Research On Feature Extraction, Comparison And Fusion Methods For 3D Model Retrieval

Along with the development of three-dimensional technology, the acquisition of 3D model has become more and more convenient. This tendency brings a rapid growth to the amount of 3D model resources. Thereby determining how to effectively browse and retrieve these resources, and then assist users to reuse the existing mature cases, has attracted increasing attention in recent times. As one of the common forms of 3D data in Engineering, triangular mesh model itself has lots of retrieval needs; such needs will become more urgent with the development of 3D printing and 3D numerical control technique. This thesis focuses on the retrieval technology for triangular mesh models, and carries out a systematical research on 3D model retrieval from its three key aspects--feature extraction, similarity comparison and multiple feature fusion.The structure of this thesis is arranged as follows:1) A compound-eye-vision-based local feature extraction method for 3D models is proposed. Following the principle of insect compound-eye vision, the method represents a 3D model as a spherical image, so as to transform the feature extraction of the 3D model into the feature extraction of the spherical image. Since the projection for the 3D model from multiple viewpoints is not required, the proposed method fundamentally avoids the shortcomings(e.g. informationredundancy) of the existing human-vision-based methods. According to the characteristicof the spherical image, aspherical SIFT feature extraction algorithm is designed for extracting the salient local features on the spherical image. These features are then utilized for describing the content of the original 3D model. The experimental results show a better retrieval performance is achieved by our method compared to the other competing methods.2) A 3D model feature extraction method using local shape distributions is presented, which can implement a more detailed description for 3D models during extracting their features. Based on the concept of probability and statistics, the method starts by randomly sampling on a 3D model so as to divide the surface of the model into a series of sub regions; the local shape of each region is then represented as a local shape distribution histogram. The experimental results show the proposed method, with a good efficiency of implementation, is effective in reflecting the detailed difference between two 3D models during their comparison.3) A novel 3D model similarity comparison method is proposed, which applies the spatial relations of the local features of 3D models tothe similarity measure of different models. Given a 3D model, the method begins with building a spatial-relation graph to record the spatial relations of its local features; the graph is then converted and deconstructed, so that the shape-word cliques hidden in the graph can be revealed; the similarity comparison of 3D models finally is transformed into the likeness calculation of their corresponding shape-word clique histograms. The experimental results show the retrieval accuracy is lifted by using our method, with its effective utilization of the spatial relations of 3D model local features; the method also displays a good efficiency to meet the practical needs of retrieval applications.4) An optimal fusion method for multiple features of 3D models is presented, which addresses the selection problem of the optimal features for different 3D model retrieval scenes. Initially, a modified Partial Swarm Optimization algorithm is proposed, which is equipped with a supervisory mechanism to prevent the premature convergence of the algorithm; by building on such algorithm, we then apply a supervised learning mode toseeking the optimal features and calculating their optimal weights in sequence using the specific training sets for different retrieval scenes, so as to determine the optimal feature fusion scheme for these scenes. The experimental results show the effectiveness of the proposed method in fusing multiple features conforming to the characteristic of different 3D model retrieval scenes.Based on the above work, a 3D model retrieval prototype system--Northwestern Polytechnical University-Model Retrieval Prototype System(NPU), is developed so as to examine the proposed methods.Its functional architecture and some running results are presented in this thesis.