Unsupervised training of neural network for high dynamic range image compression

Patent metadata

Field	Value
Publication number	US-11295423-B2
Application number	US-201916694286-A
Country	US
Kind code	B2
Filing date	Nov 25, 2019
Priority date	Nov 25, 2019
Publication date	Apr 5, 2022
Grant date	Apr 5, 2022

Abstract

Techniques are provided for unsupervised training of a neural network to perform compression of a high dynamic range (HDR) image. A methodology implementing the techniques according to an embodiment includes performing global tone mapping on an HDR training image to generate a low dynamic range (LDR) training image. The method also includes applying the neural network to the HDR training image and the LDR training image to generate a delta image representing image detail lost in the global tone mapping operation. The method further includes summing the delta image with the LDR training image to generate an output training image, and generating a loss function calculated from a weighted sum of a contrast loss and a compression loss. The contrast loss is based on the output training image and the HDR training image, and the compression loss is based on the output training image and the LDR training image.

First claim

What is claimed is:

1. A processor-implemented method for training a neural network to perform high dynamic range (HDR) image compression, the method comprising:
performing, by a processor-based system, global tone mapping on an HDR training image to generate a low dynamic range (LDR) training image;
applying, by the processor-based system, the neural network to the HDR training image and the LDR training image, the neural network to generate a delta image, the delta image representing image detail lost between the HDR training image and LDR training image;
summing, by the processor-based system, the delta image with the LDR training image to generate an output training image;
calculating, by the processor-based system, a contrast loss, the contrast loss based on the output training image and the HDR training image;
calculating, by the processor-based system, a compression loss, the compression loss based on the output training image and the LDR training image; and
calculating, by the processor-based system, a loss function based on a weighted sum of the contrast loss and the compression loss.

2. The method of claim 1 , further including performing iterative backpropagation training of the neural network based on the loss function, such that the output training image converges to a compressed HDR image.

3. The method of claim 2 , wherein the weighted sum is based on a parameter selected to control a ratio of compression to preservation of contrast in the compressed HDR image.

4. The method of claim 1 , wherein the contrast loss is calculated as a rectilinear distance between a calculated contrast of the output training image and a calculated contrast of the HDR training image.

5. The method of claim 4 , wherein the calculated contrast of an image is calculated as a logarithm of a ratio of intensity of adjacent pixels of the image.

6. The method of claim 1 , wherein the compression loss is calculated as a rectilinear distance between a logarithm of the output training image and a logarithm of the LDR training image.

7. The method of claim 1 , wherein the neural network is a convolutional neural network.

8. The method of claim 1 , wherein the training of the neural network is unsupervised.

9. At least one non-transitory computer readable storage medium comprising instructions that, when executed, cause at least one processor to at least:
perform global tone mapping on an HDR training image to generate a low dynamic range (LDR) training image;
apply a neural network to the HDR training image and the LDR training image, the neural network to generate a delta image, the delta image representing image detail lost between the HDR training image and the LDR training image;
sum the delta image with the LDR training image to generate an output training image;
calculate a contrast loss, the contrast loss based on the output training image and the HDR training image;
calculate a compression loss, the compression loss based on the output training image and the LDR training image; and
calculate a loss function based on from a weighted sum of the contrast loss and the compression loss.

10. The computer readable storage medium of claim 9 , wherein the instructions, when executed, cause the at least one processor to perform iterative backpropagation training of the neural network based on the loss function, such that the output training image converges to a compressed HDR image.

11. The computer readable storage medium of claim 10 , wherein the weighted sum is based on a parameter selected to control a ratio of compression to preservation of contrast in the compressed HDR image.

12. The computer readable storage medium of claim 9 , wherein the contrast loss is calculated as a rectilinear distance between a calculated contrast of the output training image and a calculated contrast of the HDR training image.

13. The computer readable storage medium of claim 12 , wherein the calculated contrast of an image is calculated as a logarithm of a ratio of intensity of adjacent pixels of the image.

14. The computer readable storage medium of claim 9 , wherein the compression loss is calculated as a rectilinear distance between a logarithm of the output training image and a logarithm of the LDR training image.

15. The computer readable storage medium of claim 9 , wherein the neural network is a convolutional neural network.

16. The computer readable storage medium of claim 9 , wherein the training of the neural network is unsupervised.

17. An apparatus comprising:
memory;
instructions;
at least one processor to execute the instructions to:
perform global tone mapping on an HDR training image to generate a low dynamic range (LDR) training image;
apply the neural network to the HDR training image and the LDR training image, the neural network to generate a delta image, the delta image representing image detail lost between the HDR training image and the LDR training image;
sum the delta image with the LDR training image to generate an output training image;
calculate a contrast loss, the contrast loss based on the output training image and the HDR training image;
calculate a compression loss, the compression loss based on the output training image and the LDR training image; and
calculate a loss function based on a weighted sum of the contrast loss and the compression loss.

18. The apparatus of claim 17 , wherein the processor is to perform iterative backpropagation training of the neural network based on the loss function, such that the output training image converges to a compressed HDR image.

19. The apparatus of claim 18 , wherein the weighted sum is based on a parameter selected to control a ratio of compression to preservation of contrast in the compressed HDR image.

20. The apparatus of claim 17 , wherein the contrast loss is calculated as a rectilinear distance between a calculated contrast of the output training image and a calculated contrast of the HDR training image.

Assignees

• Intel Corp
• Intel Corportation

Inventors

• Bugdary Shlomo
• Zatzarinni Rony

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/10004

  Still image; Photographic image · CPC title

• G06T9/002

  using neural networks · CPC title