The Researches On The 3D Model Retrieval Techniques Based On Clustering Analysis And Based On Semantic

3D model is a natural and direct way to illustrate the real-world objects. With the proliferation of 3D models, it becomes an emergency task to obtain the desired models from the existing. This topic can facilitate the modeling process, and has great value in manufacturing, military, virtual reality, simulation etc.. For instance, 80% new CAD models can be obtained directly or after minor modification from the existing ones.Therefore, 3D model retrieval emerges as an important field in multimedia retrieval, which aims at retrieving the desired models correctly, quickly and conveniently. Previous researches of 3D model retrieval concentrate on the specific type of 3D model, such as CAD model and 3D protein model. The researches focusing on the general type model begin at 1997 and the research groups in this field come from American, German, Japan and China.The context-based 3D model retrieval is research focus at present due to the drawbacks of text-based retrieval. Nowadays, the context-based retrieval technique has many improvements in the theory and the applications. Researches on the context-based retrieval cover the topics including: (1) the feature extraction and similarity computation, the normalization of 3D models, etc.; (2) the classification and organization of 3D model database; (3) retrieval method and retrieval interface; (4) the construction of context-based 3D model retrieval system.Since the feature of 3D model determines the performance of a 3D model retrieval system, it is widely accepted that feature extraction of 3D model is the key problem of context-based method. Among the proposed feature extraction methods, most are shape-based methods.However, several different research groups prove that no feature-extraction method is perfect. Therefore, the researchers start to combine different 3D model features, while pursuing the perfect feature extraction method. The proposed methods combine different features by determining their weight. But all of these combination methods have their own shortcomings. Moreover, the fundamental topics related to the feature combination are still unnoticed, such as"the more the feature combined, the better"or otherwise.Current researches do not pay much attention to the classification and the organization of 3D models. Researchers categorize 3D models manually, which is not only expensive but also reflecting human's understanding of 3D model. For the organization of 3D model database, current researches use the traditional index which is not suitable for the high dimensionality of the feature vector of 3D models.Since model's shape feature only reflects the physics information of model and can not represent its semantics, the shape-based method doesn't perform quite well due to the influence of Semantic Gap. Using the studies in the image or video retrieval for reference, semantic-based way can greatly improve the retrieval performance. However, this is still a novel topic in 3D model retrieval. Some researches adopt the relevance feedback to improve the performance, but do not solve the fundamental problems, like the semantic representation method of model. To solve these problems, the thesis conducts researches in the context-based 3D model retrieval and the semantic-based 3D model retrieval.For the context-based 3D model retrieval, the thesis concentrates on the application of clustering analysis techniques in 3D model retrieval. The contributions of the thesis are stated as follows:(1) Realizes and analyzes 3 kinds of shape feature extraction methods and their variance. The experiments conducting on the classic 3D model database Princeton Shape Benchmark show that each kind of feature has its limitation which can not be breached through improving the particularity degree of the feature; and no method is best for all kinds of 3D models. These conclusions show the importance of feature combination.(2) Analyzes the influence of the feature weight, the feature type and the classification of 3D model on the performance of feature combination. Then, the thesis proposes the method to automatically decide the weight of different shape features and combines these features selectively; states the feature combination method based on iterative clustering in case no classification information is available. The experiments conduct on PSB show that the combined feature obtains the R-Precision 10%-16% higher than that of the best single feature.(3) Proposes to apply the clustering analysis technique in analyzing, classification and organization of 3D models. The thesis states two key problems for the application of clustering techniques, that is"finding the clustering algorithms which are suitable for the shape vector of 3D models"and"finding the effective way to classify and organize 3D models based on clustering result".(4) Aiming at the characteristics of 3D model's feature, the thesis proposes a novel hierarchical clustering mechanism which integrates with the outlier mining. The clustering methods ASHCA, CURED and AS-ROCK are constructed following this mechanism. The new methods determine the traditional parameter k of hierarchical algorithm automatically and compensate the drawbacks of outlier detection for traditional clustering methods. The experiments conducting on the UCI datasets and PSB prove the good performance of the proposed methods in obtaining suitable number of clusters and detecting clusters with complex shape.(5) Realizes the efficient shape-based classification of 3D models based on the clustering result, which can handle 88% models in PSB. And the index ClusterTree based on the clustering result is applied in organizing the 3D model database. And some improvements are made on ClusterTree. The experiments conducted on PSB show that the retrieval time based on the original and improved ClusterTree are just 35% and 22% of that of the linear scan.For the semantic-based 3D model retrieval, the thesis explores two ways for representing 3D model's semantic. The first is based on the precise semantic of each model and constructs the semantic tree of all models of PSB. The retrieval based on semantic tree returns related results for the 94% keywords of ZOO dataset, while the traditional method just return results for 2% keywords. And the interaction between semantic-based retrieval and context-based retrieval are also discussed in the thesis.The second aims at the difficult situation that no semantic of models is available, and the method tries to obtain the semantic relationship between models from the user retrieval results.On the whole, the thesis pursues the researches of several important topics in 3D model retrieval, and particularly concentrates on the application of clustering technique in content-based retrieval and the semantic retrieval using semantic tree.In future works, the proposed methods in this thesis will be integrated in the efficiency and applicable 3D model retrieval system, in addition to perform in-depth researches based on current works.