The Shape Feature Extraction And Retrieval Of Three Dimensional CAD Models

The digital design and manufacturing techniques are the core technologies of intelligent manufacturing. In recent years, with the three-dimensional modeling software, the Internet technology, special graphics hardware widely used, the manufacturing companies have accumulated a large number of 3D CAD models for digital design and manufacturing. The digital manufacturing is facing the challenges of new manufacturing mode and new technology in full life cycle of products, network,intelligence and large-scale model database processing. How to make full use of the 3D model databases for quick retrieval and then reuse the existing product designs, modify or conduct functional test are the keys to digital manufacturing, which are aimed at shortening the product development cycle, improving the productivity and quality. The study of extracting the 3D model shape features and the development of model retrieval system is an effective way to solve this problem.At present, the shape feature based 3D model retrieval algorithms has made much progress, but there are still some key problems not solved: (1) Given the various applications of the 3D model retrieval in digital manufacutring, how to effectively express the shape characteristics of the model and meet the users' needs; (2) The shape feature of the 3D model is usually a high dimension vector, and how to design an efficient matching scheme when etrieval models from large-scale model databases.(3) The users have high expection on the retrieval efficencyand precision, and then how to deisgn a retrieval system that can meet the retrieval personality as well as high performance.Based on the extraction of shape feature of three-dimensional model, this dissertation analyzes and excavates the rich shape feature information in different types of models, and establishes an efficient model retrieval mechanism. Based on theoretical deduction and many experimental analysis, the study illustrates and validates that the proposed methods are superior to the traditional classical algorithms. The main content and innovations of this dissertation are as follows:(1) This dissertation explores the fast and effective retrieval methods in large-scale 3Dmodel database, and analyzes the characteristics and properties of six three-dimensional descriptors based on statistics (D1, D2, D3, A3 shape distribution,string method and convex hull). The efficiency and results of the model retrieval are evaluated by experiments. It has shown that utilizing the simple geometric information of statistical models is a fast, simple and low cost method, and suitable for large-scale model database retrieval.(2) This study proposes a new model expression with 2D slice to represent the 3D model feature distribution. In the view of discrimination deficiency of shape distribution, this approach developes an integrated 2D slice representation of 3D CAD models, and with the refined local 2D slice representation it improves the feature discrimination and robustness. This appraoch reduces the matching error and improves the performances of model matching against traditional classic algorithms as shown from experiemnts.(3) This dissertation develops an algorithm for low intrinsic dimensional embedding of high dimentional feature space. In view of the fact that the three-dimensional model contains not only appearance property but also complex geometrical data, which makes it is difficult to untilize the model information efficiently, this study represents the model as adjacency graph, and the eigenvector of adjacency graph is computed as the basis to retain the original model topological information. The method of anisotropic kernel diffusion is proposed, so that the salient feature points needed to express the shape characteristics of the 3D CAD model (e.g. gear) are preserved and the eigenvector is computed. Finally, the Grovmov-Hausdorff metric method is proposed to align the spectra and solve the problem of alignment and comparison in two different scales.(4) This dissertation proposes a local-global shape descriptor. Heuristic digital design requires the semantics of models, while conventional 3D model matching approaches extract only the geometrical and topological information of the model, and can not meet the semantic requirement. By computing the wavelet kernel energy of each vertex and then using wavelet coefficients as local-global descriptors, a sparse dictionary is learned and the signatures of each non-negative coding coefficient is matched to the atoms in the dictionary to achieve the model retrieval. This approach has improved the capability of retriving model sematics as demonstrated.