Dissertation > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device

3D Model Retrieval Based on Segmentation and Multi-features Integration

Author LiDeLiang
Tutor QinMaoLing
School Shandong Normal University
Course Computer Software and Theory
Keywords 3D model retrieval model segmentation flatness statistical distribution multi-features integration feedback
CLC TP391.41
Type Master's thesis
Year 2009
Downloads 179
Quotes 0
Download Dissertation

As a new media, the application of 3D model becomes more and more widely. With the improvements of the 3D modeling tools, 3D scanning device ,hardware and software technology, a large number of models are produced and have been spread every day. So effective management and retrieval on 3D model is urgent.3D model retrieval is widely used in people’s daily life, CAD / CAM, computer animation design, and it has become an important research component of standard MPEG7. Research on 3D model retrieval has important practical significance, and it can greatly promote the development of other related fields.The content-based 3D model retrieval aims at extracting some kind of shape features in a vector or a map, according to 3D model’s geometry, spatial relationships, statistical characteristics, texture, materials and other information of the 3D model’s source file stored. We can use these characteristics mark a 3D model only. The 3D model retrieval process is to find the first K models which have the smallest distance with query model in characteristics database of the 3D models and then output these models in accordance with the similarity. Therefore, the shape feature extraction algorithm is the core field in 3D model retrieval research.This paper proposed two improved retrieval methods based on the analysis of the current retrieval systems and the summary of the existing algorithms. The main work is as follows:Firstly, we introduced the framework and the process of 3D model retrieval system, according to consulting a large number of native and foreign documents about 3D model retrieval. Then we summarized the existing feature extraction algorithms with analysis and comparison of the different algorithms.Secondly, this paper proposed a new retrieval method which is based on 3D model segmentation. To be aimed at the question that current retrieval algorithms did not fully consider partial information of 3D model, we use model segmentation to utilize 3D model’s partial information. Firstly, we obtain a stable signal calculation method by comparing different methods. Then we use stable flatness signal as height function for the watershed segmentation. In merger process preventing over-segmentation, we adopt multiple rounds of mergers based on the dynamic weights, in order to make the segmentation results suitable for retrieval requests. Finally, we model the 3D model to be a mesh map, and compare the similarity of 3D models on the basis of graph matching. At the same time, because the flatness signal in segmentation process can fully represent 3D model’s shape characteristics, we extract flatness-based feature eigenvector from idea of the statistical feature extraction methods.Finally, we proposed a method of multi-features integration. Because single feature is not enough to describe the shape of 3D models and it can not retrieve all 3D models effectively, we use multi-features integration to complement each feature’s advantages and enhance retrieval algorithm’s adaptability. In order to measure each feature’s retrieval performance, we assign different weights to each feature. Because the retrieval performance will be different when query models are different, to each query models, we set up a weights knowledge base for all the models in model database, using the idea of equivalent substitution. The formation of weights is based on user’s feedback. On the base of initial weights, through a certain number of user’s feedback training process, weights knowledge base can be stabilized. Then according to the weights, we integrate different features to retrieval.

Related Dissertations
More Dissertations