CN101719267B - A kind of denoising noise image and system - Google Patents

Abstract

The invention discloses a kind of denoising noise image, including step: the noisy image inputted is carried out mistake and cuts operation, obtain mistake and cut image; Utilize the contourlet transformation based on small echo that mistake is cut image and carry out Its Sparse Decomposition, obtain low frequency subgraph picture and high frequency subimage, each high frequency subimage is carried out noise reduction process, obtain noise reduction high frequency subimage; Low frequency subgraph picture and noise reduction high frequency subimage are carried out the Contourlet transformation by reciprocal direction based on small echo, then carries out identical mistake and cut the reverse mistake of amplitude and cut operation. The invention also discloses a kind of noisy image noise reduction process system. The present invention is effectively reduced the noise signal in image, is effectively retained the detailed information of image, is greatly promoted image noise reduction effect, improves picture quality.

Description

A kind of denoising noise image and system

Technical field

The present invention relates to technical field of image processing, be specifically related to a kind of denoising noise image and system.

Background technology

Image is gathering, is changing and in transmitting procedure, it is vulnerable to the impacts such as imaging device and external environmental noise interference and Quality Down, therefore image noise reduction is basis and necessary pre-treatment step in Image Engineering, is one of the key technology of image perception, classification and identification.

The essence of image noise reduction is pattern classification: namely irregular " noise pattern " that have point-like unusual separated from regular " image model ". The object function unusual containing point-like is optimum base by small echo, but in higher-dimension situation, wavelet analysis can not make full use of the specific geometric properties of data itself, is not therefore the function representation method of optimum " the most sparse " in other words. Contourlet transformation solves wavelet transformation can not effectively represent two-dimentional or more higher-dimension singularity shortcoming, exactly the edge in image can be captured in the subband of different scale, different frequency, different directions, it is thus possible to effectively process the line singularity in image.

At present, image denoising method is broadly divided into airspace filter, transform domain filtering and transform domain statistical modeling and analyzes three major types. Traditional most of filtering method belongs to the former, such as mean filter, medium filtering etc., actually adopts various smooth function that image is carried out process of convolution, it is simple to hardware realizes, and while attenuating noise, image useful information has been also carried out smooth. Transform domain statistical modeling is analyzed method and coefficient in transform domain is carried out statistical modeling, obtains good noise reduction, however it is necessary that more prior information, sets up the model being suitable for and is trained, and computation complexity is significantly high. And in transform domain filtering method, most representative with the collapse threshold noise-reduction method based on wavelet transformation that Donoho and Johnstone proposes. Owing to signal is after wavelet transformation, signal is concentrated mainly on the wavelet coefficient that minority absolute amplitude is bigger, and noise is then dispersed on the wavelet coefficient that some absolute amplitude are less, therefore, can utilize collapse threshold that wavelet coefficient is carried out noise reduction, reach the purpose of noise reduction. But, wavelet shrinkage threshold value easily produces distortion, is referred to as Pseudo-Gibbs artifacts.

Chinese Patent Application No. is CN200610030756, denomination of invention is " image denoising method in a kind of contourlet transformation territory " and number of patent application be CN200610030745, denomination of invention is the method that " based on the morphologic image denoising method of transforming domain mathematics " discloses 2 kinds of image noise reductions, both approaches is inherently utilize translation operator to improve Contourlet noise reduction, it provides a translation invariant tight frame. But, only simple translation, not only can not make full use of the directivity of contourlet transformation, and destroy the texture continuity of image, be restricted the improvement of noise reduction is also natural.

Summary of the invention

The main technical problem to be solved in the present invention is to provide a kind of denoising noise image and system, is effectively reduced the noise signal in image, improves picture quality.

For solving above-mentioned technical problem, the present invention provides a kind of denoising noise image, including step:

The noisy image inputted is carried out mistake and cuts operation, obtain mistake and cut image;

Utilize the contourlet transformation based on small echo that mistake is cut image and carry out Its Sparse Decomposition, obtain low frequency subgraph picture and high frequency subimage, each high frequency subimage is carried out noise reduction process, obtain noise reduction high frequency subimage;

Low frequency subgraph picture and noise reduction high frequency subimage are carried out the Contourlet transformation by reciprocal direction based on small echo, then carries out identical mistake and cut the reverse mistake of amplitude and cut operation.

In embodiments of the present invention, the noisy image of input is carried out mistake to cut operation and specifically include step:

Mistake is set and cuts amplitude and mistake cuts amplitude delta, the noisy image of input is carried out mistake multi-direction, certain and cuts the repeatedly mistake of amplitude and cut operation.

