Non-rigid3D Model Retrieval And Tagging Based On HKS

3D model resources are rapidly growing on the Internet, and the research about3D model retrieval and semantic annotation technology can help organize and manage the existing3D model data resources, so it received widespread attention by the researchers. Although model retrieval technology developed rapidly in recent years, the majority of existing researches are for rigid models, they can't solve the model matching problem of non-rigid objects with active joints, so the non-rigid3D model retrieval is still a major challenge in the field of3D model retrieval. The main research of this paper is non-rigid3D model matching. Using the HKS (heat kernel signature) feature of model vertices as the model features content, not only solved the problem of model retrieval with isometric transform, but also has good robustness to model transformations. And based on this, we proposed a new non-rigid model retrieval method that the model feature is made up by'Visual words'and'Visual phrases' according to the single-value evaluation. At the same times, based on 'Visual words' and fuzzy classification, we research non-rigid model semantic annotation.The main contents of this paper are as follows:1. The majority of non-rigid deformation can approximate isometries, by comparison and analysis the existing3D model feature extraction methods, this paper uses HKS feature which reflects the heat distribution change of model surface to describe the vertices characteristics of3D models. This is a good solution to the problem of non-rigid model retrieval, as well as small distortion of the model deformation process, have better robustness.2. Since the Bag-of-words model ignores the spatial relationships of the'Visual words', we purposed to add spatial relationships into'Visual words', and becomes 'Visual phrases'. Then we composed these two features according to the single evaluation value, for describing the shape of the non-rigid model better.'Visual phrases'is formed by combinations of features of model vertexes and the1-ring neighborhoods. Comparison with'Visual words','Visual phrases'considered the spatial relationships between features. Single evaluation value reflects the retrieval effectiveness of different features, so it can determine the weight of two features. The experimental results show that our algorithm has better results in accuracy of the nearest neighbor. 3. In order to achieve the semantic annotation of non-rigid model, we use Fuzzy K-nearest neighbor classification algorithm to get probabilistic classification results based on'Visual words'. It can semi-automatically annotate3D models. This research improved the text annotation information, as well as the precision and recall of text-based3D model retrieval.