Multi-Scale and Kernel-Predicting Convolutional Networks for Monte Carlo Denoising

Monte Carlo rendering has been widely used in many fields, such as movies, which pursue the photorealistic rendering effect. Monte Carlo rendering needs high sampling rates to get an accurate rendering effect, but the calculation cost is expensive. To keep costs down, one solution is to reduce the noise of the rendered image at reduced sampling rates. Because the traditional denoising method is based on higher and higher order regression models, it is prone to overfitting noise in the input. The Monte Carlo denoising method based on deep learning shows a certain denoising value. In this paper, we propose a kernel-predicting convolutional network with a multi-scale residual structure. Compared with previous methods, our method can extract features and perform residual learning at different scales, which can further remove low-frequency noise and improve the denoising quality.