3D Model Retrieving Based On Local Multi-scale GABOR Feature Of Best Sketch

The 3D models retrieving is a hot issue in recent researches, and the direct idea of this problem is based on the example model.It means that we input the target model and a model is returned which is most similar to the target model. This is very inconvenient.Image-based 3D model retrieving framework is proposed in recent years. Image based retrieval framework contains projection image based method and sketch-based approach. In contrast, sketch-based methods to is more convenient.The3 D model is projected to a plurality of 2D sketches in different direction. Users can retrieve sketch using this method at any time.The problem is transformed into 2D sketch retrieval problem. Of course, this method need extract the image features offline.In this paper, we do two works.At first,we propose a best view selection method,reducing the number of views.The second,We propose the multi-scale GABOR local feature,which contains the spatial distribution of the sketch local lines.Best view selection method proposed in this paper is choosing the projection direction which contain the most information using the information entropy of each viewpoint.Viewpoint entropy can capture corresponding area of perspective projection and the number of projection plane. Combined with the view stability based on depth image,we select the projection direction which contains more information,and the direction is in keeping with human drawing habit.There are a large number of projections, and the viewpoint evenly distributed in the sphere of the bounding sphere.When people render the sketch,they prefer to render in some directions which are three quarters of the whole directions. Some researches show that people tend to draw sketches in these directions,because these directions are stable. Combined with these two algorithms, we calculate the weighted score of eachviewpoint, the higher the score as the best view.Using this method,the redundant views are removed,and the number of views is less.Local multi-scale GABOR local feature is based on the GABOR filter extraction method related to local features.We propose the multi-scale GABOR local feature,making local descriptor contains multiple scales.The area around the feature points is partitioned.It is divided into a number of different scales.The extracted feature vector,which contains the spatial distribution of the sketch local lines. We call this local feature multi-scale GABOR feature. Then using the bag of word model to cluster the local features.Then using the related data structure, such as K-D tree, to find the nearest clustering center of each local feature vector. At last, the global feature vector of the sketch is formed.Using the selection of the best view and improved sketch feature,the number of projection views is reduced, which greatly reduces calculation of the feature extraction in exacting features,and enriching the space information. By using the partial model of the model library of Shrec12, we have carried out the experiment under the same conditions, and compared with other methods. In the retrieval results,after calculation, for the same recall, our method is more precise.It means that adding ketch line local position distribution information to the feature vector is helpful to improve the search results.