In embodiments of the present invention, described method still further comprises step:

Reverse mistake is cut all images of obtaining of operation and carries out linear averaging, obtain final noise-reduced image.

In embodiments of the present invention, utilize the contourlet transformation based on small echo that mistake is cut image and carry out Its Sparse Decomposition, obtain low frequency subgraph picture and high frequency subimage specifically includes step:

Arrange based on the wavelet decomposition number of plies in the contourlet transformation of small echo and the Directional Decomposition number in every layer;

Utilize the contourlet transformation based on small echo that mistake is cut image and carry out Its Sparse Decomposition multiple dimensioned, multidirectional, obtain low frequency subgraph picture and a series of high frequency subimage with different resolution.

In embodiments of the present invention, each high frequency subimage is carried out noise reduction process to specifically include:

Threshold parameter is set, utilizes soft-threshold function, each high frequency subimage is carried out threshold deniosing process.

In embodiments of the present invention, described threshold function table is set before still further comprise step:

According to the wavelet decomposition number of plies, Directional Decomposition number in every layer and high frequency band coefficient, adopt the mediant estimation of robustness, obtain the noise criteria on each direction of high frequency subimage poor;

Poor according to the noise criteria on each direction, calculate and obtain threshold parameter.

In embodiments of the present invention, described soft-threshold function is soft-threshold function Λ (*)=sgn (*) max (*, T_B), T_BFor threshold parameter.

A kind of noisy image noise reduction process system, including:

Wrong cut unit, carries out mistake to the noisy image inputted and cuts operation, obtain mistake and cut image; And to cutting the reverse mistake of amplitude and cut operation through carrying out identical mistake again based on the image of Contourlet transformation by reciprocal direction of small echo;

Based on the contourlet transformation unit of small echo, mistake is cut image and carries out Its Sparse Decomposition, obtain low frequency subgraph picture and high frequency subimage; And carry out the Contourlet transformation by reciprocal direction based on small echo to low frequency subgraph picture with through the noise reduction high frequency subimage of noise reduction process;

Noise reduction processing unit, carries out noise reduction process to each high frequency subimage, obtains noise reduction high frequency subimage.

In embodiments of the present invention, described system may further comprise:

Linear averaging unit, for cutting all images that operation obtains and carry out linear averaging through carrying out reverse mistake based on the contourlet transformation unit of small echo, obtaining final noise-reduced image.

In embodiments of the present invention, described noise reduction processing unit is threshold deniosing processing unit, for each high frequency subimage is carried out threshold deniosing process, obtains noise reduction high frequency subimage.

Compared with prior art, the present invention is more careful and thorough to the decomposition of noisy image, is effectively reduced the noise signal in image, the detailed information such as texture edge remaining image, improves the quality of image.

Accompanying drawing explanation

A kind of denoising noise image flow chart that Fig. 1 provides for the specific embodiment of the invention;

A kind of noisy image noise reduction process system block diagram that Fig. 2 provides for the specific embodiment of the invention;

Fig. 3 is the image schematic diagram of the different-effect adopting different disposal method to obtain.

Detailed description of the invention

The present invention is described in further detail in conjunction with accompanying drawing below by detailed description of the invention.

It is desirable to provide the image denoising method of a kind of improvement and system, first the mistake noisy image of input carrying out certain amplitude is cut, utilize the contourlet transformation based on small echo that noisy image carries out Its Sparse Decomposition multiple dimensioned, multidirectional, and utilize minimum Bayesian risk function to estimate that the noise threshold of the sub-high frequency subimage of high frequency is thus carrying out threshold denoising at transform domain, next carries out the Contourlet inverse transformation based on small echo and corresponding mistake is cut the inverse mistake of amplitude and cut operation, obtains the noise-reduced image after this mistake is cut. Then repeat step above, and the noise-reduced image every time obtained is carried out linear averaging, obtain final noise-reduced image, reach the purpose of image noise reduction.

In the embodiment of the present invention, mistake is cut the transformation matrix of operation and can be expressed as:

$A_{x,a}=\left(\begin{array}{cc}1& 0\\ a& 1\end{array}\right),$ $A_{y,a}=\left(\begin{array}{cc}1& a\\ 0& 1\end{array}\right)$

Wherein, A_{X, a}And A_{Y, a}Represent the mistake along x-axis and y-axis respectively and cut operation. Parameter a is called that mistake cuts amplitude, so that mistake cuts operation does not introduce too much irregular frequency, and | a |≤1 in the embodiment of the present invention.

