted. Thus, our performances are easily compared with other institutions'. Secondly, the work is aimed to the representative of models within one category, which is greatly different from other current methods. Although there are some shortcomings in mold, our method still has following advantages: effectiveness, intuition and simplicity. Our work also makes a foundation of related work.