【作者】 偶春生；

【导师】 张佑生；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2008， 博士

【摘要】 作为CAD/CAM、虚拟现实、增强现实、科学计算可视化等领域的基础和核心,场景的建模和绘制一直是计算机图形学中最活跃的研究热点。随着计算机图形学应用领域的不断扩展,绘制的场景越来越复杂,对模型精度和绘制真实感的要求也越来越高,这使传统的建模和绘制技术面临着严峻的挑战。近年来,研究人员不断地探索和研究新的场景建模与绘制技术,相继提了基于图像的建模与绘制(IBMR)、基于点的建模与绘制、逆向工程等技术。这些技术基于全新的理论和方法,以满足复杂场景的高真实感实时显示的需要为目标,绘制效率不断提高,建模与绘制质量不断改进,在虚拟现实、科学计算可视化等领域得到愈来愈广泛的应用,同时也展现出更加广阔的研究空间。本论文针对这些建模与绘制技术中的一些关键问题,包括基于图像的建模与绘制中的立体匹配与图像变形和基于散乱点集的曲面重建与优化等,进行了较深入的研究,所取得的成果对于相关技术的进一步发展,具有重要的理论意义和应用前景。论文的主要工作和创新之处如下:(1)针对IBMR技术中的关键问题——多幅图像的立体匹配问题进行了研究,提出一种基于窗口匹配和遗传算法的立体匹配算法,先用基于窗口的匹配算法快速求出图像的匹配,再考虑立体匹配中的唯一性约束和顺序性全局约束,使用遗传算法重新对错误匹配点求解出正确的匹配关系。此外,通过对立体匹配中重要的算法——协同算法和Zitnick的3D协同算法的分析,找出了其中影响匹配计算速度的原因,并对其中的三个关键因素进行了改进,大大减少了计算量,提高了立体匹配的速度。(2)通过对IBR中图像变形的常用算法的分析和研究,提出一种基于区域边界的变形算法,针对变形算法中特征指定、变形效果、计算代价三个关键方面给出了相应的改进策略。在特征指定方面采用任意形状的封闭环,不仅定义非常方便,还有着表现力丰富和适用性强的特点。在变形效果方面,算法重点关注特征区域,使扭曲局部化,易于用户对特征区域变形的预测和控制,可保证关键区域的轮廓精度。图像重建采用空间可变线性颜色插值,保证了变形图像的颜色精度。在计算代价方面,将复杂的区域自动剖分为简单三角形区域,使区域扭曲函数简单、统一,加快了计算速度。同时利用连贯性实现了像素点和三角形区域的位置关系的快速判定,降低了计算代价。(3)在基于散乱点集的曲面重建方面,提出了曲面法向量场驱动和曲面曲率驱动的重建方法,分别采用法向量和曲率来度量网格和待建曲面的误筹,给出了基于法向量和曲率的三角形面片生成规则。该方法在重建曲面拓扑的同时,可自适应地优化网格,即在曲面平坦(曲率较小)的区域生成较大的三角面片,在曲率较大的区域生成较小的面片。两种方法都显式地给出了三角形面片与曲面之间误差的度量,以用直观、统一的重建精度参数方便地控制曲面重建的整体精度。(4)通过采用网格局部区域的法向量来度量边折叠代价,提出了特征保持的基于边折叠的网格简化方法。该方法无论是在模型的高曲率区还是低曲率区都能在高简化率下保持网格的特征。提出了基于特征的质点—弹簧模型对简化后的网格进行局部优化,采用边的平均曲率作为弹簧的弹性系数,使得网格能够根据形状特征进行优化,从而保持网格的原有细节特征。更多还原

【Abstract】 Modeling and rendering method of scenes is one of the most active research fields in Computer Graphics. As more and more high precision and detailed models are required in many applications, it is difficulty to create accurate model and realistic image with traditional modeling and rendering methods on popular hardware platforms. Some new modeling and rendering methods such as Image Based Modeling and Rendering, Point Based Modeling and Rendering, Reverse Engineering are presented recently. Several key techniques in those new methods are discussed in the dissertation. The main research contents and innovations are as follows:(1) Stereo matching between two real images is a key component in IBMR, in this dissertation, we present a new hybrid stereo matching method based on genetic algorithm. The method is composite of two parts: A local window-based stereo matching algorithm is employed to match image rapidly though there are some mismatches in the result. After filtering those mismatches, the genetic algorithm is used to correct those mismatches by global constraints. Additionally, we also analysis another typical stereo matching method which is called cooperative algorithm presented by C. L. Zitnick. By improving the algorithm in three aspects, the computation cost of new algorithm is much lower than Zitnick’s algorithm.(2) A new image morphing algorithm based on boundary of regions is presented. The study focus on three key point of image morphing: Features of images are specified conveniently by arbitrary polygons, those polygons also divide the whole image into some sub-regions which are classified into feature regions and non-feature regions. Feature regions contain details of images. In order to get higher quality of morphing, warping functions in feature region are determined only by its own boundary while warping functions in non-feature region are determined by boundary of several non-feature regions. Dividing the polygon region into triangle mesh automatically can not only simplify the warping functions but also reduce the computation cost. The space-variant liner image interpolation is used to reconstruct a high quality image.(3) Reconstructing a mesh surface from a given scatter points has become an important modeling method, most of previous reconstruction method are only based on the coordinate information of the scatter points. We present a new surface reconstruction method based on normal and curvature of sampled surface. Normal and curvature as key features of surface can be computed from the coordinate information of scatter points. Driven by normal and curvature, the mesh surface can be rapidly reconstructed and is also optimized for its detailed maintaining the shape feature.(4) Most of the simplification algorithms use a distance metric as simplification criteria. The distance metric is not good for preserving the shape of the original mesh. In the dissertation, we present a new mesh simplification algorithm based on vertex’s normal, which use edge collapse method to reduce the density of meshes. As the normal is useful to enhance the shape description, the algorithm can simplify the mesh while preserving the features of the surface. A model based on mass-spring system is used to optimize the simplified mesh, taking average curvature of the edge of mesh as the tension of the spring, the vertex of the mesh can be optimized and most of feature of original mesh are preserved.更多还原