Assume that the noise image observed is

I=f+n

Wherein f is original image, and n is independent identically distributed white Gaussian noise signal N (0, σ²)。

Refer to shown in Fig. 1, a kind of noisy image processing method that Fig. 1 provides for the embodiment of the present invention, including step:

Step 10: Initialize installation, makes mistake cut amplitude a=-1, and mistake cuts amplitude delta Δ a=0.05; Concurrently set based on the wavelet decomposition number of plies K in the contourlet transformation of small echo and the Directional Decomposition number L in every layer_k;

Step 11: the noisy image I of input is carried out in x-axis and y-axis direction effective mistake and cuts the mistake of amplitude and cut operation, obtain mistake and cut image S_{X, a}=C_{X, a}And S (I)_{Y, a}=C_{Y, a}(I);

Step 12: the mistake obtained is cut image S_{X, a}And S_{Y, a}Carry out the Contourlet Its Sparse Decomposition based on small echo respectively; Namely

$[S_{\mathrm{lx}},S_{\mathrm{hx}}^{\left(\mathrm{1,1}\right)},...,S_{\mathrm{hx}}^{(1,L_1)},S_{\mathrm{hx}}^{\left(\mathrm{2,1}\right)},...,S_{\mathrm{hx}}^{(k,L_k)}]=T\left(S_{\mathrm{xa}}\right)$

With $[S_{\mathrm{ly}},S_{\mathrm{hy}}^{\left(\mathrm{1,1}\right)},...,S_{\mathrm{hy}}^{(1,L_1)},S_{\mathrm{hy}}^{\left(\mathrm{2,1}\right)},...,S_{\mathrm{hy}}^{(k,L_k)}]=T\left(S_{\mathrm{ya}}\right)$

Wherein, T (*) is the contourlet transformation based on small echo; Thus obtaining low frequency subgraph as S_lxAnd S_ly, and a series of high frequency subimage with different resolution, wherein k ∈ (1, K) and l ∈ (1, L_k) indicate that subimage is positioned at the l Directional Decomposition sub-band of kth layer wavelet decomposition;

Step 13: to each the high frequency subimage in step 12WithCarry out threshold deniosing process and obtain noise reduction high frequency subimage, obtain noise reduction high frequency subimageWith

$S_{\mathrm{Dhx}}^{(k,L_k)}=\mathrm{\Λ}(S_{\mathrm{hx}}^{(k,L_k)},T_B)$ With $S_{\mathrm{Dhy}}^{(k,L_k)}=\mathrm{\Λ}(S_{\mathrm{hy}}^{(k,L_k)},T_B);$

Wherein, Λ (*) is threshold function table, and the present invention selects soft-threshold function Λ (*)=sgn (*) max (*, T_B), T_BFor threshold parameter. Choosing of threshold parameter is most important, and owing to the Contourlet domain coefficient of image obeys generalized Gaussian distribution, the assumed conditions-signal meeting Bayesian method of estimation obeys generalized Gaussian distribution. Therefore, utilize herein based on the Bayesian threshold estimation method estimated, estimate threshold parameter. Based on the Bayesian adaptive threshold estimated, namely $T_B={\mathrm{\σ}}_n^2/{\mathrm{\σ}}_x.$ Concrete estimating step is:

Step 130: for noise criteria difference σ_n, adopt the mediant estimation of robustness, ${\widehat{\mathrm{\σ}}}_n=\frac{1}{0.6745L_K}\underset{i=1}{\overset{L_K}{\mathrm{\Σ}}}\mathrm{median}\left(\right|S_{\mathrm{hx}}^{(K,i)}\left|\right)$ With ${\widehat{\mathrm{\σ}}}_n=\frac{1}{0.6745L_K}\underset{i=1}{\overset{L_K}{\mathrm{\Σ}}}\mathrm{median}\left(\right|S_{\mathrm{hy}}^{(K,i)}\left|\right);$

Wherein S_hx ^{(K, i)}And S_hy ^{(K, i)}(i=1 ..., L_K) for high-frequency sub-band coefficient;

Step 131: by ${\mathrm{\σ}}_y^2={\mathrm{\σ}}_x^2+{\mathrm{\σ}}_n^2,$ Have ${\widehat{\mathrm{\σ}}}_x=\sqrt{\max ({\widehat{\mathrm{\σ}}}_y^2-{\widehat{\mathrm{\σ}}}_n^2,0)},$

Wherein, ${\widehat{\mathrm{\σ}}}_y^2=\frac{1}{\mathrm{MN}}\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{S}_{\mathrm{hf}}^{(K,i)}(m,n),$ S_hf ^{(K, i)}For S_hx ^{(K, i)}Or S_hy ^{(K, i)}, it is the high-frequency sub-band coefficient considered;

