The Implementation And Researches Of 3D Model Retrieval System Based On Semantics

We live in a three-dimensional world; three-dimensional models are the most intuitive way to understand the world. In recent years, the continuous improvement of computer performance, as well as software that used to generate three-dimensional models made a continuous development, resulting in appearing a large number of three-dimensional models. How to identify the target model from a three-dimensional model base becomes a pressing issue. Research to the problem contribute to rapid model generation as well as building three-dimensional scenes, so in the medical, industrial manufacturing, virtual reality, War simulation areas has broad application prospects. For example in the field of medicine, we can significantly improve the opportunity to discover new drugs through the search of similar to the molecular structure.3D model retrieval technology came into being. It has become a hot research topic. Its goal is "accurate, rapid, and convenient" to help users find the relevant three-dimensional model. 3D model retrieval is generally regarded as a part of multimedia information retrieval by researchers.In early state, the research related of three-dimensional models always aimed at a specialized field of model, such as CAD models, the protein model. The first article about 3D model retrieval arise before and after the emergence in 1997, followed by the United States, Germany, Japan, many Chinese universities and research institutions enter this field.This paper is mainly about method of 3D model retrieval, based on the content and semantic-based , especially semantic-based retrieval. We present a new method of semantic retrieval through the research about the relevance among models, and then build a content-based and semantic-based 3D model retrieval system.Chapter I introduced the history of development of the 3D model retrieval technology, as well as the status of the current study, briefly introduced the two main methods of the 3D model retrieval :based on content, based on semantics; focused on the main structure of this article and research significance.Chapter II is mainly divided into two parts, focus on the process of content-based and semantic-based retrieval of 3D model retrieval. The first part includes 5 sectors, used to carefully introduce the knowledge of content-based retrieval. Section I introduce the process of the content-based 3D model retrieval, the major search objective: the value of three-dimensional model geometry characteristics, the main steps: preprocessing, feature extraction, compare, return result. The feature extraction is the main factor that affects the retrieval accuracy. In the second section, we introduce several eigenvalue extracting method of three-dimensional model. In the third section we introduce a variety of criteria that used to evaluation the effect of the eigenvalue used, including: Recall, Precision, R-Precision indicators. In the fourth section, we introduce the other steps of content-based retrieval rather than feature extractions, include the process of implementation and their impact on the retrieval. The fifth section is mainly about the sources and characteristics of PSB model base that used in the research of 3D model retrieval. The second part of this chapter is mainly based on the introduction of the research in the semantic retrieval. The Web search engines provide similar retrieval method. It is the simplest, and the most convenient retrieval mode. However, during semantic retrieval process needs the computer to identify the semantics, and the models have to be tagged perfectly. So currently, the effect of semantic-based 3D model retrieval is unsatisfactory.Chapter III introduced the motivation and main functions of our 3D model retrieval system. Our 3D model retrieval system has been mainly used to display the technology of 3D model retrieval, provide a development platform for the research of 3D model retrieval. The design of the content-based retrieval of 3D model, seek to achieve modular, so that can be facilitate future upgrades, and carry out experiments to facilitate easily. Of the system, we propose a new kind of semantic tagging method in the process of semantic-based retrieval process, in Chapter IV for details.In chapter IV we detailed descript the main implementation process of model relevance based retrieval method. The relevance of the model are defined from the people's understanding of the model , Some models have a semantic correlation between each other means the majority of people think that in the process of retrieval they should appear in the search results together. The stronger of semantic correlation between models, the higher probability of these models appear in the search results together. Since the correlation between the models is defined from the perspective of person, so the process we obtain the correlation of the models must have persons involved. We developed a model relevance collection system, through 3 testers to identify the original model base 907 model's correlation, and then we analysis the test results. Found that people's understanding of the model difference because of the differences in perspective on the relevance of the models, but after aggregated the search results, we can see that indeed exist model group because of the relevance between the models, and this situation can be generally accepted by people. In order to use the relevance of the models in the model's automatic semantic tagging, first ,we use the x-means clustering algorithm to the original model base, the model based on the model database is divided into a number of relevant model cluster, the models in one model cluster with the character that the correlation between each other are higher. Then, through tagging some models manually, achieve the auto-tagging of all the models of the model base. The automatic semantic tagging is used in the system, when the retrieval term was entered, first, the system search the term in the word list of tagging manually, if matched, the system automatically return the models tagged with the term; if manually tagging semantic word matching failure, then the system return the three words that was judged with similar semantic by the WordNet. When the result models are returned, the system based on the similar level of three words and the retrieval term, the higher level of similar, the more correspond modules are returned.Chapter V present the content-based retrieval method used in the system. In the process of the content-based retrieval, we have adopted the combined feature DSR472 that discovered by Germany CCCC Group to extract the model's eigenvalue. DSR472 is combined of the depth of field eigenvectors, the outline of eigenvectors and eigenvector-ray. When we choose the eigenvectors to combine, the vectors should descript the model from different angles, and at the same time they should complement each other. DSR472 combines the three feature eigenvalue: depth of field through the extraction model eigenvalue at x, y, z axis projection of plus or minus direction, resulting in six high projection of the model's surface; based on the outline of the models , through three direction projection contour of the model; Finally eigenvector of ray-based launch many ray from the centroid of the model to all angles, and the ball surround the model collect the projection points, provided three-dimensional outline information of the model. Three eigenvalues of the model describe the different characteristics of the model. In the process of DSR472 is computed, in accordance with the capacity of the three eigenvalues describe, depth of field eigenvalue selected 186 bits, select the outline of the 150 bits eigenvalues, select ray Eigenvalue 136 bits, constitute the DSR472-dimensional eigenvalue. The system extracts the DSR472 of the model after the model was uploaded, then the system compares the model's eigenvalue of the models in the model base to the eigenvalue of the object model and returned the first 16 similar models.In this paper, We developed a 3D model retrieval system based on Princeton PSB generic model base, presented a new method of automatic semantic tagging based on the relevance of the models, and can greatly improve efficiency, get better retrieval effect. In the process of the content-based retrieval of our system, we used the DSR472 to extract the model's feature eigenvalue and get a high accuracy.