Tree-based linear regression for denoising

Patent metadata

Field	Value
Publication number	US-9342870-B2
Application number	US-201314060076-A
Country	US
Kind code	B2
Filing date	Oct 22, 2013
Priority date	Oct 22, 2013
Publication date	May 17, 2016
Grant date	May 17, 2016

Abstract

Image denoising techniques are described. In one or more implementations, a denoising result is computed by a computing device for a patch of an image. One or more partitions are located by the computing device that correspond to the denoising result and a denoising operator is obtained by the computing device that corresponds to the located one or more partitions. The obtained denoising operator is applied by the computing device to the image.

First claim

What is claimed is:

1. A method comprising:
locating a noisy patch by a computing device within an image to be denoised;
computing an initial denoising result by the computing device for the noisy patch;
locating one or more of a plurality of partitions by the computing device that correspond to the initial denoising result for the noisy patch, each of the plurality of partitions formed as clusters based on clean patches taken from training samples of training images that do not have noise that is to be removed;
obtaining a denoising operator by the computing device associated with the located one or more partitions that correspond to the initial denoising result for the noisy patch and that is learned from the training samples of the training images; and
applying the obtained denoising operator by the computing device to the noisy patch to compute a denoising result for the noisy patch.

2. A method as described in claim 1 , wherein the denoising operator is locally linear.

3. A method as described in claim 1 , wherein the partitions are formed using a hierarchical k-means clustering technique, k-means clustering technique, agglomerative clustering, spectral clustering, or affinity-based clustering.

4. A method as described in claim 1 , wherein the denoising operator of respective said partitions is formed using a piece-wise or locally linear regression technique.

5. A method as described in claim 1 , wherein the partitions are arranged in a tree-like structure.

6. A method as described in claim 1 , wherein the initial denoising result for the noisy patch is computed using a self-similarity technique.

7. A method as described in claim 6 , wherein the self-similarity is based at least in part on patch recurrence in the image.

8. A system comprising:
one or more modules implemented at least partially in hardware, the one or more modules configured to perform operations comprising:
forming a plurality of partitions of training samples of training images according to clean patches taken from the training images that do not have noise that is to be removed;
calculating a denoising operator for each said partition; and
providing the calculated denoising operator for use in performing a denoising operation for a noisy patch of an image to be denoised in response to identification of at least one said partition that corresponds to an initial denoising result for the noisy patch.

9. A system as described in claim 8 , wherein the partitions are formed using a piece-wise or locally linear regression technique.

10. A system as described in claim 8 , wherein the denoising operator is locally linear.

11. A system as described in claim 8 , wherein the partitions are arranged in a tree-like structure.

12. A system as described in claim 8 , wherein the initial denoising result is based at least in part on patch recurrence.

13. A method comprising:
forming a plurality of partitions of training samples of training images by one or more computing devices according to clean patches taken from the training images that do not have noise that is to be removed;
calculating a denoising operator by the one or more computing devices for each said partition; and
providing the calculated denoising operator by the one or more computing devices for use in performing a denoising operation for a noisy patch of an image in response to identification of at least one said partition that corresponds to an initial denoising result for the noisy patch.

14. A method as described in claim 13 , wherein the partitions are formed using a piece-wise or locally linear regression technique.

15. A method as described in claim 13 , wherein the denoising operator is locally linear.

16. A method as described in claim 13 , wherein the partitions are arranged in a tree-like structure.

17. A method as described in claim 13 , wherein the initial denoising result is based at least in part on patch recurrence.

18. A method as described in claim 1 , wherein mean values of the initial denoising result for the noisy patch are added to the denoising result for the noisy patch after the denoising result is computed for the noisy patch.

19. A method as described in claim 1 , wherein the initial denoising result for the noisy patch is computed using patches from within the image.

Assignees

• Adobe Systems Inc

Inventors

• Lin Zhe
• Jin Hailin

Classifications

• G06T5/002 Primary

  Physics · mapped topic

• G06T2207/20021

  Dividing image into blocks, subimages or windows · CPC title

• G06T5/70 Primary

  Denoising; Smoothing · CPC title