【作者】 李瑞瑞；

【导师】 秦开怀；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2014， 博士

【摘要】 全局光照模拟绘制是真实感图像生成不可缺少的组成部分。随着真实感图形绘制应用的飞速发展，全局光照绘制方法正同时受到来自真实感和实时性两个方面更高要求的挑战。在有限的绘制资源下，针对完全动态、包含大量几何数据的场景，现有的方法不能实时交互地生成包含多种光照效果的真实感图像。本文对包含各种光照效果的交互绘制问题展开研究，针对大规模的、完全动态的场景，提出一种快速可交互地绘制包含反射、折射、焦散以及间接漫反射光照等多种效果的全局光照绘制方案，形成GPU上的完整的绘制流程，并通过实验验证方法的有效性。主要的研究成果包括如下几个方面：针对间接光照的可见性判断的瓶颈，研究了体加速结构的实时构建方法，提出了目前最快的实时体素化的方法，得到的体素结果占用存储少，没有空洞。在该基础上，进一步提出建立体层次结构以及基于层次结构快速判断光线交点的方法，明显减少迭代计算的开销。基于层次组织的体素数据，研究并改进了VGI中提出的基于体的近景间接光照计算方法，提出通过体远近划分来实时计算间接光的方法，并进一步利用多分辨率光照收集加速间接光照的计算。针对动态场景，该方法不仅可以达到每秒绘制20帧以上，且得到的绘制质量较VGI更高。对于包含镜面反射/折射体的全局光照绘制问题，研究并提出了包含反射、折射、焦散的全局光照计算方法，该方法能够模拟高频和低频的全局光照以及它们之间的交互，绘制速度可以接近或者达到实时，比基于GPU的近似光子映射方法快2～5倍。论文研究并解决了基于图像近似的焦散绘制方法所带来的失真问题，提出基于光学映像虚物体的焦散绘制方法。该方法通过GPU加速的几何变换方法来快速确定光子的位置，同时提出基于光子路径映射图构建焦散映射图的方法。对于几何面片表示的复杂场景，该方法能够模拟经多次递归弹射产生的焦散，并解决欠采样失真的问题，具有更强的真实感。论文也对超大规模的几何数据的真实感绘制进行了探索，提出基于NPU快速图像合成的Sort-last并行绘制系统NPUPR。该系统负载均衡好，并行开销小，灵活可扩展，具有一定的应用前景和实用价值。更多还原

【Abstract】 Global illumination is a vital part of realistic image synthesis. With the rapiddevelopment of graphics applications, it becomes a challenge to ask the globalillumination methods to render photorealistic images in realtime. Current researches onglobal illumination cannot render multiple global illumination effects interactively forscenes with both large geometries and fully dynamic animations. In this dissertation,based on the survey on interactive realistic rendering, we propose an interactive globalillumination approach, which simulates reflections, refractions, caustics, and diffuseinter-reflections simultaneously for large and dynamic scenes. The approach provides acomplete rendering process on the GPU. In the experiments, the validity of the approachis verified. The contributions of the work are summarized as follows.For the bottleneck of indirect occlusion tests, we first study the way to dynamicallyconstruct the volume-based accelerating structure. In this work, we propose a real-timebinary voxelization which may be the fastest voxelization method until now. Thevoxelization can get high-quality voxel-represented scenes with less error and using lessmemory storage. We also introduce the way to build a hierarchical structure on thevoxels. With its help, rays stop marching early and we find the intersections of the rayand objects in the scene fast.Based on the hierarchically organized voxels, the dissertation extends the work ofvoxel-based near-filed indirect illumination described in the VGI and proposes areal-time indirect illumination method by the near-far irradiance decomposition. In thiswork proposed a multi-resolutional gathering and a multi-pass geometric basedup-sampling is to get higher performance. It is proved in the experiment that for largedynamic scenes, the method can get high realistic images in a speed of more than20frames per second.For global illumination with reflections, refractions and caustics, the dissertationprovides an approach to simulate low-frequency lights, high-frequency lights and theirinteraction effects. The method can render interactively. Even compared with the fastGPU-acclerated photon mapping, it renders2~5times faster.The dissertation also analyzes and solves the problem of the artifacts caused by the approximation in the image-based caustic rendering. Based on the concept of virtualobjects in the optical maping theory, we introduce the virtual-object-based photon beamtracing for caustic rendering. Taking advantages of GPU techniques, the method canfind the correct positions of photons fast by the geometry transformation computation.The method also introduces the path-based photon mapping to construct causticmapping. For complex geometry-based scenes, the method elliminates artifacts causedby undersampling and renders caustics caused by multi-bounces reflections orrefractions realistically.Finally, the dissertation explores the way to render large-scale geometry data inparallel using a distributed clustered system. We propose a fast NPU-based imagecomposition for the sort-last parallel rendering system, called NPUPR. NPUPR hasgood load-balance, low parallel cost and scalable features. It can be applied to a widerange of applications.更多还原