IP Library › Granted Patent US 12664694
Granted Patent B1
US 12664694 · App. 17/707,633 · Granted Jun 23, 2026

Neural network-based image compression

Inventor: Vipul Parashar (Pune, IN)
Assignee: NVIDIA Corporation
G06T9/002G06T5/60G06V10/761
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12664694
App. No.
17/707,633
Granted
Jun 23, 2026
Kind
B1
Abstract

Apparatuses, systems, and techniques are presented to compress data. In at least one embodiment, one or more images are generated based, at least in part, upon one or more compressed images and image enhancement data generated using one or more neural networks.

Claims (44)

1 . A processor, comprising:

one or more circuits to:

decode one or more compressed images including first residual information determined in a pixel domain and based, at least in part, on one or more other images;

decode, using one or more neural networks, image enhancement data including second residual information associated with the one or more compressed images and determined in a neural network feature space; and

generate one or more images based, at least in part, upon the decoded one or more compressed images and the decoded image enhancement data such that the generated one or more images are based, at least in part, upon the first residual information determined in the pixel domain and the second residual information determined in the neural network feature space.

2 . The processor of claim 1 , wherein the one or more circuits are further to generate the one or more compressed images using a compression scheme that is indicated to be supported by one or more recipients of the one or more compressed images.

3 . The processor of claim 1 , wherein the one or more circuits are further to generate the image enhancement data, at least in part, by comparing in a feature space of the one or more neural networks the one or more other images to one or more uncompressed versions of the one or more compressed images, wherein the image enhancement data is inferred to increase a similarity between the one or more other images and the one or more uncompressed versions of the one or more compressed images.

4 . The processor of claim 3 , wherein the one or more circuits are further to transmit the image enhancement data as message data encoded with image data for the one or more compressed images.

5 . The processor of claim 4 , wherein the one or more circuits are further to enable an application receiving the image data to decode the one or more compressed images and the image enhancement data, and to modify one or more pixel values of the one or more compressed image using the image enhancement data to generate, for presentation, the one or more images that comprise one or more enhanced compressed images.

6 . The processor of claim 3 , wherein the one or more circuits are further to train the one or more neural networks using a loss function that includes at least a first loss term for similarity and a second loss term for data size, wherein the one or more networks are to be trained to infer a maximum similarity increase within an allowable value of bits per pixel.

7 . A system, comprising:

one or more processors to:

decode one or more compressed images including first residual information determined in a pixel domain and based, at least in part, on one or more other images;

decode, using one or more neural networks, image enhancement data including second residual information associated with the one or more compressed images and determined in a neural network feature space; and

generate one or more images based, at least in part, upon the decoded one or more compressed images and the decoded image enhancement data such that the generated one or more images are based, at least in part, upon the first residual information determined in the pixel domain and the second residual information determined in the neural network feature space.

8 . The system of claim 7 , wherein the one or more processors are further to generate the one or more compressed images using a compression scheme that is indicated to be supported by one or more recipients of the one or more compressed images.

9 . The system of claim 7 , wherein the one or more processors are further to generate the image enhancement data, at least in part, by comparing in a feature space of the one or more neural networks the one or more other images to one or more uncompressed versions of the one or more compressed images, wherein the image enhancement data is inferred to increase a similarity between the one or more other images and the one or more uncompressed versions of the one or more compressed images.

10 . The system of claim 9 , wherein the one or more processors are further to transmit the image enhancement data as message data encoded with image data for the one or more compressed images.

11 . The system of claim 10 , wherein the one or more processors are further to enable an application receiving the image data to decode the one or more compressed images and the image enhancement data, and to modify one or more pixel values of the one or more compressed image using the image enhancement data to generate, for presentation, the one or more images that comprise one or more enhanced compressed images.

12 . The system of claim 9 , wherein the one or more processors are further to train the one or more neural networks using a loss function that includes at least a first loss term for similarity and a second loss term for data size, wherein the one or more networks are to be trained to infer a maximum similarity increase within an allowable value of bits per pixel.

13 . A method, comprising:

decoding one or more compressed images including first residual information determined in a pixel domain and based, at least in part, on one or more other images;

decoding, using one or more neural networks, image enhancement data including second residual information associated with the one or more compressed images and determined in a neural network feature space; and

generating one or more images based, at least in part, upon the decoded one or more compressed images and the decoded image enhancement data such that the generated one or more images are based, at least in part, upon the first residual information determined in the pixel domain and the second residual information determined in the neural network feature space.

14 . The method of claim 13 , further comprising:

generating the one or more compressed images using a compression scheme that is indicated to be supported by one or more recipients of the one or more compressed images.

15 . The method of claim 13 , further comprising:

generating the image enhancement data, at least in part, by comparing in a feature space of the one or more neural networks the one or more other images to one or more uncompressed versions of the one or more compressed images, wherein the image enhancement data is inferred to increase a similarity between the one or more other images and the one or more uncompressed versions of the one or more compressed images.

16 . The method of claim 15 , further comprising:

transmitting the image enhancement data as message data encoded with image data for the one or more compressed images.

17 . The method of claim 16 , further comprising:

enabling an application receiving the image data to decode the one or more compressed images and the image enhancement data, and to modify one or more pixel values of the one or more compressed image using the image enhancement data to generate, for presentation, the one or more images that comprise one or more enhanced compressed images.

18 . The method of claim 15 , wherein the one or more processors are further to train the one or more neural networks using a loss function that includes at least a first loss term for similarity and a second loss term for data size, wherein the one or more networks are to be trained to infer a maximum similarity increase within an allowable value of bits per pixel.

19 . An image generation system, comprising:

one or more processors to:

decode one or more compressed images including first residual information determined in a pixel domain and based, at least in part, on one or more other images;

decode, using one or more neural networks, image enhancement data including second residual information associated with the one or more compressed images and determined in a neural network feature space; and

generate one or more images based, at least in part, upon the decoded one or more compressed images and the decoded image enhancement data such that the generated one or more images are based, at least in part, upon the first residual information determined in the pixel domain and the second residual information determined in the neural network feature space; and

memory for storing network parameters for the one or more neural networks.

20 . The image generation system of claim 19 , wherein the one or more processors are further to generate the one or more compressed images using a compression scheme that is indicated to be supported by one or more recipients of the one or more compressed images.

21 . The image generation system of claim 19 , wherein the one or more processors are further to generate the image enhancement data, at least in part, by comparing in a feature space of the one or more neural networks the one or more other images to one or more uncompressed versions of the one or more compressed images, wherein the image enhancement data is inferred to increase a similarity between the one or more other images and the one or more uncompressed versions of the one or more compressed images.

22 . The image generation system of claim 21 , wherein the one or more processors are further to transmit the image enhancement data as message data encoded with image data for the one or more compressed images.

23 . The image generation system of claim 22 , wherein the one or more processors are further to enable an application receiving the image data to decode the one or more compressed images and the image enhancement data, and to modify one or more pixel values of the one or more compressed image using the image enhancement data to generate, for presentation, the one or more images that comprise one or more enhanced compressed images.

24 . The image generation system of claim 21 , wherein the one or more processors are further to train the one or more neural networks using a loss function that includes at least a first loss term for similarity and a second loss term for data size, wherein the one or more networks are to be trained to infer a maximum similarity increase within an allowable value of bits per pixel.