Methods for deep-learning based super-resolution using high-frequency loss

1. A method comprising:

obtaining a plurality of reference images each of which includes high-frequency texture information and low-frequency texture information;

applying one or more degradations of a first type on the reference images to generate corresponding degraded images, the degraded images including a reduced amount of high-frequency texture information as compared to the corresponding reference images;

generating, from at least a subset of the degraded images, corresponding predicted images using a mapping function associated with a machine learning process to at least partially remove the one or more degradations of the first type;

applying one or more degradations of a second type to each of the predicted images to generate corresponding low-frequency predicted images, the low-frequency predicted images including a reduced amount of high-frequency texture information as compared to the corresponding predicted images;

applying the one or more degradations of the second type to each of the reference images to generate corresponding low-frequency reference images, the low-frequency reference images including a reduced amount of high-frequency texture information as compared to the corresponding reference images;

pixel-wise subtracting each of the low-frequency predicted images from the corresponding predicted images to generate a first set of images;

pixel-wise subtracting each of the low-frequency reference images from the corresponding reference image to generate a second set of images;

computing, based on the first and second set of images, a loss value associated with the machine learning process;

determining that the loss value satisfies a threshold condition; and

responsive to determining that the loss value satisfies a threshold condition, updating the mapping function.

2. The method of claim 1 , further comprising computing, based on the predicted images and corresponding reference images, a pixel-loss value associated with the machine learning process, and wherein computing the loss value associated with the machine learning process is also based on the pixel-loss value.

3. The method of claim 1 , wherein the one or more degradations of the first type include at least one of noise, blur, or resolution downscaling.

4. The method of claim 1 , wherein the one or more degradations of the second type includes a Gaussian blur.

5. The method of claim 1 , wherein at least a portion of the plurality of reference images includes one or more ocular features.

6. The method of claim 1 , wherein the mapping function is configured to restore high-frequency texture information.

7. The method of claim 1 , wherein updating the mapping function includes adjusting one or more parameters of the mapping function to reduce the loss value.

8. The method of claim 1 , wherein computing the loss value includes performing a pixel-wise subtraction between each image in the first set of images and a corresponding image in the second set of images.

9. The method of claim 1 , wherein the first set of images and the second set of images include high-frequency texture information.

10. The method of claim 1 , wherein a degradation of the first type is substantially same as a degradation of the second type.

11. A system, comprising:

a computer-readable memory comprising computer-executable instructions; and

at least one processor executing the computer executable instructions to provide a machine learning module, wherein training of the machine-learning module comprises:

obtaining a plurality of reference images each of which includes high-frequency texture information and low-frequency texture information;

applying one or more degradations of a first type on the reference images to generate corresponding degraded images, the degraded images including a reduced amount of high-frequency texture information as compared to the corresponding reference images;

generating, from at least a subset of the degraded images, corresponding predicted images using a mapping function associated with a machine learning process to at least partially remove the one or more degradations of the first type;

applying one or more degradations of a second type to each of the predicted images to generate corresponding low-frequency predicted images, the low-frequency predicted images including a reduced amount of high-frequency texture information as compared to the corresponding predicted images;

applying the one or more degradations of the second type to each of the reference images to generate corresponding low-frequency reference images, the low-frequency reference images including a reduced amount of high-frequency texture information as compared to the corresponding reference images;

pixel-wise subtracting each of the low-frequency predicted images from the corresponding predicted images to generate a first set of images;

pixel-wise subtracting each of the low-frequency reference images from the corresponding reference image to generate a second set of images;

computing, based on the first and second set of images, a loss value associated with the machine learning process;

determining that the loss value satisfies a threshold condition; and

responsive to determining that the loss value satisfies a threshold condition, updating the mapping function if the loss value satisfies a threshold condition.

12. The system of claim 11 , wherein the machine learning module is configured to: compute, based on the predicted images and corresponding reference images, a pixel-loss value associated with the machine learning process, and wherein computing the loss value associated with the machine learning process is also based on the pixel-loss value.

13. The system of claim 11 , wherein the one or more degradations of the first type include at least one of noise, blur, or resolution downscaling.

14. The system of claim 11 , wherein the one or more degradations of the second type includes a Gaussian blur.

15. The system of claim 11 , wherein at least a portion of the plurality of reference images includes one or more ocular features.

16. The system of claim 11 , wherein the mapping function is configured to restore high-frequency texture information.

17. The system of claim 11 , wherein updating the mapping function includes adjusting one or more parameters of the mapping function to reduce the total loss value.

18. The system of claim 11 , wherein computing the loss value includes performing a pixel-wise subtraction between each image in the first set of images and a corresponding image in the second set of images.

19. The system of claim 11 , wherein the first set of images and the second set of images include high-frequency texture information.

20. One or more non-transitory machine-readable storage devices encoded with instructions configured to cause one or more processing devices to execute operations comprising:

obtaining a plurality of reference images each of which includes high-frequency texture information and low-frequency texture information;

applying one or more degradations of a first type on the reference images to generate corresponding degraded images, the degraded images including a reduced amount of high-frequency texture information as compared to the corresponding reference images;

generating, from at least a subset of the degraded images, corresponding predicted images using a mapping function associated with a machine learning process to at least partially remove the one or more degradations of the first type;

applying one or more degradations of a second type to each of the predicted images to generate corresponding low-frequency predicted images, the low-frequency predicted images including a reduced amount of high-frequency texture information as compared to the corresponding predicted images;

applying the one or more degradations of the second type to each of the reference images to generate corresponding low-frequency reference images, the low-frequency reference images including a reduced amount of high-frequency texture information as compared to the corresponding reference images;

pixel-wise subtracting each of the low-frequency predicted images from the corresponding predicted images to generate a first set of images;

pixel-wise subtracting each of the low-frequency reference images from the corresponding reference image to generate a second set of images;

computing, based on the first and second set of images, a loss value associated with the machine learning process;

determining that the loss value satisfies a threshold condition; and

responsive to determining that the loss value satisfies a threshold condition, updating the mapping function.