Non-local means image denoising with detail preservation using self-similarity driven blending

Patent metadata

Field	Value
Publication number	US-9489720-B2
Application number	US-201414494163-A
Country	US
Kind code	B2
Filing date	Sep 23, 2014
Priority date	Sep 23, 2014
Publication date	Nov 8, 2016
Grant date	Nov 8, 2016

Abstract

System, apparatus, method, and computer readable media for texture enhanced non-local means (NLM) image denoising. In embodiments, detail is preserved in filtered image data through a blending between the noisy input target pixel value and the NLM pixel value that is driven by self-similarity and further informed by an independent measure of local texture. In embodiments, the blending is driven by one or more blending weight or coefficient that is indicative of texture so that the level of detail preserved by the enhanced noise reduction filter scales with the amount of texture. Embodiments herein may thereby denoise regions of an image that lack significant texture (i.e. are smooth) more aggressively than more highly textured regions. In further embodiments, the blending coefficient is further determined based on similarity scores of candidate patches with the number of those scores considered being based on the texture score.

First claim

What is claimed is:

1. A computer-implemented image noise reduction method, comprising:
receiving input pixel values for a target pixel and a plurality of pixels within a spatial neighborhood of the target pixel;
computing a texture score indicative of a level of texture within the neighborhood based on the input pixel values;
computing a non-local mean of the target input pixel based on the input pixel values;
modulating a weighting of the non-local mean relative to the target input pixel value inversely with the level of texture by computing a filtered target pixel value that is a blend of the non-local mean and the target input pixel value based on the texture score; and
storing the filtered target pixel value to an electronic memory.

2. The method of claim 1 , wherein computing the non-local mean of the target input pixel further comprises determining a number of candidate patches within the neighborhood from which the non-local mean is computed based on the texture score to include fewer candidate patches in response to detecting a greater level of texture.

3. The method of claim 1 , further comprising:
computing a patch weight for one or more candidate patch within the neighborhood by performing a comparison between each candidate patch and a target patch of pixels within the neighborhood that contains the target pixel;
computing a blending coefficient based on the patch weights and the texture score; and
blending the non-local mean with the target input pixel value by interpolating between the non-local mean and the target input pixel based on the blending coefficient.

4. The method of claim 3 , further comprising:
selecting a set of the greatest patch weights based on the texture score;
wherein computing the blending coefficient based on the patch weights and the texture score further comprises computing a blending coefficient based on a patch weight statistic indicative of the central tendency of the selected set of patch weights; and
wherein interpolating between the non-local mean and the target input pixel based on the blending coefficient further comprises taking a weighted sum of the non-local mean and the target input pixel, the non-local mean and the target input pixel weighted complementarily by the patch weight statistic.

5. The method of claim 4 , wherein selecting the set of the patch weights based on the texture score further comprises determining an integer number of the candidate patches to represent in the set of weights by at least one of:
evaluating a monotonically decreasing closed form function of the texture score ranging between the number of candidate patches and a non-zero minimum; or
accessing a lookup table (LUT) using the texture score as an index value, the LUT associating different numbers of candidate patches with different texture scores.

6. The method of claim 5 , wherein selecting the set of patch weights based on the texture score further comprises determining the number of candidate patches to represent in the set by evaluating a monotonically decreasing closed form function of the texture score, and of the total number of candidate patches.

7. The method of claim 3 , further comprising:
selecting a set of the largest patch weights based on the texture score; and
wherein computing the blending coefficient based on the patch weights and the texture score further comprises:
computing an initial blending coefficient from a patch weight statistic indicative of the central tendency of the selected set of patch weights; and
computing a final blending coefficient through a mapping of the initial blending coefficient, wherein the mapping is a function of the texture score.

8. The method of claim 7 , wherein computing the final blending coefficient further comprises at least one of:
evaluating a monotonically decreasing closed form function of the texture score; or
accessing a lookup table (LUT) using the texture score as an index value, the LUT associating different texture scores with different mappings between initial and final blending coefficients.

9. The method of claim 1 , wherein computing the texture score further comprises:
computing an estimator of dispersion across all, or a subset, of the pixel values within the neighborhood; and
normalizing the dispersion estimator by signal intensity.

10. The method of claim 1 , wherein computing the non-local mean of the target input pixel further comprises:
computing a patch weight for each candidate patch within the neighborhood by performing a comparison between a number of candidate patches and a target patch of pixels within the neighborhood that contains the target pixel; and
computing an average patch by averaging the candidate patches as weighted by their corresponding patch weight.

11. A computer-implemented image noise reduction method, comprising:
receiving input pixel values for a target pixel and a plurality of pixels within a spatial neighborhood of the target pixel;
computing a texture score indicative of a level of texture within the neighborhood;
computing a patch weight for one or more candidate patch within the neighborhood by performing a comparison between each of the candidate patches and a target patch of pixels that contains the target pixel;
selecting a set of the greatest patch weights representing an integer number of the candidate patches by evaluating a monotonically decreasing closed form function of the texture score, and of the total number of candidate patches;
computing an initial blending coefficient from a patch weight statistic indicative of the central tendency of the selected set of patch weights;
computing a final blending coefficient through a mapping of the initial blending coefficient, wherein the mapping is a function of the texture score;
selecting a subset of candidate patches within the neighborhood based on the texture score to include fewer candidate patches in response to detecting a greater level of texture;
computing a non-local mean of the target input pixel from an average patch determined by averaging the subset of candidate patches as weighted by their corresponding patch weight;
blending the non-local mean with the target input pixel value by modulating a weighting of the non-local mean relative to the target input pixel value inversely with the level of texture by computing a filtered target pixel value that is a blend of the non-local mean and the target input pixel value based on the final blending coefficient; and
storing the blend to memory as the filtered target pixel value.

12. The method of claim 11 , wherein computing the texture score further comprises:
computing an estimator of dispersion across all, or a subset, of the pixel values within the neighborhood; and
normalizing the dispersion estimator by signal intensity.

13. An image processing system, comprising:
an input device to receive input pixel values for a target pixel and a plurality of pixels within a spatial neighborhood of the target pixel;
a denoising module, comprising logic circuitry, coupled to the input device, and further including:
a texture processing module comprising logic circuitry to compute a texture score indicative of a level of texture detected within the neighborhood based on the input pixel values;
a NLM module comprising logic circuitry to compute a non-local mean of the target input pixel based on the input pixel values; and
a blending module comprising logic circuitry to modulate a weighting of the non-local mean relative to the target input pixel value inversely with the level of texture by computing a filtered target pixel value that is a blend of the non-local mean and

Assignees

• Intel Corp

Inventors

• Oron Shaul
• Michael Gilad

Classifications

• G06T5/20

  using local operators · CPC title

• G06T5/002 Primary

  Physics · mapped topic

• G06T7/40

  Analysis of texture (depth or shape recovery from texture G06T7/529) · CPC title

• G06T5/70 Primary

  Denoising; Smoothing · CPC title