Study On Key Techniques Of Content-Based 3D Model Retrieval

With the development of computer graphics, the number of 3D models grows rapidly recent years. To help people effectively manage and access numerous 3D models, we need to research the methods and systems for 3D model retrieval.Content-based 3D model retrieval techniques measure the similarity of 3D data content by extracting the information that can describe the feature of visual appearance of 3D models. This feature information is expressed as a compact data structure, which is often called feature descriptor. Feature descriptor extraction is the kernel of 3D model retrieval systems. The query interface and relevance feedback mechanism are also important parts of 3D model retrieval systems. This dissertation studies feature descriptor extraction, query interface and relevance feedback mechanism of 3D model retrieval systems. The innovation of this work is as follows.(1) Present methods for building freehand sketch query interface of generic 3D model database and mechanical 3D CAD parts database.Query interface plays an important role in 3D model retrieval systems. A usable 3D model retrieval system needs to have a user-friendly query interface. View-based feature descriptor is the basis of implementing freehand sketch query interface. The key steps to build a view-based feature descriptor include: determining projection directions, generating views and comparing views. This dissertation studies several algorithms of determining projection directions, generating views and comparing views. Especially, the retrieval performance of two rotation normalization algorithms and two types of views is analyzed. Based on the analysis, methods for building freehand sketch query interface of generic 3D model databases and mechanical 3D CAD parts databases are given. It is found that PCA rotation normalization algorithm and outer silhouette are appropriate for generic 3D model database, and MND rotation normalization algorithm and silhouette are appropriate for mechanical 3D CAD parts database.(2) Presenting a feature extraction algorithm based on voxel mode. The rotation invariant feature descriptors built by this algorithm can achieve better performance than that of Osada's D2.View-based feature descriptors can implement freehand sketch query interface, but view-based feature descriptors are rotation variant, rotation normalization being needed. Furthermore, view-based feature extraction is a kind of algorithm that works in image space, thus it does not answer the problem of how to extract feature directly from 3D models. This dissertation proposes a voxel-based rotation invariant feature extraction algorithm. For containing the information of volume of 3D objects and converting irregular space signal into regular one, voxel model can facilitate the shape analysis and feature extraction process. Firstly, a voxlization algorithm to convert 3D meshes into voxel models is proposed. Then, two voxel-based shape distribution descriptors (D2V and DIR) are constructed, both of which are rotation invariant. D2V eliminates the defect of Osada's D2 that histogram computation can be biased by model's local part with high dense area. DIR obtains a shape distribution feature different with Osada's D2 by computing the distribution of inner line segments and outer line segments. Combining DIR and D2V can obtain more shape information about 3D models and enhance the retrieval performance.(3) Present two algorithms of relevance feedback to improve the retrieval performance of 3D model feature descriptors.Relevance feedback is an important way to improve the retrieval performance of multimedia retrieval systems. Recent years 3D model retrieval systems begin to use relevance feedback to get better retrieval performance. This dissertation proposes two relevance feedback algorithms and uses them to improve the performance of the feature descriptors. The first relevance feedback algorithm employs user feedback information to build a new query that is closer to user's high-level semantics. Firstly, a set of bounding spheres that covers all relevant models is built. Then, in the next iteration, the bounding spheres are used to match with the feature vectors of models. For the bounding spheres contain user semantics, the retrieval precision will be improved after the iteration. Experiments show that this algorithm can effectively improve the retrieval precision of D2V. The second algorithm employs feedback information to combine multiple feature descriptors. This algorithm determines the weights of feature descriptors dynamically according to the distribution density of the feature vectors of relevant models. Experiments show this algorithm can optimize the weights for combining feature descriptors.Query interface, model feature extraction and relevance feedback mechanism are three main parts that compose a 3D model retrieval system. The model feature extraction is the basis, which query interface and relevance feedback mechanism depend on. The freehand sketch query interface implements the comparison of user's query sketches with model's views. Relevance feedback mechanism is to improve the retrieval precision of feature descriptors and narrow the gap between the low level feature of models and the semantics of user's query.Finally, this dissertation makes conclusions and suggests several related future work. The work of this dissertation is supported by NSFC (60573146)...