Three-dimensional Model Feature Extraction Technology Research

With rapid growth in the amount of 3D models, the research on 3D model retrieval technology is becoming increasingly important. To meet the requirements of retrieval in 3D model massive database, this dissertation investigates the key techniques of the 3D model retrieval technology including low-level feature extraction, high-level semantic feature annotation, similarity measure and relevance feedback and so on. The research has important academic significance and application value.The main works are listed as follows:(1) A novel algorithm to extract the curve-skeleton of 3d model is introduced. The algorithm firstly calculates the 3D vector fields of models represented by discrete unit, and then extracts the hierarchical curve-skeleton based on topological features of the critical curve and critical points of the vector fields. The similarity among 3D curve-skeletons is measured by using an improved Earth Mover's Distance (EMD) algorithm. The curve-skeleton extracted with this novel algorithm can be used to categorize models and implement global matching and partial matching.(2) A new architecture for extraction of 3D model features using probabilistic density estimation of local surface features is proposed. With the set of 3D local geometrical features, the local feature density of a chosen target point is evaluated using probabilistic density estimation methods. The 3D model can be described using the feature vector comprised of all local feature density values. The single-variate and multi-variate descriptors of 3D mesh model supports for the implementation of 3D model retrieval. The results show that the retrieval performance the method is better than that of the statistical feature extraction methods.(3) Two methods implementing Semantic Force Relevant Feedback(SFRF) and Pure Semantic Relevance Feedback(PSRF) are presented.3D models can be regarded as interactive charged particles. The quantity of electricity depends on the users'judge against the semantic similarity among the models. The charged elements apply forces to each other in a way that semantically clusters are formed and the retrieval quality is enhanced. The two methods were used to implement 3D model retrieval. The outcome shows that SFRF algorithm outperformed the Feature Space Warping and the PSFR algorithm, also illustrated very good performance on the performed experiments.(4) A semiautomatic semantic annotation scheme for 3D models is proposed. The training examples are selected based on the content of a 3D model. A ncurofuzzy controller set is used to estimate the attributes (categories) of each database model using knowledge obtained from manual annotations of objects suggested by the system. Additionally, two relevance feedback methods were modified and integrated for supporting and enhancing the annotation procedure. The proposed framework induces a substantial acceleration of the annotation process. The experiments results show that the proposed method is superior in terms of efficiency for the automatic, semantic annotation of 3D model databases.This research work was supported by Natural Science Foundation project of China "research on the key technology of semantic annotation and ontology based retrieval for 3D cultural model (No.60873094)".