基于光传递矩阵表示与重构的高效真实感绘制技术

发布时间：2018-09-18 12:04

【摘要】：近年来,真实感图形绘制在现代生产和生活中的应用越来越广泛。作为真实感图形绘制的重要手段,全局光照明绘制技术一直是计算机图形学的一个重要研究方向,它通过复杂的计算在位置、形状、材质、介质和光能辐射传输等方面模拟或逼近光的物理行为,构成了工业设计、影视娱乐、虚拟现实等需要高质量真实感图像的应用的根基之一。由于高精度的全局光照明绘制涉及庞大的计算量,人们提出了形形色色的算法尝试从不同思路来高效求解这一复杂问题。近年提出的大量光源全局光照绘制技术是其中非常有潜力的一个创新解决方案,它使用大量虚拟光源来模拟真实光源经物体多次反射后产生的间接光照效果,实现了高效率、可伸缩的全局光照明绘制,使得用户可以根据不同的应用需求自主调节绘制质量和效率之间的平衡。然而,该技术在大量光源的收集、out-of-core场景的绘制和视点子光路的重建等方面存在重大挑战,仍然需要数分钟甚至几十分钟,才能绘制一张高精度的图像。本学位论文主要研究大量光源全局光照明的高效计算问题,重点解决大量光源光传递矩阵和绘制点入射光场光传递矩阵的表示、采样和重建等难题,以高效地实现大规模、高质量的全局光照绘制。论文创新提出了四个新的绘制方法,即:使用稀疏采样结合矩阵分离的方法来加速非介质场景的大量光源绘制,使用自适应采样结合矩阵还原的方法来加速带介质场景的大量光源绘制,对入射光场构建光传递矩阵并进行稀疏采样与重构来加速入射光场的重建,以及借助GPU(图形处理器)的out-of-core(基于外存)绘制框架来实现大规模场景的大量光源全局光照绘制。这些方法有效克服了传统技术的不足,且效率提升了一个量级。论文的主要贡献包括:·提出了非介质场景大量光源高效绘制技。该技术创新将非介质场景大量光源收集问题转化为稀疏重建光传递矩阵元素的可见性函数问题,通过解耦光照和可见性计算来充分利用非介质场景可见性的局部耦合性,从而提高计算效率。我们的方法构造了一组可见性预测器,成功从稀疏的样本还原出所有矩阵的可见性,并引入矩阵分离方法,有效去除了稀疏的可见性误差,并还原出高频的阴影细节。该技术在保留光照细节的同时,极大加速了计算,绘制速度较现有方法提高了 7倍。·提出了含介质场景大量光源高效绘制技术。该技术创新将含介质场景大量光源收集问题转化为以虚拟点光源(VPL)和虚拟线光源(VRL)段为列、以绘制点和视线段为行的矩阵的列积分计算问题,为了挖掘介质间接折射的局部耦合性来加速计算,我们提出了自适应矩阵列采样与填充技术,仅使用稀疏样本就实现了含介质场景大量光源高精度高效绘制,相较于现有方法,效率提升了一个量级。·提出了绘制点处入射光场的高效重建技术。该技术创新提出采样光场的重要性度量来均衡离散光场,将入射光场重建问题转化为入射光场光传递矩阵的稀疏填充问题。利用相邻像素间相同入射方向的光照耦合性,自适应采样、稀疏填充并快速联合滤波入射光场光传递矩阵,实现了绘制点入射光场的高质量快速重建。与现有方法相比,无论在质量上还是效率上都有很大提升。·提出了大规模场景大量光源高效绘制技术。该技术首先将含out-of-core几何和光源数据的大规模场景大量光源绘制问题转化为光传递子矩阵的规划遍历问题,进而根据矩阵元素和几何的空间局部性,自适应划分子矩阵和几何数据,将最小化数据传输的规划绘制顺序问题转化为图路径搜索问题来处理,首次实现了支持out-of-core光源和几何数据的大量光源并行绘制,将可支持的场景规模提高了一个数量级,并比基于内存的CPU算法效率提升了一个数量级。
[Abstract]:In recent years, realistic graphics rendering is more and more widely used in modern production and life. As an important means of realistic graphics rendering, global illumination rendering technology has always been an important research direction of computer graphics. It simulates the position, shape, material, medium and light radiation transmission through complex calculations. The physical behavior of approaching light constitutes one of the foundations of applications requiring high-quality realistic images in industrial design, film and television entertainment, and virtual reality. As high-precision global illumination rendering involves a huge amount of computation, various algorithms have been proposed to solve this complex problem efficiently from different perspectives. The global illumination rendering technology with a large number of light sources is one of the most promising innovative solutions. It uses a large number of virtual light sources to simulate the indirect illumination effect of the real light source after multiple reflections of the object. It achieves high-efficiency and scalable global illumination rendering, so that users can be independent according to different application requirements. However, this technique poses great challenges in the collection of a large number of light sources, the rendering of out-of-core scenes and the reconstruction of sub-light paths of viewpoints. It still takes several minutes or even tens of minutes to produce a high-precision image. This dissertation mainly studies the global illumination of a large number of light sources. Efficient computation focuses on the representation, sampling and reconstruction of a large number of light source light transfer matrices and point-to-point incident light field light transfer matrices to achieve