Recursive conditional means image denoising

• US 8,908,989 B2
• Filed: 08/12/2013
• Issued: 12/09/2014
• Est. Priority Date: 11/02/2012
• Status: Expired due to Fees

• Current Assignee: Board of Regents of the University of Texas System (University of Texas System)
• Original Assignee: Board of Regents of the University of Texas System (University of Texas System)
• Inventors: Geisler, Wilson S., Perry, Jeffrey S.
• Primary Examiner(s): Mehta, Bhavesh
• Assistant Examiner(s): Dunphy, David F
• Application Number:
  US13/965,073
• Publication Number:
  US 20140126808A1
• Time in Patent Office: 484 Days
• Field of Search: None
• US Class Current: 382/265
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/295 Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
• CPC Class Codes: G06T 2207/20076 Probabilistic image processing
• CPC Class Codes: G06T 5/70 Denoising; Smoothing
• CPC Class Codes: G06V 10/457 by analysing connectivity, e.g. edge linking, connected component analysis or slices
• CPC Class Codes: G06V 10/85 Markov-related models; Markov random fields

Abstract

Methods and composition for denoising digital camera images are provided herein. The method is based on directly measuring the local statistical structure of natural images in a large training set that has been corrupted with noise mimicking digital camera noise. The measured statistics are conditional means of the ground truth pixel value given a local context of input pixels. Each conditional mean is the Bayes optimal (minimum mean squared error) estimate given the specific local context. The conditional means are measured and applied recursively (e.g., the second conditional mean is measured after denoising with the first conditional mean). Each local context vector consists of only three variables, and hence the conditional means can be measured directly without prior assumptions about the underlying probability distributions, and they can be stored in fixed lookup tables.

Claims

1. A method of estimating an original value of a pixel in a digital image, said method comprising:

(i) identifying noise in a high quality digital image and producing a digital image noise model;

(ii) adding model noise based on said digital image noise model to a plurality of ground-truth digital images and producing a plurality of training images;

(iii) recursively measuring and applying conditional means to a plurality of different local pixel contexts based on said training images and producing a look-up table comprising estimates of point (pixel) values;

(iv) terminating the recursive measuring and applying conditional means to a plurality of different local pixel contexts based on said training images when the convergence of the conditional means indicates an asymptote; and

(v) assigning a value to a pixel in a digital image using said estimates of point (pixel) values in said look-up table.

2. The method of claim 1, wherein said estimates of point (pixel) values are the optimal Bayesian estimates of point (pixel) values.

3. The method of claim 1, wherein said conditional means for a plurality of different local pixel contexts are derived from a number of different small pixel neighborhoods comprising context regions.

4. The method of claim 1, wherein said look-up table is a stored and fixed table.

5. The method of claim 1, wherein the context vectors for measuring and applying said conditional means consist of three 8-bit variables.

6. The method of claim 1, wherein the conditional means for each different local pixel context provides the Bayes optimal minimum mean squared error (MMSE) estimate.

7. The method of claim 1, wherein said digital image is derived from a digital video.

8. The method of claim 1, wherein said the original value of a plurality of pixels in a digital image are identified thereby denoising said digital image.

9. A system for estimating an original value of a pixel in a digital image, said system comprising:

a processor configured to;

identify noise in a high quality digital image and produce a digital image noise model;

add model noise based on said digital image noise model to a plurality of ground-truth digital images and produce a plurality of training images;

recursively measure and apply conditional means for a plurality of different local pixel contexts based on said training images and produce a look-up table comprising estimates of point (pixel) values;

terminate the recursive measuring and applying conditional means to a plurality of different local pixel contexts based on said training images when the convergence of the conditional means indicates an asymptote; and

assign a value to a pixel in a digital image using said estimates of point (pixel) values in said look-up table; and

memory configured to store said look-up table.

10. The system of claim 9, wherein said estimates of point (pixel) values are the optimal Bayesian estimates of point (pixel) values.

11. The system of claim 9, wherein said conditional means for a plurality of different local pixel contexts are derived from a number of different small pixel neighborhoods comprising context regions.

12. The system of claim 9, wherein said look-up table is a fixed table.

13. The system of claim 9, wherein the context vectors for measuring and applying said conditional means consist of three 8-bit variables.

14. The system of claim 9, wherein the conditional means for each different local pixel context provides the Bayes optimal minimum mean squared error (MMSE) estimate.

15. The system of claim 9, wherein said digital image is derived from a digital video.

16. The system of claim 9, wherein said the original value of a plurality of pixels in a digital image are identified thereby denoising said digital image.

17. A method of generating a denoised image comprising:

(a) receiving at least one input image at a processor;

(b) identifying a plurality of denoising contexts within the at least one input image, wherein the denoising contexts comprise a location and selected areas proximate to the location;

(c) identifying a Bayes optimal estimate, wherein identifying the Bayes optimal estimate comprises measuring local statistics within the plurality of denoising contexts, wherein the local statistics comprise at least one of a conditional mean and a conditional variance based on the measured local statistics;

(d) generating at least one next image based on the Bayes optimal estimate;

(e) identifying a plurality of next image denoising contexts within the at least one next image, wherein the denoising contexts comprise a location and selected areas proximate to the location;

(f) measuring next image local statistics within the plurality of next image denoising contexts, wherein the next image local statistics comprise at least one of a conditional mean and a conditional variance; and

(g) determining whether the next image local statistics and the local statistics indicate an asymptote.

18. The method of claim 17, wherein, if the next image local statistics and the local statistics do not indicate an asymptote, steps (a) through (g) are repeated, and wherein the input image of step (a) is the at least one next image.

19. The method of claim 17, wherein measuring the local statistics within the plurality of denoising contexts comprises directly measuring the local statistics from the denoising contexts.

20. The method of claim 19, wherein directly measuring the local statistics from the denoising contexts comprises generating the local statistics without use of assumptions about underlying probability distributions.