Monte Carlo rendering graph denoising model, method and device based on generative adversarial network

Abstract

The invention discloses a Monte Carlo rendering graph denoising model based on a generative adversarial network and a construction method of the Monte Carlo rendering graph denoising model. The methodcomprises the steps that a training sample is constructed, a generative adversarial network is constructed, the generative adversarial network comprises a denoising network and a discrimination network, the denoising network is used for denoising an input noise rendering graph and auxiliary features and outputting the denoising rendering graph, and the discrimination network is used for classifying the input denoising rendering graph and a target rendering graph corresponding to the noise rendering graph and outputting a classification result; according to the Monte Carlo rendering graph denoising method and device, the network parameters of the generative adversarial network are adjusted and optimized through the training sample, the denoising network determined through the network parameters serves as the Monte Carlo rendering graph denoising model, the Monte Carlo rendering graph denoising method and device are further disclosed, and denoising of the Monte Carlo rendering graph containing noise can be achieved.

Description

technical field [0001] The invention belongs to the field of image denoising, and in particular relates to a denoising model, method and device of a Monte Carlo rendering image based on a generative confrontation network. Background technique [0002] The rendering technology based on Monte-Carlo Simulation consumes a lot of time and computing resources because the convergence of the variance of the rendering image requires a large number of samples. In order to save computing resources and reduce rendering time, generally a lower sampling rate is used to render a rendered image with noise, and then a certain denoising technology is used to reduce the noise of the rendered image to obtain a noise-free, visual performance. Better renderings. [0003] At present, the more cutting-edge denoising techniques for Monte Carlo renderings are mostly based on deep learning. The most commonly used method is to use convolutional neural network to denoise the Monte Carlo rendering imag...

AI Technical Summary

Problems solved by technology

[0004] Disney's "Bako S, Vogels T, McWilliams B, et al. Kernel-predicting convolutional networks for denoising Monte Carlo renderings[J]. ACMTransactions on Graphics (TOG), 2017,36(4):97." and Nvidia's "Chaitanya C R A, Kaplanyan A S, Schied C, et al.Interactive reconstruction of Monte Carlo imagesequences using a recurrent denoising autoencoder[J].ACM Transactions on Graphics(TOG),2017,36(4):98."Since pixel-level loss is used as It is difficult to accurately describe the real human visual experience when optimizing the goal, so even if the standard is high on this optimization goal, relatively blurred or low-reduced high-frequency details will often be obtained, making the denoised Monte Card Luo's rendering lacks realism in the details, and even some places with more high-frequency details will appear dirty

For example, after denoising the interior rendering, the corners and baseboards of the suspended ceiling in the interior rendering will make places with more high-frequency details dirty.

Images