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

A method for constructing a denoising model of Monte Carlo rendering based on a Generative Adversarial Network (GAN), comprising the following steps:

• obtaining Monte Carlo rendering image containing noise as a noise rendering image, obtaining auxiliary features when generating the noise rendering image, and using the noise rendering image and the corresponding auxiliary features, and a target rendering image corresponding to the noise rendering image as a training sample;
• constructing a Generative Adversarial Network (GAN), wherein the Generative Adversarial Network (GAN) comprises:
• a denoising network and a discriminant network, wherein the denoising network is used to denoise the input noise rendering image and auxiliary features, and output the denoising rendering image, and the discriminant network is used to classify the input denoising rendering image and the target rendering image corresponding to the noise rendering image, and output the classification result;
• the training samples are used to tune the network parameters of the Generative Adversarial Network (GAN);
• after the tuning is completed, the denoising network determined by the network parameters is used as the Monte Carlo rendering image denoising model;
• wherein the denoising network comprises:
• an auxiliary graph feature extraction sub-network, wherein the auxiliary graph feature extraction sub-network is a Convolutional Neural Network (CNN) comprising at least one convolutional layer, and is used to fuse input auxiliary features and output auxiliary feature maps;
• a rendering map feature extraction sub-network, wherein the rendering map feature extraction sub-network is a Convolutional Neural Network (CNN) including at least one convolutional layer, used for extracting features of a noise rendering map, and outputting a noise feature map;
• a feature fusion sub-network, wherein the feature fusion sub-network is a neural network that uses residual thought and uses convolutional layers to fuse and extract auxiliary feature maps and noise feature maps.