Motion deblurring using image upsampling

Patent metadata

Field	Value
Publication number	US-9042672-B2
Application number	US-201213480119-A
Country	US
Kind code	B2
Filing date	May 24, 2012
Priority date	Sep 2, 2011
Publication date	May 26, 2015
Grant date	May 26, 2015

Abstract

Various embodiments of methods and apparatus for motion deblurring are disclosed. In one embodiment, an estimate of a latent image of a blurred image at a current scale from an estimate of a latent image at a previous coarse scale is generated using an upsampling super-resolution function, and a blur kernel is estimated based on the estimate of the latent image and the blurred image; and are repeated from a course to fine scale. A final image estimate is generated. The generating the final image estimate includes performing a deconvolution of the latent image using the blur kernel and the blurred image.

First claim

What is claimed is:

1. A method, comprising:
using one or more processors to perform operations comprising:
repeating from a coarse to fine scale operations comprising:
generating an estimate of a latent image of a blurred image at a current scale from an estimate of a latent image at a previous coarse scale using an upsampling super-resolution function that does not involve interpolation, the upsampling super-resolution function comprising estimating a gradient profile of the latent image at the current scale and using a function of a ratio of the gradient profile of the latent image at the current scale and a gradient profile of the latent image at the previous coarse scale to estimate the latent image at the current scale; and
estimating a blur kernel based on the estimate of the upsampled non-interpolated latent image and the blurred image; and
generating a final image estimate by performing a deconvolution of the latent image using the blur kernel and the blurred image.

2. The method of claim 1 , wherein the generating the estimate of the latent image of the blurred image using the resolution function further comprises:
estimating a gradient of a transformed latent image as a function of the gradient of an initial latent image; and
estimating the latent image at the increased scale using a sum of
a square of an absolute value of a difference of a convolution of a blur kernel with the estimate of the latent image and an initial blurred image, and
a function of a square of an absolute value of a difference of a gradient of the estimate of the latent image and the gradient of the transformed latent image.

3. The method of claim 1 , wherein the performing a deconvolution of the latent image using the blur kernel and the blurred image further comprises:
calculating a final latent image using a sum of a square of an absolute value of a difference of a convolution of the latent image with a blur kernel, and the blurred image and a regularization function.

4. The method of claim 3 , wherein the regularization function comprises a hyperlaplacian prior.

5. The method of claim 1 , wherein the upsampling super-resolution function further comprises a gradient-preserving upsampling function.

6. A system, comprising:
at least one processor; and
a memory comprising program instructions, the program instructions executable by the
at least one processor to perform operations comprising:
repeating from a coarse to fine scale operations comprising:
generating an estimate of a latent image of a blurred image at a current scale from an estimate of a latent image at a previous coarse scale using an upsampling super-resolution function that does not involve interpolation, the upsampling super-resolution function comprising estimating a gradient profile of the latent image at the current scale and using a function of a ratio of the gradient profile of the latent image at the current scale and a gradient profile of the latent image at the previous coarse scale to estimate the latent image at the current scale, and estimating a blur kernel based on the estimate of the upsampled non-interpolated latent image and the blurred image; and

generating a final image estimate by performing a deconvolution of the latent image using the blur kernel and the blurred image.

7. The system of claim 6 , wherein generating the estimate of the latent image comprises:
estimating gradient of a transformed latent image as a function of the gradient of an initial latent image; and
estimating the latent image at the increased scale using a sum of
a square of an absolute value of a difference of a convolution of a blur kernel with the estimate of the latent image and an initial blurred image, and
a function of a square of an absolute value of a difference of a gradient of the estimate of the latent image and the gradient of the transformed latent image.

8. The system of claim 6 , wherein the deconvolution of the latent image comprises:
calculating a final latent image using a sum of a square of an absolute value of a difference of a convolution of the latent image with a blur kernel, and the blurred image and a regularization function.

9. The system of claim 8 , wherein the regularization function comprises a hyperlaplacian prior.

10. The system of claim 6 , wherein the upsampling super-resolution function further comprises a gradient-preserving upsampling function.

11. A non-transitory computer-readable storage medium storing program instructions, the program instructions computer-executable to implement operations comprising:
repeating from a coarse to fine scale, operations comprising:
generating an estimate of a latent image of a blurred image at a current scale from an estimate of a latent image at a previous coarse scale using an upsampling super-resolution function that does not involve interpolation, the super-resolution function comprising:
estimating a gradient of a transformed latent image as a function of the gradient of an initial latent image; and
estimating the latent image at the increased scale using a sum of:
a square of an absolute value of a difference of a convolution of a blur kernel with the estimate of the latent image and an initial blurred image, and
a function of a square of an absolute value of a difference of a gradient of the estimate of the latent image and the gradient of the transformed latent image; and

estimating a blur kernel based on the estimate of the upsampled latent image and the blurred image; and

generating a final image estimate by performing a deconvolution of the latent image using the blur kernel and the blurred image.

12. The non-transitory computer-readable storage medium of claim 11 , wherein the deconvolution of the latent image comprises:
calculating a final latent image using a sum of a square of an absolute value of a difference of a convolution of the latent image with a blur kernel, and the blurred image and a regularization function.

13. The non-transitory computer-readable storage medium of claim 12 , wherein the regularization function implements a hyperlaplacian prior.

14. The non-transitory computer-readable storage medium of claim 11 , wherein the super-resolution function further comprises a gradient-preserving upsampling function.

15. The non-transitory computer-readable storage medium of claim 11 , wherein the blur kernel comprises a point spread function.

16. The non-transitory computer-readable storage medium of claim 11 , wherein the estimating a gradient of a transformed latent image further comprises fitting a gradient profile using a generalized Gaussian distribution.

17. The method of claim 1 , wherein the blurred image contains text and estimating the gradient profile of the latent image at the current scale involves a two-tone latent image in which the intensities of text pixels and background pixels remain consistent.

18. The method of claim 17 , wherein, prior to estimating the gradient profile of the latent image at the current scale, performing an image clustering operations based on pixel intensities.

19. The system of claim 6 , wherein the blurred image contains text and estimating the gradient profile of the latent image at the current scale involves a two-tone latent image in which the intensities of text pixels and background pixels remain consistent.

20. The system of claim 19 , wherein, prior to estimating the gradient profile of the latent image at the current scale, performing an image clustering operations based on pixel intensit

Assignees

• Wang Jue
• Cho Hojin
• Chien Jen-Chan
• Kong Sarah A
• Lee Seungyong
• Adobe Systems Inc

Inventors

• Wang Jue
• Cho Hojin
• Chien Jen-Chan
• Kong Sarah A
• Lee Seungyong

Classifications

• G06T5/003 Primary

  Physics · mapped topic

• G06T5/73 Primary

  Deblurring; Sharpening · CPC title