large-scale, high-quality global illumination rendering efficiently. A large number of light sources are rendered in the medium scene. Adaptive sampling and matrix reduction are used to accelerate the rendering of large numbers of light sources in the medium scene. The light transfer matrix is constructed for the incident light field and sparse sampling and reconstruction are used to accelerate the reconstruction of the incident light field. The out-of-core (based on external memory) rendering framework of GPU (graphics processor) is also used. These methods effectively overcome the shortcomings of traditional techniques and improve the efficiency by an order of magnitude. The main contributions of this paper are as follows: (1) Propose an efficient rendering technique for large number of light sources in non-dielectric scenes. The problem of visibility function of optical transfer matrix elements is solved by decoupling illumination and visibility calculation to make full use of the local coupling of visibility in non-media scenes and improve the computational efficiency. This technique not only saves the illumination details, but also speeds up the calculation greatly. The rendering speed is 7 times faster than that of the existing methods. An efficient rendering technique for a large number of light sources in a medium-containing scene is proposed. To solve the problem of column integral calculation of matrix with virtual point source (VPL) and virtual line source (VRL) segments as columns and drawing points and line of sight segments as rows, an adaptive matrix column sampling and filling technique is proposed to speed up the calculation by mining the local coupling of indirect refraction of media. Large amounts of light in media-containing scenes are realized by only using sparse samples. Compared with the existing methods, the efficiency of source high-precision and high-efficiency rendering is improved by an order of magnitude.... A new efficient reconstruction technique of incident light field at rendering point is proposed. The importance measure of sampling light field is proposed to balance the discrete light field. The reconstruction problem of incident light field is transformed into the sparse filling problem of the light transfer matrix of the incident light field. The light coupling between adjacent pixels in the same direction of incidence, adaptive sampling, sparse filling and fast joint filtering of the incident light transfer matrix can achieve high-quality and fast reconstruction of the incident light field at the rendered point. First, the problem of large-scale scene light source rendering with out-of-core geometry and light source data is transformed into the problem of planning traversal of light transfer sub-matrix. Then, according to the spatial locality of matrix elements and geometry, the sub-matrix and geometric data are adaptively partitioned to minimize the planning and rendering order of data transmission. The problem is transformed into a graph path search problem. For the first time, a large number of light sources supporting out-of-core and geometric data are rendered in parallel, which improves the scale of the supported scene by an order of magnitude and improves the efficiency of the memory-based CPU algorithm by an order of magnitude.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】