Systems and methods for tone mapping of high dynamic range images for high-quality deep learning based processing

1. A graphics processor, comprising:

a media pipeline to generate media requests for processing images; and

an execution resource to receive media requests from the media pipeline, the execution resource is configured to compute an auto-exposure scale for an image to effectively tone map the image, to scale the image based on multiplying pixel values with the computed auto-exposure scale, and to apply a tone mapping operator including a log function to scaled values of the image to compress a range of values for the image to generate a tone mapped image.

2. The graphics processor of claim 1 , wherein the execution resource is configured to scale the log function to generate the tone mapped image.

3. The graphics processor of claim 2 , wherein the execution resource is configured to apply a gamma correction to the tone mapped image to generate a gamma corrected tone mapped image and to make a tone mapping curve more perceptually linear for the image.

4. The graphics processor of claim 3 , wherein the execution resource is configured to provide the tone mapped image as input for a neural network.

5. The graphics processor of claim 4 , wherein the neural network processes the tone mapped image to generate an output of the neural network.

6. The graphics processor of claim 5 , wherein the execution resource is configured to apply an inverse tone mapping operator and then inverse exposure scale to this output of the neural network to generate the final output image.

7. A graphics processing unit, comprising:

memory to store graphics data for images; and

a core having a plurality of processing resources for performing graphics operations, wherein at least one processing resource is configured to downsample the image in at least one dimension, to compute an auto-exposure scale for an image to effectively tone map the image, to scale the image based on multiplying pixel values with the computed auto-exposure scale, and to apply a tone mapping operator including a log function to scaled values of the image to compress a range of values for the image to generate a tone mapped image.

8. The graphics processing unit of claim 7 , wherein the at least one processing resource is configured to scale the log function to generate the tone mapped image and to apply a gamma correction to the tone mapped image to generate a gamma corrected tone mapped image.

9. The graphics processing unit of claim 8 , wherein the at least one processing resource is configured to provide the tone mapped image as input for a convolution neural network (CNN).

10. The graphics processing unit of claim 9 , wherein the neural network processes the tone mapped image with a denoising algorithm, anti-aliasing algorithm, super resolution, or demosaicing to generate an output of the neural network.

11. The graphics processing unit of claim 10 , wherein the at least one processing resource is configured to apply an inverse tone mapping operator and then inverse exposure scale to this output of the neural network to generate an output image having a high dynamic range (HDR) of the image.

12. The graphics processing unit of claim 7 , wherein the tone mapping operator comprises (log 2(x+1)/16)1/2.2.

13. The graphics processing unit of claim 12 , wherein values of the log function are scaled by 1/16 such that floating-point values are compressed between 0 and 1.

14. A non-transitory machine-readable storage medium storing sing executable instructions which when executed by at least one processing resource cause the at least one processing resource to perform a computer implemented method for tone mapping images comprising:

computing, with the at least one processing resource of a graphics processing unit (GPU), an auto-exposure scale for an image to effectively tone map the image;

scaling the image with the computed auto-exposure scale; and

applying a tone mapping operator including a log function to scaled values of the image and scaling the log function to generate a tone mapped image.

15. The non-transitory machine-readable storage medium of claim 14 , the computer-implemented method further comprising:

applying a gamma correction to the tone mapped image to generate a gamma corrected tone mapped image and generating a tone mapping curve that is perceptually linear.

16. The non-transitory machine-readable storage medium of claim 15 , the computer-implemented method further comprising:

providing the tone mapped image as input for a convolutional neural network (CNN); and

processing including one of denoising algorithm and anti-aliasing algorithm the tone mapped image to generate an output of the CNN.

17. The non-transitory machine-readable storage medium of claim 16 , wherein the processing includes applying tone mapping to the image and target images, and computing a loss directly in the tone mapped color space.

18. The non-transitory machine-readable storage medium of claim 17 , the computer-implemented method further comprising:

applying an inverse of the tone mapping operator to the output of the CNN.

19. The non-transitory machine-readable storage medium of claim 18 , the computer-implemented method further comprising:

applying an inverse of the exposure scale to the output of the CNN to generate the final output image.

20. The non-transitory machine-readable storage medium of claim 15 , the computer-implemented method further comprising:

downsampling the image in multiple dimensions.