Spherical blue noise sampling and deep learning based de-noising techniques for Monte Carlo rendering

• Other Authors: Wong, Kin Ming (author.), Wong, Tien-Tsin (thesis advisor.), Chinese University of Hong Kong Graduate School. Division of Computer Science and Engineering. (degree granting institution.)
• Format: Text
• Language: English; Chinese
• Published: 2018
• Subjects: Computer graphics; Rendering (Computer graphics); Ray tracing algorithms; Three-dimensional display systems; Image processing > Digital techniques; T385 .W6449 2018eb; sami
• Online Access: https://julac.hosted.exlibrisgroup.com/primo-explore/search?query=addsrcrid,exact,991039750396103407,AND&tab=default_tab&search_scope=All&vid=CUHK&mode=advanced&lang=en_US; https://repository.lib.cuhk.edu.hk/en/item/cuhk-2187971

Ph.D. Photo-realistic rendering is central to a wide range of modern visualization based computer applications such as video games, computer animation, visual effects in movies and virtual reality etc. Monte Carlo based rendering methods offer a streamlined approach to synthesize high quality photorealistic images through physically accurate numerical simulation, and it has recently become the mainstream photo-realistic rendering method. However, Monte Carlo rendering method often requires a large number of samples in order to produce visually noise-free results. This thesis proposes to improve the quality of Monte Carlo renderings through two different directions. We propose a novel sampling method which actively improves the image quality during the Monte Carlo integration process, and then we further enhance the rendered image quality by post-processing using our novel deep learning based de-noising method. Given a fixed number of samples, the key to high quality Monte Carlo rendering results is the random and yet uniform distribution of sample points over the sampling domain. Current practices rely on mapping 1D or unit square uniform distribution to the sampling domain, and this often results sub-optimal uniform distributions. We propose a novel sampling method which generates high quality unstructured uniform blue noise pattern directly on the spherical surface which is the most common sampling domain for most Monte Carlo rendering methods. Our method models the sample points as a collection of identical electrically charged particles on the sami pling domain. The repulsive force exerted on the particles from its neighborhood propels the system to undergo a self-organizing process to minimize the force by movement. The resultant stationary state reveals a uniform but unstructured distribution of points with a blue noise characteristic. Unlike the number sequence or jitter based methods, our physically based approach works with arbitrary number of sample points. In addition, the distribution ...