Filter And Sampling Algorithms For Global Illumination

Global illumination algorithms can synthesize realistic images. Both direct lighting and indirect lighting can be calculated with global illumination algorithms. Direct lighting is the direct contribution of light source on objects; indirect lighting is the contribution of light which is reflected or refracted several times. Global illumination algorithms are widely used in areas such as architecture, design, entertainment and so on. In these areas, both the geometries and the properties of materials are very complex. Monte Carlo ray tracing based global illumination techniques are able to handle this complexity. However, the rendered images are usually noisy when the number of samples is inadequate. There are two ways to reduce noise. One way is filtering after sampling; the other is improving the efficiency of sampling algorithm. Proceeding from the above two aspects, novel filter and sampling algorithms were presented in the thesis.In filter area, existing algorithms often blur the borders of objects and textures, and low frequency noise remains in the filtered images. A noval filter algorithm based on reusing incident radiance samples was proposed in the thesis. The borders of objects and textures were not blured because incident radiance samples were reused in the filter process. The distance information around the sampling point was taken into accont and large filter window was used to remove low frequence noise. Adaptive reuse probability was calculated in order to reduce time cost of the filter process. The validity of the algorithm was verified by four scene tests. Compared to existing algorithms, our algorithm achieved better image quality.Two kinds of sampling algorithms were proposed in this thesis. One is the hemispherical iterative importance sampling; the other is the image plane adaptive sampling. Sampling according to the incident radiance distribution is an efficient method to improve the efficiency of hemispherical sampling. However, existing algorithms dose not take into account the occlusion of light source by the objects in the scene, and can not be applied in indirect lighting. In our algorithm, the scene was divided into subregions with a 5D tree. Within each subregion, the incident radiance distribution was calculated with samples which were recorded in the sampling process. Two kinds of tree-style importance functions were proposed: spherical 2D tree and spacial 3D tree, which were used to describe the incident radiance distribution of indirect lighting and direct lighting respectively. In order to improve sampling efficiency, mixed transition kernel was formed with tree-style importance function, bidirectional reflection distribution function and light energy distribution function in Population Monte Carlo framework.The variance and time cost of direct and indirect lighting were taken into account in image plane adaptive sampling. The sample numbers of direct and indirect lighting for each pixel were set adaptively according to Lagrange multiplier method, so as to minimize the variance of the image in a given time. The hemispherical sampling algorithm and the image plane sampling algorithm were combined together to further reduce variance. The validity of the algorithm was verified by four scene tests. The sampling efficiency of the new algorithm was markedly improved compared to the existing algorithms.