Research On Multiple Feature Fusion Of 3D Model Retrieval Technology

With the rapid development of multimedia technology and computer graphics technique, 3D models has been used in many fields, including computer aided design, medical diagnosis, 3D movies, 3D games, etc. At the same time, there is an exponential increase in the number of 3D models. Therefore, how to retrieve a 3D model for repeated use from existing 3D databases has been a hot pot in recent years. For 3D model retrieval methods, feature extraction is the most major step.Because the single feature extraction algorithm was lack of comprehensive description for a 3D model, many researchers paid attention to the multiple feature fusion to improve the ability of description for a 3D model. However, how to improve the ability of the description for a 3D model and propose more effective methods is still a tough problem. In addition, how to select features for feature fusion is a problem worth studying. For these two problems, the main research of this dissertation can be summarized as following:(1) We propose a new method called combining local and global features for 3D model retrieval. We fuse local feature and global feature to get a new feature that can describe a 3D better to improve the retrieval precision. Our approach selects the SIFT feature that can reflect the local feature of a 3D model and the multiscale Fourier descriptor feature which can reflect the global feature of a 3D model to represent a 3D model. Multiscale analysis method is used to reduce the impact of noise on gray image Fourier transform. Then we utilize the two different features to calculate the similarity between the query model and the models in the database respectively. Finally, the features are integrated via the linear combination of the distance values they produced with adaptive weights.(2) We propose a 3D model retrieval based on weighted BOF. Tens of thousands of sift feature points can be extracted from a 3D model, and each feature point is a 128-dimensional vector. If we use SIFT feature for feature matching directly, it will slow down the whole retrieval process. To solve this problem, we propose a 3D model retrieval based on weighted BOF. Our method utilizes word bag model to train SIFT features to generate a visual dictionary, which can greatly reduce the feature dimension of a model. To further improve the discrimination between feature vectors, we utilize the TF-IDF algorithm to weight the BOF. Then the weighted BOF is utilized to calculate the similarity between the query model and the models in the database. Finally, we apply adaptive weights to fuse the distance values calculated by multiscale Fourier descriptor feature.We made experiments on two data sets: the test dataset of Princeton Shape Benchmark(PSB) and the Shape Retrieval Contest 2012(SHREC12GTB). We compared the result of our method with five methods on the SHREC12 GTB and four methods on the PSB respectively to evaluate the performance of our method respectively. Experimental results demonstrated the proposed method was effective.