Research On Model Query Interface And Feature Extraction Algorithm In 3D Model Retrieval

With the development of computer graphics application domains, such as industrial design, digital medical care and entertainment, the amout of 3D models experiences an explosive growth. As the text search emerged from information expansion of internet, a 3D model retrieval system with high-efficiency is highly expected. For the time being, research interests of 3D model retrieval focus on query interfaces and feature extraction algorithms.Compared with traditional text query interface, model query interface can provide intuitive 3D model information, especially in the condition lacking text description. How to construct a fast, perception adaptive model query interface, which can offer model feature with high distinction, is the key of query interface research.Model type provided by query interface is corresponding to feature extraction algorithm. 3D model type includes surface model and solid model. Compared with surface model, solid model can provide volume information with low cost. Voxel model, which is a sort of solid model, is attracting more research interests in 3D model retrieval due to its regular solid information.Under the funding of National Natural Science Foundation (60573146, 61073086), this dissertation is devoted to theoretic and application research on 3D model retrieval system based on voxel model. Through research of sketch modeling and image based modeling, both of these two aspects, which can establish voxel model, are introduced to model query interface in 3D model retrieval. The work and contribution of this research are as follows:1. Present a 3D model retrieval framework based on voxel model. The core processing element of this framework is voxel model information. In theoretic level, this framework includes acquisition and retrieval of 3D regular information provided by voxel model. In application level, this framework includes model query interface, voxel model database, feature extraction and comparision. From the search of papers, retrieval system consturcted under this framework is the first one to introduce certain scale voxel modeling to 3D model query interface.2. For sketch modeling query interface, a virtual canvas based normal voxel generation method and a seed adjacency based voxel modification method are prsented.The virtual canvas based normal voxel generation method consists of normal equal altitude algorithm and normal gradient algorithm. The seed adjacency based voxel modification method consists of half-closed curve algorithm and closed curve algorithm. This set of voxel generation method and voxel modification method with real-time feedback support the voxel model query interface. This voxel model query interface, which is from the intuitionistic visual cognition of human, maps 2D drawing information into 3D model information, neatly constructs topology structure and regular voxel information of voxel model, solves the lack of space information in pure 2D query interface. From experiments, this interactive query interface, based on its ease of use, can provide quick access to voxel model, which is adaptive for volume feature extraction.3. Present a 3D model query interface based on image modeling, which is based on images in sequence and does provide a wide range of model features.This query interface is constucted under the image modeling method based on images in sequence. Meanwhile, this image modeling method constucts 3D model in the way of stack of cross-section shapes. In detail, this modeling method firstly introduces classic SIFT algorithm to produce primary feature mathing between views and ground view. Then, through an advanced RANSAC algorithm presented by this dissertation, more resonable and global mathing pairs will be generated. From these matching pairs, the homography on ground and cross section will be caculated. The silhouettes of model in every view will be cut out by a silhouette extraction algorithm based on Lab color space. After that, these silhouettes will be transfomed by homographies and crossover each other to generate cross-section shapes of real model. Finally, a voxel modeling method based on binary grid is presented to build voxel model from these cross-section shapes. Compared with usual image based modeling method, this 3D reconstruction method can provide various model features as cross-section shapes, volume and topology. As the constructed result of this modeling method, voxel model, which is full of volume feature, can be used by model query interface. Experiments show that this model query interface is practical.4. A feature extraction algorithm based on voxel model is presented.This feature extraction algorithm,which is a statistical method on volume level, adds up the voxel distribution function from voxel model. From this algorithm, a rotation-invariant descriptor called VD2, which can simultaneously provide common features among models in the same class and differential features among models in different classes, is generated. Compared with classic D2 and GD2, which are feature descriptors based on surface model, VD2 has higher retrieval efficiency. In addition, for the VD2 extraction from surface model, a voxelization method based on radial weight is presented. This voxelization method can repair surface degeneration as cavums and holes in just one-step. Compared with traditional method, this voxelization method can provide voxel models with more robust volume information.All these four research aspects are correlative and independent. They are inplemented under the voxel based 3D model retrieval framework and do research on query interface and feature extraction. They work as an organic whole. Relative experiments show the practical value of the 3D model retrieval system.