Method and apparatus for improving quality and realism of rendered image
View Patent ↗Disclosed herein is a method for improving the quality and realism of a rendered image. The method includes receiving training data including a real image and a rendered image, generating a low-quality image using the training data, generating a high-quality image using the low-quality image, generating a realistic image using the high-quality image, and training a neural network using an error calculated based on the high-quality image and the realistic image.
1 . A method for improving quality and realism of a rendered image, comprising:
receiving training data, including a real image dataset and a rendered image dataset, wherein the real image dataset includes a real image and the rendered image dataset includes a rendered image generated from 3D graphics rendering;
generating a low-quality image using the training data, wherein the low-quality image is generated by degrading the quality of the image of the training data;
generating a high-quality image using the low-quality image;
generating a realistic image using the high-quality image; and
training a neural network using an error calculated based on the realistic image, wherein the error is defined differently depending on whether the training data is the real image or the rendered image,
wherein the degrading includes performing a color distortion, an addition of Gaussian noise, an image compression and a decrease in resolution on the training data,
wherein the error includes an adversarial generation error, a pixel restoration error and a perceptual restoration error,
wherein the adversarial generation error is calculated based on the real image and the realistic image, the perceptual restoration error is calculated based on the real image, the rendered image and the realistic image, and
wherein the neural network is trained using the adversarial generation error the pixel restoration error and the perceptual restoration error to enhance both image quality and photorealism simultaneously for real-time performance capture applications.
2 . The method of claim 1 , wherein the error calculated based on the high-quality image and the realistic image includes a quality restoration error between the generated high-quality image and training data transformed to correspond to the high-quality image.
3 . The method of claim 1 , wherein the adversarial generation error is calculated using a generative adversarial network structure, which uses a neural network for generating the realistic image as a generator.
4 . The method of claim 3 , wherein the generative adversarial network structure uses a real image as a ground truth when an image input thereto is a low-quality image generated based on the real image, and uses an arbitrary real image in the training data as a ground truth when the image input thereto is a low-quality image generated based on a rendered image, because a real image corresponding thereto is not present.
5 . The method of claim 1 , wherein the pixel restoration error is calculated by applying a distance equation in pixel units between the real image, the rendered image and the realistic image, and the perceptual restoration error is calculated based on respective feature vectors of the real image, the rendered image and the realistic image.
6 . The method of claim 5 , wherein the pixel restoration error is calculated based on a first distance equation when the training data is a real image, and is calculated based on a second distance equation, which differs from the first distance equation, when the training data is a rendered image.
7 . The method of claim 1 , wherein the error calculated based on the high-quality image and the realistic image includes an identity preservation error between the low-quality image and the realistic image, which is calculated using a pretrained identification neural network.
8 . An apparatus for improving quality and realism of a rendered image, comprising:
a processor;
a low-quality image generator configured to generate, via the processor, a low-quality image using training data, including a real image dataset and a rendered image dataset, wherein the real image dataset includes a real image and the rendered image dataset includes a rendered image generated from 3D graphics rendering;
an image restorer configured to generate, via the processor, a high-quality image using the low-quality image, wherein the low-quality image is generated by degrading the quality of the image of the training data;
a realistic-image generator configured to generate, via the processor, a realistic image using the high-quality image; and
an error analyzer configured to train, via the processor, a neural network using an error calculated based on the realistic image, wherein the error is defined differently depending on whether the training data is the real image or the rendered image,
wherein the degrading includes performing a color distortion, an addition of Gaussian noise, an image compression and a decrease in resolution on the training data,
wherein the error includes an adversarial generation error, a pixel restoration error and a perceptual restoration error,
wherein the adversarial generation error is calculated based on the real image and the realistic image, the perceptual restoration error is calculated based on the real image, the rendered image and the realistic image, and
wherein the neural network is trained using the adversarial generation error, the pixel restoration error and the perceptual restoration error to enhance both image quality and photorealism simultaneously for real-time performance capture applications.
9 . The apparatus of claim 8 , wherein the error analyzer includes a quality error analyzer configured to calculate, via the processor, a quality restoration error between the generated high-quality image and training data transformed to correspond to the high-quality image.
10 . The apparatus of claim 8 , wherein the error analyzer includes an adversarial generation error analyzer configured to calculate, via the processor, the adversarial generation error using a generative adversarial network structure, which uses a neural network for generating the realistic image as a generator.
11 . The apparatus of claim 10 , wherein the generative adversarial network structure uses a real image as a ground truth when an image input thereto is a low-quality image generated based on the real image, and uses an arbitrary real image in the training data as a ground truth when the image input thereto is a low-quality image generated based on a rendered image, because a real image corresponding thereto is not present.
12 . The apparatus of claim 8 , wherein the error analyzer includes a realistic-image error analyzer configured to calculate, via the processor, the pixel restoration error by applying a distance equation in pixel units between the real image, the rendered image and the realistic image, and the perceptual restoration error based on respective feature vectors of the real image, the rendered image and the realistic image.
13 . The apparatus of claim 12 , wherein the realistic-image error analyzer calculates, via the processor, the pixel restoration error based on a first distance equation when the training data is a real image, and calculates, via the processor, the pixel restoration error based on a second distance equation, which differs from the first distance equation, when the training data is a rendered image.
14 . The apparatus of claim 8 , wherein the error analyzer includes an identity error analyzer configured to calculate, via the processor, an identity preservation error between the low-quality image and the realistic image using a pretrained identification neural network.