Research On Knowledge Extraction And Representation Of 3D Model

3D models represents the shape and appearance of real-world objects.Users can have a comprehensive and accurate understanding of their shape and appearance through a multi perspective observation.The knowledge acquisition is mainly text-based,and the more expressive non-text resources are difficult to describe in natural language.It is still very difficult to build an easy-to-use knowledge base of 3D models.Several studies support extracting of descriptive by low-level feature extraction and high-level semantic annotation.However,there is still a lack of study on organizing text description in the form of knowledge and the abstract expression of image and salient region.There are countless perspectives to observe the model.In addition,how to select a small number of views that represent complete visual information and extract semantics as much as possible,has no unified objective criterion.It is still a challenge to build a visual knowledge base though extracting salient region of 3D model for high-level semantic annotation and effectively representing relationships between different forms describe data.Aiming at the above problems,this dissertation studies the knowledge extraction and representation of 3D model,so as to improve the capability of 3D model annotation,reuse and intelligent management.The main contributions of this dissertation are summarized as follows:1.Aiming at the optimization of viewpoints for contour observation of 3D models and semantic extraction,a multi-level view descriptor is proposed to provide a set of 2D view selection based on contour visual information for 3D models.Moreover,semantic extraction rules are established to realize the transformation from features to semantics.The uniform spherical observation viewpoint model is first developed,and the up,front and right viewpoints with intuitive semantic information are considered as the initial viewpoints.According to the chamfer distance comparison of the projection edge of the model,the optimal viewpoint is then obtained as the seed point of the next layer.It continuously subdivided in order to obtain the combination of observation viewpoints which is better than the uniform selection view at different levels.Finally,the roundness,rectangularity,visual axis and symmetry are calculated,and the transformation rules are designed to realize semantic extraction.2.Aiming at the problem of depth view optimization with more abundant information and corpus extraction of observation 3D model,two information entropy measures of depth change entropy and distribution entropy on the orthogonal projection depth map of 3D model,are proposed.The optimal view is selected according to the amount of information.The global representative view is then automatically selected using the decentralized viewpoint selection algorithm.Afterwards,a threshold visual word histogram abstract description method based on word bag mechanism,is proposed.The effects of different view-based retrieval algorithms are compared on the public dataset to evaluate the recognition ability of this abstract description.Experiment show that the selected depth view is better than the uniform selection view.The principal component analysis and block view acquisition model design in this method,ensure the rotation invariance of the algorithm,and provide a method and abstract description of determining a small number of observation views,based on the depth information of the 3D model.3.Aiming at the problem of complex and nonintuitive operation of salient region extraction in 3D space,a 2D geometric image representing the 3D mesh,is proposed.The image retains the neighborhood topology information of 3D mesh points.In addition,it can intuitively segment the significant region of the original 3D model from the image,in order to support advanced semantic annotation.The Euler loop cutting path calculation method is proposed to obtain a smooth cutting path.The 3D mesh is cut into a simply connected surface,and parameterized into a flat unit rectangle.The geometric optimization step of capacity constrained Delaunay triangulation method is then used to obtain the point distribution with the characteristics close to the blue noise spectrum.A mesh relaxation algorithm is also proposed to keep the vertex topological relationship moving to the integer coordinate points without repetition and patch folding.Finally,the blank part of the image is filled by the local affine transformation of the corresponding triangular surface.This method performs one-to-one mapping between the 3D surface mesh and points in the 2D image.It can simultaneously generate corresponding images of vertex curvature,normal direction,shape index and other attributes.It also performs 3D model salient region extraction using 2D image processing algorithms thouth curve evolution,which reduces the difficulty of extracting salient region from 3D models.4.Aiming at the problem of knowledge representation of different types semantic description data,a hierarchical visual knowledge ontology representation method which can be extended,reasoned and shared,is proposed.The top-level structure of macro domain classification is considered as the basic class of 3D model visual knowledge ontology.An entity dictionary and relational dictionary are developed,a knowledge representation is carried out in combination with weighted directed graph,and a sparse matrix is used for storage.The upper classified entities,specific instances and attribute values are uniformly set as vertices.Relationship words and attribute names are uniformly set as edges.The basic ontology of 3D model visual knowledge is developed from top to bottom.Specific 3D model instances,relationships and attribute values are then added from bottom to top.This method performs the expression of relational characteristics to support reasoning and strengthen description.It can obtain the relevant knowledge of conceptnet common sense base and enrich the knowledge base.In addition,based on the visual knowledge base of 3D model,rules are set to verify the efficiency of semantic retrieval,put forward the intelligent application of 3D reconstruction of first person sketch terrain.This show 3D model visual knowledge base reuse and effective support for intelligent development and application.