CN106127729A - A kind of picture noise level estimation method based on gradient - Google Patents

Abstract

The invention discloses a kind of picture noise level estimation method based on gradient, including: the noisy image of noise level to be estimated is divided into the image block of multiple fixed size by (1)；(2) calculate the comprehensive gradient metric of each pixel in each image block, choose the image block of predetermined quantity according to the interval distribution of comprehensive gradient metric；(3) neutral net is utilized, the variance yields before and after each image block choosing out is carried out denoising and calculates denoising；(4) variance yields of minimum is chosen as final picture noise horizontal estimated value.The present invention, by the way of using statistical picture gradient, selects the image block that corresponding texture is more weak, improves the computational accuracy of final variance；Utilize neutral net, the image block of selection is first carried out a denoising, then calculates the variance of error image block so that algorithm more robust, the scope of application is wider.

Description

A kind of picture noise level estimation method based on gradient

Technical field

The present invention relates to Computer Image Processing field, be specifically related to a kind of picture noise horizontal estimated side based on gradient Method.

Background technology

The fast development of multimedia technology so that we live in the world of a full digital picture and video.So And, these images are usually because the impact of the factor such as electronic component, illumination condition, and the when of meeting imaging, meeting is with noise.Generally We research noise be independently of image itself, the additive noise of Gaussian distributed.Many Denoising Algorithm are by noise Variance as a priori to carry out denoising process, it is apparent that this is inapplicable to produce actual way, so making an uproar from band Image calculates noise level automatically, is the most necessary.

In recent years, the algorithm having many noise levels to estimate is suggested, including using wavelet transformation etc..Based on filtering Method is used to first carry out image denoising, then by the side of the error image of image after calculating noisy image and denoising Difference is using as final noise level estimated value.But this class method has its limitation, because in the error image obtained The most also the marginal information of original image is comprised, so the variance calculated can be affected.

In the recent period, new noise level algorithm for estimating is also had to be suggested.Tian and Chen utilizes graph model and ant group optimization Technology selects the image block for carrying out calculating noise level.Jiang and Zhang proposes a kind of quickly based on statistics vacation If noise Estimation Algorithm.Pyatykh proposes a kind of new noise level algorithm for estimating based on image texture.

Summary of the invention

The invention provides a kind of picture noise level estimation method based on gradient, by using statistical picture gradient Mode, selects the image block that corresponding texture is more weak, improves the computational accuracy of final variance；Utilize neutral net, will select Image block first carry out a denoising, then calculate the variance of error image block so that algorithm more robust, the scope of application is more Extensively.

A kind of picture noise level estimation method based on gradient, including:

(1) noisy image of noise level to be estimated is divided into the image block of multiple fixed size；

(2) calculate the comprehensive gradient metric of each pixel in each image block, divide according to the interval of comprehensive gradient metric The image block of predetermined quantity chosen by cloth；

(3) neutral net is utilized, the variance yields before and after each image block choosing out is carried out denoising and calculates denoising；

(4) variance yields of minimum is chosen as final picture noise horizontal estimated value.

In step (1), the noisy image of noise level to be estimated divides according to pre-fixed step size, and adjacent image block Between with overlapping region；

Described in step (2), comprehensive gradient metric is:

$measure\left(P_i\right(x,y\left)\right)=\sqrt{{\left(\frac{\∂P_i(x,y)}{x}\right)}^2+{\left(\frac{\∂P_i(x,y)}{y}\right)}^2}\×\mathrm{\θ}\left(P_i\right(x,y\left)\right)$

Wherein:

P_i(x y) represents image block P_iIn (x, y) pixel of position,

Represent image block P_iIn (x, y) the horizontal direction gradient of position pixel,

Represent image block P_iIn (x, y) the vertical direction gradient of position pixel.

