Sectional Multi-Layer Convolutional Neural Network Based Real-Time Denoiser for Low Sampling Rate Rendering

The application scenarios of global illumination rendering have become more and more extensive. However, due to the limited computing capability of computer hardware, high samples-per-pixel (SPP) rendering will lead to terrible user experience. We offered a solution using the deep learning method to restore low SPP pictures through sectional convolutional neural networks (CNNs) and auxiliary inputs. Our model achieved better restoration effects and real-time denoising under the premise of reducing sampling cost. Specifically, we used a structure with multiple skip connections based on recurrent convolutional neural network (RCNN) to handle the input picture sequence, which overcame the shortage of picture denoiser in terms of time stability. To deal with the limited reconstruction ability of RCNN, we used multiple denoising convolutional neural network (DnCNN) layers to further process the output data. In order to obtain a better real-time denoising effect, we introduced auxiliary inputs to our model, including albedo maps, normal vector maps, depth maps and shadowing maps gathered from the low SPP rendering process. Our model synthesized the virtue of picture denoiser and video denoiser. On the 5 different customized scenes we made, our model shows gratifying time stability and outperforms the widely used OptiX denoiser. Using the structural similarity (SSIM) as objective indicator, our model shows 5.8%, 12.2%, 1.5%, 4.7%, 1.8% improvement on the 5 different scenes respectively when compared with OptiX denoiser.