Step 132: therefore can obtain, threshold value $T_B={\mathrm{\σ}}_n^2/{\mathrm{\σ}}_x.$

Step 14: all noise reduction high frequency subimages that step 13 is obtainedWithWith the corresponding low frequency subgraph obtained in step 12 as S_lxAnd S_lyImplement based on the Contourlet inverse transformation of small echo, obtain in x-axis and y-axis direction wrong respectively cutting the noise-reduced image after a:

$S_{x,a}^{\′}=T^{-1}(S_{\mathrm{lx}},S_{\mathrm{Dhx}}^{\left(\mathrm{1,1}\right)},...,S_{\mathrm{Dhx}}^{(1,L_1)},S_{\mathrm{Dhx}}^{\left(\mathrm{2,1}\right)},...,S_{\mathrm{Dhx}}^{(k,L_k)})$

With $S_{y,a}^{\′}=T^{-1}(S_{\mathrm{ly}},S_{\mathrm{Dhy}}^{\left(\mathrm{1,1}\right)},...,S_{\mathrm{Dhy}}^{(1,L_1)},S_{\mathrm{Dhy}}^{\left(\mathrm{2,1}\right)},...,S_{\mathrm{Dhy}}^{(k,L_k)})$

Wherein, T^-1(*) it is the Contourlet inverse transformation based on small echo;

Step 15: the image S that step 14 is obtained_{X, a} ^nxAnd S_{Y, a} ^nyCarry out corresponding mistake to cut the reverse mistake of amplitude and cut operation, obtain:

I_{X, a}=C_{X ,-a}(S′_{X, a}) and I_{Y, a}=C_{Y ,-a}(S′_{Y, a});

Step 16: make a+=Δ a, repeats step 11 to step 15, until a=1, stops repeating;

Step 17: to all I obtained_{X, a}And I_{Y, a}(| a |≤1) is averaging, and obtains final noise-reduced image:

${\hat{I}}_D=\frac{1}{2(2/\mathrm{\Δa}+1)}\underset{a=-1,\mathrm{\Δa}=0.05}{\overset{1}{\mathrm{\Σ}}}(I_{\mathrm{xa}}+I_{\mathrm{ya}}).$

Referring to shown in Fig. 2, Fig. 2 is a kind of noisy image noise reduction process system block diagram provided by the invention, including:

Wrong cut unit 20, carries out mistake to the noisy image inputted and cuts operation, obtain mistake and cut image; And to cutting the reverse mistake of amplitude and cut operation through carrying out identical mistake again based on the image of Contourlet transformation by reciprocal direction of small echo;

Based on the contourlet transformation unit 21 of small echo, mistake is cut image and carries out Its Sparse Decomposition, obtain low frequency subgraph picture and high frequency subimage; And carry out the Contourlet transformation by reciprocal direction based on small echo to low frequency subgraph picture with through the noise reduction high frequency subimage of noise reduction process;

Noise reduction processing unit 22, carries out noise reduction process to each high frequency subimage, obtains noise reduction high frequency subimage.

It should be noted that in the embodiment of the present invention, described noise reduction processing unit 22 is threshold deniosing processing unit, it is preferable that each high frequency subimage is carried out threshold deniosing process, obtains noise reduction high frequency subimage.

In the embodiment of the present invention, described system may further comprise:

Linear averaging unit 23, for cutting all images that operation obtains and carry out linear averaging through carrying out reverse mistake based on the contourlet transformation unit of small echo, obtaining final noise-reduced image.

Referring to shown in Fig. 3, Fig. 3 illustrates the image schematic diagram of the different-effect adopting different disposal method to obtain.

In Fig. 3, (a) is artwork; B () is noisy image, σ=60, PSNR=18.41dB; C () is the denoising result picture through Contourlet soft-threshold, σ=60, PSNR=23.69dB; D () is through translation invariant Contourlet soft-threshold denoising result picture, σ=60, PSNR=24.05dB; E () is for cutting constant Contourlet soft-threshold denoising result picture, σ=60, PSNR=25.20dB through the mistake based on small echo. Contrast it can be seen that (e) figure visually, (e) figure remains the detailed information such as image texture edge, and picture quality is substantially good than (b), (c).

Table 1 below gives the result (Y-PSNR (PSNR) value) that under various different noise criteria difference, different noise reduction process methods obtain.

