Research On 3D Model Retrieval Based On Feature Combination And Manifold Ranking

3D models have been widely used in many areas. With the rapidly emerging of 3D models, content-based 3D model retrieval technology has become a key issue to find out the desired models.The accuracy of 3D model retrieval depends on three main factors, feature extraction, feature coding and similarity measurement. We have proposed three features for high accuracy 3D model retrieval along with two future combination algorithms and one manifold ranking method after extensively research on discriminative power feature extraction, future combination, efficient feature coding and the effects of manifold ranking in similarity measurement.3D models can be divided into two categories, the rigid and non-rigid. We mainly concentrate on rigid 3D models in this paper. With extensively research on this topic, we have proposed three features, two future combination algorithms and one manifold ranking method for 3D retrieval. The main contributions of the dissertation are depicted as follows:(1)We have proposed three features for 3D model retrieval to solve problems on existing features.Shape distribution algorithm may suffer from the problem that different models contain the same feature without considering the difference between the bounding box information and normal area distribution information. To solve the problem, we integrate the information into shape distribution algorithm for the compensation of shape distribution feature to enhance 3D model retrieval performance.View-based 3D model retrieval methods are the state of art in rigid 3D model retrieval. However, these methods encode all the depth images from one model without considering the viewpoint distribution information, thus may lead to failure in some situation. Therefore, we propose a novel 3D model retrieval method based on multiresolution depth image with fisher vector coding. We firstly extract multiresolution depth images from a given 3D model, and then divide them into two parts according to the depth image whether belongs to orthographic view or not. Next, we use a Gaussian Mix Model to form a visual dictionary of each part for feature coding. We encode these features from two parts with fisher vector coding to get two vectors. At last, the model feature vector is the concatenation of the two vectors.Furthermore, we cannot capture the hidden information of 3D models with the schema of view-based feature extraction. Therefore, we propose another new 3D local feature for the compensation the information loss of view-based feature. The new 3D local feature is invariant for occlusion, and the two features are complementary to each other.(2) We propose two feature combination algorithms for 3D model retrieval. The first algorithm mainly focus on distance combination by adjust weight and discusses the influence of complementation between different features for feature combination. However, there are also some problems such as the determination of weights and the normalization process might be effected by data distribution. So the second algorithm which based on the adaptive weight adjustment on the retrieval ranking order is used for alleviate the problem comes from the feature combination process. We first select the best feature as the standards, and then the weight value is related to the correlation between the candidates and the standards. Later, we calculate affinity values from the ranking order and update the values by nearest neighbor information. At last, the feature combination process is carried on the adaptive weights and the affinity values. Because of the ranking order is invariant to the magnitude of the original data, the normalization process is not needed in our algorithm. Experimental results show the proposed algorithm with high accuracy on 3D model retrieval.(3) Since manifold ranking may discover the manifold structure behind the original data, we integrate it into our 3D retrieval process. However, the affinity matrix fully depends on the original distance matrix, so the outlier may lead to unstable results. In addition, some models are removed from the correct result set due to the transition process. Therefore, we use the ranking order that is invariant to the exception and outliers to construct the affinity matrix, and then we proposed a new transform term on the transition result to solve the problem that models in correct result might be lost after transition process. Experimental results show the proposed method is also suitable for combined feature and the performance exceed other manifold ranking based methods.