$\mathrm{\θ}\left(P_i\right(x,y\left)\right)=\arctan (\frac{\∂P_i(x,y)}{y}/\frac{\∂P_i(x,y)}{x})+\frac{\mathrm{\π}}{2}.$

Horizontal direction gradient and the vertical direction gradient of pixel are respectively as follows:

$\frac{\∂P_i(x,y)}{x}=\[P_i(x+1,y)-P_i(x,y)+P_i(x+1,y+1)-P_i(x,y+1)\]/2,$

$\frac{\∂P_i(x,y)}{y}=\[P_i(x,y+1)-P_i(x,y)+P_i(x+1,y+1)-P_i(x+1,y)\]/2.$

For each image block in step (2), the comprehensive gradient metric of each pixel is sorted from small to large；

Wherein maximum is max (measure), and minima is min (measure)；

Interval between max (measure)-min (measure) is divided into nine parts, adds up combining on each subinterval Close gradient metric number, i.e. corresponding number of pixels, be designated as In (m) (m=1,2 ..., 9)；By In (m) (m=1,2 ..., 9) maxima and minima in takes out, the corresponding ratio of calculating:

$In\_ration=\frac{max\left(In\right(m\left)\right)}{min\left(In\right(m\left)\right)}$

This corresponding for all of image block ratio is sorted from big to small, and the descending image choosing predetermined quantity Block.

Compared with prior art, the invention have the benefit that

(1) a kind of based on gradient the picture noise level estimation method that the present invention proposes, by using statistical picture ladder The mode of degree, selects the image block that corresponding texture is more weak, improves the computational accuracy of final variance；

(2) a kind of based on gradient the picture noise level estimation method that the present invention proposes, utilizes neutral net, will select Image block first carry out a denoising, then calculate the variance of error image block；With this so that algorithm more robust, it is suitable for Wider.

Detailed description of the invention

Below in conjunction with specific embodiment, the present invention is described in detail.Present invention picture noise based on gradient level Method of estimation mainly comprises the following steps:

(1) noisy image of given noise level to be estimated, (in the present embodiment, I size is 256 × 256 pictures to be designated as I Element), block (when being embodied as, N=28) that I is divided into size be N × N (on image I according to interval (step-length) 2 pixels, from From left to right, top-down rule scan, and obtain the image block that multiple size is N × N).

(2) each image block P that step (1) is obtained_i(i=1,2 ..., n), wherein n is image block number, carries out Following operation:

(2-1) according to equation below, image block P is calculated_iIn the horizontal direction of each pixel and the gradient of vertical direction:

$\frac{\∂P_i(x,y)}{x}=\[P_i(x+1,y)-P_i(x,y)+P_i(x+1,y+1)-P_i(x,y+1)\]/2$

$\frac{\∂P_i(x,y)}{y}=\[P_i(x,y+1)-P_i(x,y)+P_i(x+1,y+1)-P_i(x+1,y)\]/2$

Wherein, P_i(x y) represents image block P_iIn (x, y) pixel of position,Represent image block P_iIn (x, y) The horizontal direction gradient of position pixel,Represent image block P_iIn (x, y) the vertical direction gradient of position pixel.

(2-2) gradient of each pixel obtained according to step (2-1), calculates each pixel according to equation below and repaiies Positive gradient direction:

$\mathrm{\θ}\left(P_i\right(x,y\left)\right)=\arctan (\frac{\∂P_i(x,y)}{y}/\frac{\∂P_i(x,y)}{x})+\frac{\mathrm{\π}}{2}$

Wherein, P_i(x y) represents image block P_iIn (x, y) pixel of position,The most above-mentioned ask The gradient obtained, arctan represents arc tangent trigonometric function.

(2-3) according to equation below, image block P_iIn the comprehensive gradient metric of each pixel be expressed as:

$measure\left(P_i\right(x,y\left)\right)=\sqrt{{\left(\frac{\∂P_i(x,y)}{x}\right)}^2+{\left(\frac{\∂P_i(x,y)}{y}\right)}^2}\×\mathrm{\θ}\left(P_i\right(x,y\left)\right)$

Wherein, P_i(x y) represents image block P_iIn (x, y) pixel of position,The most above-mentioned ask The gradient obtained, θ (P_i(x, y)) is calculated gradient direction in step (2-2).

(2-4) by comprehensive gradient metric obtained above according to sorting from small to large, it is assumed that maximum is max (measure), minima min (measure), the interval between max (measure)-min (measure) is divided into nine parts, Add up the comprehensive gradient metric number on each subinterval, i.e. corresponding number of pixels, be designated as In (m) (m=1,2 ..., 9)。

(2-5) by (2-4) calculates In (m) (m=1,2 ..., 9) in maxima and minima take out, it is right to calculate Answer ratio:

$In\_ration=\frac{max\left(In\right(m\left)\right)}{min\left(In\right(m\left)\right)}$

Wherein, calculate during max (In (m)) is step (2-4) comprises the interval that comprehensive gradient metric number is most Contained pixel count, corresponding min (In (m)) is in step (2-4) interval comprising gradient metric minimum number calculated Contained pixel count；

(3) each image block is through the calculating of step (2), the most all can obtain ratio I n_ration, will be all This ratio corresponding to image block sort from big to small, take its front K numerical value (in being embodied as, K=25), and it be corresponding to record it Image block.

