Efficient Denoising Method For Monte Carlo Rendering Algorithm

In recent years,physically based realistic image synthesis techniques have been widely used in industries such as games,animation,and film and TV special effects.Among them,Monte Carlo rendering algorithms have become the mainstream image synthesis techniques because they can meet strict quality requirements,are easy to implement,and are robust.However,to produce high-quality,noise-free results,Monte Carlo rendering algorithms often need to set very high sampling rates,which means high computational costs and huge computation time.Fortunately,various denoising methods have been shown to help Monte Carlo rendering algorithms achieve high quality rendering results at lower sample rates.In the field of offline rendering,Monte Carlo rendering algorithms have long been the mainstay of the field,and there has been a wealth of research on denoising methods for unbiased Monte Carlo rendering algorithms,such as path tracing.These post-processing noise reduction methods in image space can effectively remove the small noise common to unbiased Monte Carlo rendering algorithms at low sampling rates.However,such denoising methods tend to perform poorly when applied to biased Monte Carlo rendering algorithms(e.g.,stochastic progressive photon mapping)with their multi-scale noise,while there are few efficient noise reduction methods specifically designed for biased Monte Carlo rendering algorithms.On the other hand,in the field of real-time rendering,real-time Monte Carlo rendering algorithms(e.g.,real-time ray tracing)have become a hot research topic as the demand for realistic rendering continues to increase.Real-time rendering applications generally require more than 30 frames per second,which requires Monte Carlo rendering algorithms to use very low sampling rates,and the result will introduce a lot of noise.Therefore,the key to the success of real-time Monte Carlo rendering algorithms is to incorporate denoising methods.Recently,a series of time-domain denoising methods have been proposed,inspired by the time-domain inverse sampling methods of traditional rasterization algorithms,which use a motion vector to capture inter-frame valid information for noise reduction.However,when the motion vector fails,the noise reduction effect will be greatly reduced.In this paper,we propose a stochastic progressive photon mapping denoising method based on multiple residual networks and photon related features and a time-domain denoising method based on real-time ray tracing with accurate motion vector calculation for the noise problem in Monte Carlo rendering calculations.The main work of this paper as follows:We propose an efficient denoising method for the biased Monte Carlo rendering algorithm to address the multiscale noise problem prevalent in Monte Carlo stochastic progressive photon mapping algorithms.The method uses multiple residual blocks with multiple residual functions,which can effectively handle the multiscale noise in the random progressive photon mapping rendering results;a variety of photon-related auxiliary features are used to better preserve important lighting details such as caustics.We also propose a series of motion vectors for difficult cases such as shadows,glossy reflections and occlusions for the time-domain denoising problem of Monte Carlo real-time ray tracing.They capture and exploit the effective time-domain information between frames to effectively solve the problems of ghosting artifacts and lagging illumination details;compared with the state-of-the-art methods in the field,the method proposed in this paper is more reliable and stable in the time domain,suitable for complex scenes with a large number of geometric occlusion relations,for fast moving light sources and cameras,and with almost no additional overhead.In summary,a stochastic progressive photon mapping denoising method and a real-time ray tracing time-domain denoising method based on multiple motion vectors are proposed in this paper.The related method research effectively solves the noise problem in the results of biased and real-time Monte Carlo rendering algorithms,and provides ideas for denoising research in this field;the related method can be widely used in film and animation,games and 3D simulation.