Table 1

Noise criteria is poor	Contourlet Conversion	Translation invariant Contourlet Conversion	Mistake based on small echo is cut not Become contourlet transformation
				σ=20	29.32	28.55	30.05
σ=30	27.24	26.75	28.13
				σ=40	25.97	25.19	27.23
σ=50	24.11	24.38	26.04
				σ=60	23.69	24.05	25.20

Visible, cut, based on the mistake of small echo, the image PSNR index that image PSNR index is above contourlet transformation, translation invariant contourlet transformation obtains that the constant contourlet transformation noise reduction process to noisy image obtains under difference noise criteria difference of the present invention. Cut constant contourlet transformation based on the mistake of small echo and picture breakdown is had the Directional Decomposition than contourlet transformation, translation invariant contourlet transformation more horn of plenty, utilize mistake to cut operation makes the more mistake that is lost in of image detail texture information cut in image and be compensated simultaneously, thus being greatly promoted image noise reduction effect, improve picture quality.

Above content is in conjunction with specific embodiment further description made for the present invention, it is impossible to assert that specific embodiment of the invention is confined to these explanations. For general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, it is also possible to make some simple deduction or replace, protection scope of the present invention all should be considered as belonging to.

Claims (9)

1. a denoising noise image, it is characterised in that include step:

The noisy image inputted is carried out mistake and cuts operation, obtain mistake and cut image;

Arrange based on the wavelet decomposition number of plies in the contourlet transformation of small echo and the Directional Decomposition number in every layer;

Utilize the contourlet transformation based on small echo that mistake is cut image and carry out Its Sparse Decomposition multiple dimensioned, multidirectional, obtain low frequency subgraph picture and a series of high frequency subimage with different resolution;

Each high frequency subimage is carried out noise reduction process, obtains noise reduction high frequency subimage;

Low frequency subgraph picture and noise reduction high frequency subimage carrying out the Contourlet transformation by reciprocal direction based on small echo, then carries out identical mistake and cut the reverse mistake of amplitude and cut operation, described mistake cuts the absolute value of amplitude less than 1.

2. method according to claim 1, it is characterised in that the noisy image of input is carried out mistake and cuts operation and specifically include step:

Mistake is set and cuts amplitude a and mistake cuts amplitude delta Δ a, | a |≤1;

The noisy image inputted is carried out repeatedly mistake multi-direction, cut amplitude by a=a+ Δ a mistake and cuts operation, until a=1.

3. method according to claim 1 and 2, it is characterised in that still further comprise step:

Reverse mistake is cut all images of obtaining of operation and carries out linear averaging, obtain final noise-reduced image.

4. method according to claim 1, it is characterised in that each high frequency subimage is carried out noise reduction process and specifically includes:

Threshold parameter is set, utilizes soft-threshold function, each high frequency subimage is carried out threshold deniosing process.

5. method according to claim 4, it is characterised in that described threshold parameter is set before still further comprise step:

According to the wavelet decomposition number of plies, Directional Decomposition number in every layer and high frequency band coefficient, adopt the mediant estimation of robustness, obtain the noise criteria on each direction of high frequency subimage poor;

Poor according to the noise criteria on each direction, calculate and obtain threshold parameter.

6. method according to claim 4, it is characterised in that described soft-threshold function is soft-threshold function Λ (*)=sgn (*) max (*, T_B), T_BFor threshold parameter.

7. a noisy image noise reduction process system, it is characterised in that including:

Wrong cut unit, carries out mistake to the noisy image inputted and cuts operation, obtain mistake and cut image; And to cutting the reverse mistake of amplitude and cut operation through carrying out identical mistake again based on the image of Contourlet transformation by reciprocal direction of small echo, described mistake cuts the absolute value of amplitude less than 1;

Contourlet transformation unit based on small echo, for arranging based on the wavelet decomposition number of plies in the contourlet transformation of small echo and the Directional Decomposition number in every layer, utilize the contourlet transformation based on small echo that mistake is cut image and carry out Its Sparse Decomposition multiple dimensioned, multidirectional, obtain low frequency subgraph picture and a series of high frequency subimage with different resolution; And carry out the Contourlet transformation by reciprocal direction based on small echo to low frequency subgraph picture with through the noise reduction high frequency subimage of noise reduction process;

Noise reduction processing unit, carries out noise reduction process to each high frequency subimage, obtains noise reduction high frequency subimage.

8. system according to claim 7, it is characterised in that may further comprise:

Linear averaging unit, for cutting all images that operation obtains and carry out linear averaging through carrying out reverse mistake based on the contourlet transformation unit of small echo, obtaining final noise-reduced image.

9. system according to claim 7, it is characterised in that described noise reduction processing unit is threshold deniosing processing unit, for each high frequency subimage is carried out threshold deniosing process, obtains noise reduction high frequency subimage.