(4) for K the image block determined in step (3), list of references " Agostinelli F, Anderson M R, Lee H.Adaptive Multi-Column Deep Neural Networks with Application to Robust Image Denoising (2013) " the middle AMC-SSDA technology proposed, utilize neutral net, to each image block P_kGo Make an uproar:

(4-1) size of image block here is 28 × 28 in a particular embodiment, and the AMC-SSDA god mentioned in literary composition Also it is 28 × 28 through the input dimension of network, therefore can directly utilize the neutral net trained and carry out relevant treatment；Here AMC-SSDA be the linear combination of a series of SSDA (stacking coefficient denoising automatic coding machine)；

(4-2) for image block P_k, as the input of the AMC-SSDA neutral net trained, first pass around C (specifically real Shi Shi, C=20) individual SSDA, each of which SSDA is trained for the specific noise type of one, and output result is C Image block after corresponding denoisingThe size of each image block and P_kUnanimously；

(4-3) image block of above-mentioned generationIt is multiplied by the weight coefficient s of correspondence₁,s₂,...,s_CIt is added, this Weight coefficient is the same a bit is automatically generated by the neutral net trained, and final weighted results is corresponding noise-reduced image piece

Obtain the noise-reduced image of correspondenceThe advantage of the method is to obtain on the premise of unknown noise level value To good denoising effect, this is also the feature of neutral net.

(5) P is calculated according to equation below_kWithDifference variance:

$\mathrm{var}iance(P_k-P_k^{*})=\frac{1}{S^2}\underset{i=1}{\overset{S}{\Σ}}\underset{j=1}{\overset{S}{\Σ}}\left(\right(P_k-P_k^{*}\left)\right(i,j)-\mathrm{\μ})$

Wherein, S is image block P_kNumber of pixels (being 256 in being embodied as), μ isAverage, be defined as:

$\mathrm{\μ}=\frac{1}{S^2}\underset{i=1}{\overset{S}{\Σ}}\underset{j=1}{\overset{S}{\Σ}}(P_k-P_k^{*})(i,j)$

(6) K variance yields calculated in step (5) is sorted from small to large, take minimum that as final Noise level estimated value

Claims (5)

1. a picture noise level estimation method based on gradient, it is characterised in that including:

(1) noisy image of noise level to be estimated is divided into the image block of multiple fixed size；

(2) the comprehensive gradient metric of each pixel in each image block is calculated, according to the interval distribution choosing of comprehensive gradient metric Take the image block of predetermined quantity；

(3) neutral net is utilized, the variance yields before and after each image block choosing out is carried out denoising and calculates denoising；

(4) variance yields of minimum is chosen as final picture noise horizontal estimated value.

2. picture noise level estimation method based on gradient as claimed in claim 1, it is characterised in that step is treated in (1) Estimate that the noisy image of noise level divides according to pre-fixed step size, and with overlapping region between adjacent image block.

3. picture noise level estimation method based on gradient as claimed in claim 1, it is characterised in that institute in step (2) Stating comprehensive gradient metric is:

Wherein:

P_i(x y) represents image block P_iIn (x, y) pixel of position,

Represent image block P_iIn (x, y) the horizontal direction gradient of position pixel,

Represent image block P_iIn (x, y) the vertical direction gradient of position pixel.

4. picture noise level estimation method based on gradient as claimed in claim 3, it is characterised in that the level side of pixel It is respectively as follows: to gradient and vertical direction gradient

5. picture noise level estimation method based on gradient as claimed in claim 1, it is characterised in that pin in step (2) To each image block, the comprehensive gradient metric of each pixel is sorted from small to large；

Wherein maximum is max (measure), and minima is min (measure)；

Interval between max (measure)-min (measure) is divided into nine parts, adds up the comprehensive ladder on each subinterval Degree metric number, i.e. corresponding number of pixels, be designated as In (m) (m=1,2 ..., 9)；By In (m) (m=1,2 ..., 9) in Maxima and minima take out, calculate corresponding ratio:

This corresponding for all of image block ratio is sorted from big to small, and the descending image block choosing predetermined quantity.