Research On Key Technology In 3D Model Retrieval

With the development of 3D data acquisition and modeling technique and hardware, more and more 3D models are produced. How to retrieve the needed 3D models from the internet and 3d models database is becoming an important issue.Feature extraction is the key problem in 3D model retrieval. Most of the previous methods are about the feature extraction of the general 3D models. In this context, we focus on two aspects: first the feature extraction and similarity measurement of the general 3D models in mesh representation, second the feature extraction of the 3D structure of protein molecule. The existing methods cannot provide the real geometrical feature of 3D models such as the orientation and the area and position of the faces of 3D models. In this thesis, we propose two feature extraction methods for the general 3D models in mesh representation : one is phase-encoded Fourier Transform based on multi viewpoint range image of models, another is complex sphere feature map. Simultaneously, we propose the Hausdorff fractal dimension and Information fractal dimension as the feature of 3D structure of the protein molecule.Firstly, the feature extraction method from Phase-encoded Fourier Transform based on multi-viewpoint range image of 3D models is proposed, which comes from the statement "If two models are similar, they also look similar from all viewing angles.". Two techniques are used in this method: Principal Component Analysis and Phase-encoded Fourier Transform, a set of two dimensional statistical histgoram about the orientation and area of the face of 3D model are used for feature vectors, which are invariant to the translation, rotation, and scaling of the model and insensitive to the noise. Similarity measurement between two models are obtained by using the image correlation techniques. The experimental results show that the proposed method is suitable for the classification of 3D models.Secondly, a feature extraction method based on Complex Sphere Feature Map is suggested. The orientation, area and position of the face of 3D model are mapped ontothe complex point on the unit sphere, then a spherical histogram about the geometrical feature of 3D model is constructed. This method only needs the low computation and storage cost. By using spherical correlation techniques, the similarity measurement can be obtained. The experimental results show this method is good at the recognition of 3D models and invariant to the translation and rotation of the models.Thirdly, in order to avoid the complexity of the existing methods in rigid body alignment and high dimension feature vectors, we propose two fractal dimensions as the feature of the backbone of the protein molecules in the feature extraction for the three-dimensional structure of protein molecules, one is Hausdorff dimension and the other is Information dimension. Compared to the existing methods, this method needs lower computation cost and lower storage, and it is suitable for the pre-classification of 3D structures of protein molecules.