Generation of high dynamic range visual media

1. A method for visual media processing, comprising:

generating a first set of images based on performing an alignment of a set of input images having a first dynamic range, the first set of images having the first dynamic range;

generating a second set of images having a second dynamic range that is greater than the first dynamic range;

aligning the second set of images to generate a third set of images having the second dynamic range; and

generating, using a first convolutional neural network (CNN) and based on at least the third set of images, a final image having the second dynamic range,

wherein the alignment of the set of input images is performed using a plurality of sub-CNNs, wherein each of the plurality of sub-CNNs operates at a distinct resolution, and wherein at least one of the plurality of sub-CNNs is followed by a rectified linear unit (ReLU).

2. The method of claim 1 , wherein each of the first set of images having the first dynamic range have a same exposure or different exposures.

3. The method of claim 1 , wherein performing the alignment of the set of input images is based on an optical flow algorithm.

4. The method of claim 1 , wherein performing the alignment of the set of input images is further based on a plurality of optical flows generated using a corresponding sub-CNN of the plurality of sub-CNNs.

5. The method of claim 4 , further comprising:

generating, using the first CNN with an input comprising the first set of images, a set of estimated images having the second dynamic range; and

minimizing an error between a set of ground truth images and the set of estimated images.

6. The method of claim 1 , wherein a last of the plurality of sub-CNNs is followed by a linear activation function.

7. An apparatus for visual media processing, comprising:

a processor; and

a memory that comprises instructions stored thereupon, wherein the instructions when executed by the processor configure the processor to:

generate a first set of images based on performing an alignment of a set of input images having a first dynamic range, the first set of images having the first dynamic range;

generate a second set of images having a second dynamic range that is greater than the first dynamic range;

align the second set of images to generate a third set of images having the second dynamic range; and

generate, using a first convolutional neural network (CNN) and based on at least the third set of images, a final image having the second dynamic range,

wherein the alignment of the set of input images is performed using a plurality of sub-CNNs, wherein each of the plurality of sub-CNNs operates at a distinct resolution, and wherein at least one of the plurality of sub-CNNs is followed by a rectified linear unit (ReLU).

8. The apparatus of claim 7 , wherein each of the first set of images having the first dynamic range have a same exposure or different exposures.

9. The apparatus of claim 7 , wherein performing the alignment of the set of input images is based on an optical flow algorithm.

10. The apparatus of claim 7 , wherein performing the alignment of the set of input images is further based on a plurality of optical flows generated using a corresponding sub-CNN of the plurality of sub-CNNs.

11. The apparatus of claim 7 , wherein the processor is further configured to:

generate, using the first CNN with an input comprising the first set of images, a set of estimated images having the second dynamic range; and

minimize an error between a set of ground truth images and the set of estimated images.

12. The apparatus of claim 7 , wherein a last of the plurality of sub-CNNs is followed by a linear activation function.

13. A non-transitory computer readable program storage medium having code stored thereon, the code, when executed by a processor, causing the processor to implement a method for visual media processing, the method comprising:

generating a first set of images based on performing an alignment of a set of input images having a first dynamic range, the first set of images having the first dynamic range;

generating a second set of images having a second dynamic range that is greater than the first dynamic range;

aligning the second set of images to generate a third set of images having the second dynamic range; and

generating, using a first convolutional neural network (CNN) and based on at least the third set of images, a final image having the second dynamic range,

wherein the alignment of the set of input images is performed using a plurality of sub-CNNs, wherein each of the plurality of sub-CNNs operates at a distinct resolution, and wherein at least one of the plurality of sub-CNNs is followed by a rectified linear unit (ReLU).

14. The non-transitory computer readable program storage medium of claim 13 , wherein performing the alignment of the set of input images is further based on a plurality of optical flows generated using a corresponding sub-CNN of the plurality of sub-CNNs.

15. The non-transitory computer readable program storage medium of claim 14 , wherein the method further comprises training the first CNN based on:

generating, using the first CNN with an input comprising the first set of images, a set of estimated images having the second dynamic range; and

minimizing an error between a set of ground truth images and the set of estimated images.

16. The non-transitory computer readable program storage medium of claim 13 , wherein each of the first set of images having the first dynamic range have a same exposure or different exposures.

17. The non-transitory computer readable program storage medium of claim 13 , wherein performing the alignment of the set of input images is based on an optical flow algorithm.