The Research For 3D Model Retrieval And Feedback System

With the rapid development of multimedia information technique and three-dimensional (3D) technique, the research of 3D model retrieval system has become a hotspot of the academe, as well as an important part of the context of MPEG-7 standard. This technique is widely used in many fields such as CAD/CAM, computer animation and game design, and has a bright future. The text-based retrieval technique has some limitations, so the content-based retrieval technique is an important issue in this field. The precision, speed and robusticity are three most important criterions in measuring a 3D model retrieval system.A typical 3D model retrieval system has three modules: a pre-computation module, a retrieval module and a feedback module. This paper will place emphases on the algorithm and system design of both retrieval and feedback modules.The retrieval module is used to compute the similarity of 3D models, and its core issue is the method for geometrical feature extracting. So far many methods have been presented, such as outline-based algorithms, visual-based algorithms, and topology-based algorithms. After comparing many methods, in this paper, we present a new 3D model feature extracting algorithm based on feature binary tree and Zernike descriptor. This algorithm firstly computes the model's projections on a series of homocentric sphere, then computes the 3D Zernike descriptors of these projections, at last matches the models with the feature vectors of Zernike descriptors. Experimental results show that this algorithm can effectively improve the precision and robusticity.The feedback module is used to analyze and learn the result of first retrieval round, to help us improve the precision after several feedback rounds. At present the common feedback methods are traditional relevance feedback method, and feedback method based on Support Vector Machine(SVM) or Neural Network(NN). The former is much more mature and easier to implement. This paper presents a new relevance feedback method which combines model similarity and texture similarity. The latter is complex both in theory and implementation. In this paper we design a new method based on SVM, whose main idea is using probability to predict the class of a model. The result of this algorithm is very good.At last, this paper studies the multimedia information searching technique under MPEG-7 framework. With describing the algorithm based on feature binary tree and Zernike descriptor under this framework, we show that our algorithm can meet the requirement of the MPEG-7 standard.