Denoising Of Monte Carlo Renderings Based On Kernel Prediction And Feature-preserved GAN

Monte Carlo rendering is popular in photorealistic image generation due to its flexibility and universality.However,the algorithm is too time-consuming to generate high-quality image when needing to trace a huge number of light paths.Besides,it is inevitable to generate noises when sampling a limited light path randomly.To overcome this problem,a compromise is to increasing the sampling rate and rendering time,another is to generate a noisy scene by the Monte Carlo method fed with a limited number of rays,and then apply a specific filter on the noisy image to obtain high-quality rendering results.This paper proposes a novel end-to-end denoising network for Monte Carlo renderings.There are two recent two major camps of learning-based methods for denoising:estimating the filter kernel,and direct learning how to produce a denoised output from an noisy input.The key idea of this paper is to combine these two ideas:train a GAN to denoise a noisy rendering,and apply a learned kernel for further improvement.This network consists of three components that can fulfill the filtering.Firstly,a kernel prediction denoising network(KPDnet)is employed to estimate a filtering kernel for each pixel of the noisy image according to the auxiliary features information.Secondly,a Generative Adversarial Direct Denoising Network(GAND~2Net)component is used to perform preliminary denoising,responsible for extracting structure and detailed features from the noisy image.Finally,a reconstruction part is designed to apply the kernels to the preliminary-denoised image which outputs the final result.To make these components work together,this paper also propose a new GAN loss function that penalizes the deviation of the reconstructed result from the ground truth.Besides,this paper also use a random scene generation method for constructing a powerful dataset,which enhance the stability of the training process.Experiments demonstrate that,compared to the state-of-the-art approaches,our algorithm is more effective in recovering scene structure and details while less sensitive to the sampling rate of the noisy image and keeping a stable running time.