Application Of Model Segmentation In Content-based 3d Model Retrieval

With the rapid development of computer and network techniques, and the increasing perfection of the computer graphics theory, 3D model has gradually become the fourth type of multimedia data following the sound, image and video. Their number is exploding. Since it is very difficult to make high fidelity 3d model it may take a lot of time and energy. If you can fully reuse the existing data resources of 3d model,you can also greatly reduce the workload of designing new models. This requires sound methods for the classification and retrieval of 3d model.3D model retrieval includes text-based retrieval and content-based retrieval. Text-based retrieval technology is very mature and widely used, but it takes lots of human and financial resources. Subjective factors are involved in text-based retrieval technology, so that it tends to fail. Thereby it is not suitable for 3d model retrieval. Content-based 3d model retrieval firstly automatically calculates and extracts the feature extraction from the model data, and then creates 3d models of multi-dimensional information index and computes the similarity of query models and target model, which is used to browse and retrieve the 3d model database. Since content-based 3d model retrieval technology involves less manual intervention, 3d model retrieval research mainly focuses on content-based retrieval technology.The general process of 3d model retrieval: Calculate shape features of the search target, and compare them with all model features in features library, then we get several models that have the closest feature. Enabled shape features reflect some features of the human's visual perception, which is a new idea to solve the problem of shape features description in 3d model retrieval.In recent years, some researchers have proposed that based on the component recognition theory in cognitive psychology, segmentation could be used to divide the models into several significant components. the relationship between the parts is then analyzed to obtain the feature description, according to the measure of the Similarity of 3d model.In this thesis, the author introduces the current study on 3D model retrieval and discusses its prospects for application, analyses the significance of the study on 3D model retrieval system and the evaluation of search properties, and summarizes feature extraction method and similarity matching. It also briefly introduces the main content, the purpose and the significance of this thesis. At the same time, the author shows the need for standardized pretreatment of the model, and preprocessing steps are described in detail: Translation normalization, Scaling normalization and rotation normalization. To achieve rotation normalization, we use the method of principal component analysis (principal component analysis,PCA) to determine the spindle of 3d model.Making a comprehensive overview to 3D model retrieval techniques, the author proposes a new concept and three methods. The concept is application of model segmentation in content-based 3D model retrieval. Three methods are 3D model retrieval based on partition projection, based on bounding box segmentation for 3D model retrieval and 3D model retrieval based on concentric sphere segmentation.In 3D model retrieval based on partition projection method, We take the centroid of model as the origin, and takes the surface formed by system axis as the section, divides 3d model into 8 model components, and projects each 3d model component onto two-dimensional plane, gets the 2D projection points set all directions. And then segment these 2D projection points in the method of fan segmentation, extract the maximum distance between 2D projection points and the centre of projection points in all fan-shaped area as feature. In based on bounding box segmentation for 3D model retrieval method, we use axis surface to segment the model. There are three axis surfaces, so we can get six model segmentations from model segmentation. We design bounding box and take each model into the box, and then the bounding boxes are divided into several regions. In each region, we compute and search the point which is the farthest/ nearest away from the segmentation surface in the 3D model points set, as the first/second feature vector. Finally we compute ratio of the nearest point and the farthest point as the feature vector weigh in each region. According to the feature vectors, we get the 3D model vector similarity distance. In 3D model retrieval based on concentric sphere segmentation, we select the centroid of 3d model as the center of the concentric sphere and select the point which is the farthest away from centroid as the maximum radius of the concentric sphere. There are several model segmentations from which the 3D model is divided by the surfaces of concentric sphere. We use concentric sphere ring as the extract feature vectors regional unit. In every region, we select the point which is the nearest away from the surface of out concentric sphere ring as the feature point, and calculate distance between two feature points in adjacent region, which is used to set the weighting coefficient between two compared models. The advantages that base on concentric sphere segmentation for 3D model retrieval is that we can eliminate the rotation normalization operation from preprocessing step. Thereby it reduces the time complexity of the algorithm.The author has been through programming to implement the three 3D models retrieval methods and has got a lot of the experimental data. The experiments indicate that the three methods show more accurate retrieval performance.