# US 2016/0253787 A1 · Methods and systems for denoising images

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2016/0253787 A1
- Document ID: US20160253787A1
- Status: Published application
- Kind: A1
- Application number: 14/759,061
- Filed: 2014-03-25
- Published: 2016-09-01
- Earliest claimed: 2014-03-25
- Assignees: Spreadtrum Communications (Shanghai) Co., Ltd.
- Applicants: Spreadtum Communications (Shanghai) Co., Ltd.
- Inventors: Minjie Chen; Le Zhang; Haijun Niu; Fuhuei Lin
- Web page: https://patentgrep.com/patent/US20160253787A1

## CPC

- G06T 5/70: Denoising; Smoothing
- G06T 5/10: using non-spatial domain filtering
- G06T 5/50: using two or more images, e.g. averaging or subtraction
- G06T 2207/10024: Color image
- G06T 2207/20052: Discrete cosine transform [DCT]
- G06T 2207/20064: Wavelet transform [DWT]
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL

## Abstract

Shrinkage function model associated training schemes are used to facilitate obtaining denoised patches, and more particularly, to removing noise from images.

## Claims

### Claim 1

A method for removing noise from an image, the method comprising:
selecting one or more patches of a first image having a noise profile as a set of noisy patches, each noisy patch having a set of n by n pixels;
selecting one or more patches of a second image without the noise profile as a set of clean patches, each clean patch having a set of n by n pixels;
estimating one or more patch types for the set of noisy patches;
constructing a shrinkage function model based on the one or more patch types and the set of n by n pixels of the noisy patch and the clean patch respectively;
applying the shrinkage function model to a third image to obtain a set of n by n denoised patches; and
reconstructing the set of n by n denoised patches to remove noise from the third image.

### Claim 2 (depends on 1)

The method for removing noise from an image of claim 1, further comprising selecting one or more patches with the same color configuration.

### Claim 3 (depends on 1)

The method for removing noise from an image of claim 1, further comprising selecting one or more patches with different color configurations respectively.

### Claim 4 (depends on 1)

The method for removing noise from an image of claim 1, wherein estimating one or more patch types further comprises estimating spatial distribution of the one or more patches of the first image.

### Claim 5 (depends on 1)

The method for removing noise from an image of claim 1, wherein estimating one or more patch types further comprises calculating patch distances between one or more adjacent patches of the first image.

### Claim 6 (depends on 1)

The method for removing noise from an image of claim 1, wherein the shrinkage function model further comprising n by n shrinkage coefficients for each patch type respectively.

### Claim 7 (depends on 1)

The method for removing noise from an image of claim 1, wherein estimating one or more patch types further comprises classifying patch types and grouping similar patch types based on the calculated patch distances.

### Claim 8 (depends on 7)

The method for removing noise from an image of claim 7, wherein classifying patch types further comprises a fast block matching algorithm.

### Claim 9 (depends on 1)

The method for removing noise from an image of claim 1, wherein estimating the patch types further comprises rotating the first patch and the second patch in order to have a same color configuration of the first patch and the second patch.

### Claim 10 (depends on 1)

The method for removing noise from an image of claim 1, wherein estimating patch types further comprises an n-dimensional vector represents number of matched patches in each direction of the fast block matching algorithm.

### Claim 11

A method for removing noise from an image, the method comprising:
selecting a patch of an image, the patch having a first set of n by n pixels, each pixel having a red green blue (RGB) value;
determining a mean RGB value for the patch;
estimating a patch type for the patch;
subtracting each of the RGB values of the n by n pixels of the patch from the mean RGB value to obtain a second set of n by n pixels;
performing a 2D transformation on the second set of n by n pixels to obtain n by n frequency domain coefficients;
performing a 3D transformation on the n by n frequency domain coefficients for s similar patches to obtain n by n by s 3D coefficients;
obtaining a set of shrinkage functions associated with a noise profile of an image sensor and the patch types of the patch;
performing a shrinkage process using the set of shrinkage functions for the n by n by s set of 3D coefficients to obtain an n by n by s set of denoised frequency domain values;
performing one or more inverse transformations on the n by n by s denoised frequency domain values to obtain an n by n set of denoised pixel values.

### Claim 12 (depends on 11)

The method for removing noise from an image of claim 11, wherein the n by n by s set of 3D coefficients further comprising a unique shrinkage curve for each of the n by n by s set of 3D coefficients.

### Claim 13 (depends on 11)

The method for removing noise from an image of claim 11, wherein obtaining the set of shrinkage functions further comprises constructing a shrinkage function model.

### Claim 14 (depends on 13)

The method for removing noise from an image of claim 13, wherein the shrinkage function model further comprising n by n shrinkage coefficients for each patch type respectively.

### Claim 15 (depends on 11)

The method for removing noise from an image of claim 11, further comprising selecting one or more patches with the same color configuration.

### Claim 16 (depends on 11)

The method for removing noise from an image of claim 11, further comprising selecting one or more patches with different color configurations respectively.

### Claim 17 (depends on 11)

The method for removing noise from an image of claim 11, the 3D transformation further comprises a 2D discrete cosine transform (2D-DCT).

### Claim 18 (depends on 11)

The method for removing noise from an image of claim 11, wherein the 3D transformation comprises a 2D discrete wavelet transform (2D-DWT) transform following with a 1-D Haar/Hadamard transform.

### Claim 19 (depends on 11)

The method for removing noise from an image of claim 11, wherein estimating the patch type further comprises estimating spatial distribution of the one or more patches.

### Claim 20 (depends on 11)

The method for removing noise from an image of claim 11, wherein estimating the patch type further comprises calculating patch distances between one or more adjacent patches.

### Claim 21 (depends on 11)

The method for removing noise from an image of claim 11, wherein estimating a patch type further comprises classifying patch types and grouping similar patch types based on the calculated patch distances.

### Claim 22 (depends on 21)

The method for removing noise from an image of claim 21, wherein classifying patch types further comprises a fast block matching algorithm.

### Claim 23 (depends on 11)

The method for removing noise from an image of claim 11, wherein estimating the patch type further comprises rotating one or more patches in order to have a same color configuration.

### Claim 24 (depends on 11)

The method for removing noise from an image of claim 11, wherein estimating the patch type further comprises an n-dimensional vector represents number of matched patches in each direction of the fast block matching algorithm.

### Claim 25 (depends on 11)

The method for removing noise from an image of claim 11, wherein determining a mean RGB value further comprises a least square training scheme.

### Claim 26 (depends on 11)

The method for removing noise from an image of claim 11, wherein obtaining the set of shrinkage functions further comprises a green-red green-blue (GrGb) correction process.

### Claim 27 (depends on 11)

The method for removing noise from an image of claim 11, wherein obtaining the set of shrinkage functions further comprises a chromatic value correction process.
