Research On 3D Model Retrieval Based On View Features

Nowadays, more and more three-dimensional technology has been widely applied, such as 3D printing,3D scanning, and 3D design.3D technology is increasingly showing great potential for development in information sharing. Features extraction and model retrieval is one of the focus fields of computer graphics and is a very challenging open issue, which has great theoretical and practical value.The core issue of 3D model retrieval system is how to build an efficient feature extraction algorithm to retrieve complex Internet massive 3D model. However, the complexity of the model and the challenges of the feature extraction still exist. Firstly, although some features (geometry features, visual features and topological features, etc.) of 3D models have the ability to identify a specific range of models, but it may not for all models. Secondly, various types of features are not isolated, unconnected, but co-existence and interaction. It is a very difficult problem that how to banlance advantages of a single feature and the complex dimensions of features. Finally, the existing single feature or multiple features can not always an accurate representation the process of artificially model recognition.Domestic and foreign scholars have done a lot of work, which so far have not been effectively addressed. To solve the above problems, based on the view features extraction methods conducted a series of studies, several effective and new methods are proposed as follows:1 A new method to obtaining a perspective image of 3D model is proposed. The parameter θ andφ of the 3D Radon transform determine the projection plane. The projection is the view image of the 3D model in the perspective. The information generated by the function transformation is view image, which not only reduce the dimension of features, but also improve the model retrieval efficiency.2 A new method was proposed by the Radon transform algorithm and BOVF method. After obtaining the multi-view image of each model by Radon transform, a visual dictionary library is generated by using Bag-of-Features. And then the frequency of the characteristics of the visual vocabulary is counted as a feature of 3d model. Finally, the feature was used for model retrieval. The experimental results show that:(1) the feature extracted by this method can accurately describe the 3D model; (2) the retrieval performance of the method is superior to the comparison algorithms; (3) the retrieval performance is stable and less affected by the system parameters.3 Construct a three-dimensional Radon normalized central moment, and applied to extract features of 3D model. Firstly, a normalized 3D Radon central moment is constructed by combining PCA method. The moment is invariant to translation, rotation and scaling. Secondly, the visual features of a three-dimensional model are obtained by Normalized 3D Radon centers moments, and the feature vectors of 3D model are generated. This method needn’t obtain the model’s view image, and can be directly extracted view features by moments from the specific view. It is greatly simplifies the process of extracting features of the existing methods. Experimental results show that this method has a good shape description, shorter time of feature extraction and stability retrieval performance, and the retrieval performance is less affected by the moment parameters.4 Proposed a histogram extraction method of protein structure. Firstly, a method of obtaining the view-image of protein structure is designed by regular polygon, which can automatically select model, load model, rotate the model and capture the images. Secondly, the local feature of each view-image was extracted by SURF algorithm, and a visual dictionary was generated by K-means algorithm, and then vocabulary histogram is built. Finally, the view-image’s histogram is feature vectors of protein structures. The Experimental results show that:(1) the correct of protein structure prediction of proposed method was significantly higher than the comparison algorithm; (2) the proposed method has a higher recall and precision rate; (3) the retrieval performance is less affected by the number of dictionary size and the sample set.