Interactive Global Illumination With The Modified Radiance Regression Functions

Global illumination is very important part in computer graphic and photorealistic Rendering, because the photorealistic of an image is judged by the lighting and shadow effect. Nowadays, most of the global illumination can produce an image which can fake human eyes, but they all need to spend a lot of times. Some applications which have a high require on render time always need to reduce the effect in order to increase the speed, so interactive real time global illumination which have good effect is always the hot and difficult question on research.This thesis analyzes a lot of algorithms in global illumination about their advantage and disadvantage. An algorithm called Transparent Supported Radiance Regression Function(TsRRF) is designed and realized which can support interactive global illumination in real-time.The main research and contributions of this paper include the following aspects:(1) Design the framework of the global illumination algorithm supporting object moving and transparent object rendering. This algorithm which based on radiance regression function, solve the imperfection of primary algorithm which can only support opaque object, and by include new parameter, the object moving is also supported.(2) Using Multi-layer deep learning architecture replace the neural network which will be caught into local optimum easily. In this way, this algorithm becomes more general and universalistic. It can also decrease the over fitting problem which may cause some bias result in rendering.(3) Importing the space dividing method by property. Using different deep learning architectures to render different lights like reflect light and refract light. Detailing the classification in deep learning, in order to make the reflect light and refract light rendered without the influence of diffuse light, and in this way, the transparent object can be rendered successfully.(4) Building all of the algorithms on GPU and using the parallel computing of GPGPU to speed up the process of ray tracing, in order to make the ray tracing in real time when the iterations are few.(5) By experiment, the algorithm is proved that it can describe the reflected and refracted effect effective. And the improved algorithm not only keeps the advantage of the radiance regression function, but also supports object moving and transparent object rendering in real time. At last, it can render a full global illumination effects with the model of Sponza in 512*512 images at 21~25 FPS.