Research And Application Of Several Shape Feature Extracting Methods In 3D Model Retrieval

With the development of 3D model acquisition, modeling methods, and hardware technology, 3D models are more and more widely used in many areas. Not only increasing number of 3D models are produced, did the quantity and scale of 3D model databases. Since constructing a new 3D model is a time-expending task as well as a energy-consuming job, it becomes more and more important to reuse the existing 3D models. In order to fully make use of the existing model resources, and find the models that one needs accurately and efficiently, the research on building the 3D model search engine is an urgent issue.A complete 3D model retrieval system typically includes feature extraction, similarity matching, index structure, and query interface. Among them, feature extraction is the most important for the 3D model retrieval. Hence 3D model feature extraction is the key technology in 3D model retrieval, and it is also the focus of this paper.The main job of this paper is to research and implement the technology of 3D model feature extraction, the innovation is that it proposes and implements three new methods of feature extraction:1. The first method proposed in this paper is a 3D model geometric shape matching based on 2D projective point sets. It is different from the method of Min which compared the 3D shapes based on 2D contour map, also different from the means of Loffer which used the technology of of 2D image retrieval. This method compares the 3D shapes by measuring the statistical characteristic of 2D projective point sets, and it has low complexity. It is the first innovation of this paper.2. Make use of multi-feature weighted distance to match the 3D models. Combine two characteristics, which is: the boundary feature of 2D projective point sets the former method had extracted, and the vertex density of triangle mesh of 3D models. The technology of the multi-feature weighted distance combining 2D boundary feature with 3D vertex density is the second innovation of this paper.3. Introduce the curvature of discrete points into the 3D models matching. Extract the boundary conture of 2D projective point sets, compute the feature which is the product of the two: the curvature of discrete points on the boundary conture, and the distance between these points and the center. This method is the third innovation of this paper.This paper is organized as follows:In chapter1 we first present the application perspective and the significance of 3D model research engine study, as well as 3D model research technology and 3D model retrieval based on the shape feature, then review the current 3D model search technology, summarize the fruit in shape feature extraction and similarity matching. We also give some examples of typical 3D model search engine. Finally we introduce performance evaluation methods of the retrieval system.In chapter2 we describe the necessary of normalization, then introduce the method of normalization preprocessing from three aspects which are transformation, rotation, and scale.In Chapter3 a 3D geometric similarity matching algorithm based on 2D projective point sets is proposed, which comes from the statement"If two models are similar, they also look similar from all viewing angles."Based on this, we first project 3D model onto 2D planes, then extract the feature of 2D projective point sets. By calculating the distance between shape descriptors of 2D projections, a faithful 3D similarity measurement is achieved. The feature of 2D projective point sets mentioned here refers to the maximum distances between 2D projective points and the center in every areas after being segmentated by sectors. Experiment results show that this method achieves satisfactory performance for the rough classification.In chapter4 we first point out the deficiency of the former method which only compare the boundary feature of 2D projective points and just suitable for rough classification, then bring forward a method using multi-feature weighted distance. We compare the boundary feature of 2D projective points first. Secondly, we compare the feature of vertex density of 3D models. Then combine the distances of the two features to obtain the final distance. The experiment results show that the multi-feature weighted distance combining 2D boundary and 3D vertex density has a better performance than the former.In chapter5 we first introduce the concept of curvature and curvature calculation of the discrete points. This method also projects 3D models onto 2D planes. Then extract the boundary contour of 2D projective point sets. Calculate the curvature of every points on the contour, and the distances between the points and the center. Then regard the product of the two as the feature descriptor of 2D projective points. Experiments show that the feature extraction with curvature has a better performance than the former approaches.We conclude the whole thesis in Chapter6, with a brief discussion of future research directions.