CN103310424A - Image denoising method based on structural similarity and total variation hybrid model - Google Patents
Image denoising method based on structural similarity and total variation hybrid model Download PDFInfo
- Publication number
- CN103310424A CN103310424A CN2013102852139A CN201310285213A CN103310424A CN 103310424 A CN103310424 A CN 103310424A CN 2013102852139 A CN2013102852139 A CN 2013102852139A CN 201310285213 A CN201310285213 A CN 201310285213A CN 103310424 A CN103310424 A CN 103310424A
- Authority
- CN
- China
- Prior art keywords
- image
- omega
- sigma
- model
- dxdy
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Abstract
The invention discloses an image denoising method based on a structural similarity and total variation hybrid model. The method comprises (1), designing a functional E (u); (2), introducing a new auxiliary variable to the E (u) and turning the original model into two simple sub-models by using an alternate iterative method; (3), performing numerical solution on the two sub-models respectively by using a gradient descent method and a chambolle projection method to obtain a discrete mathematical model; (4), inputting a noisy image f; (5), performing iteration denoising on the f by using the discrete mathematical model; and (6), stopping until the iteration reaches set end conditions and outputting the denoised image. By means of the image denoising method, structural information of images can be well maintained while denoising is performed effectively, visual effects of the images are improved, and the method is applicable to denoising of natural images.
Description
Technical field
The present invention relates to the image technique process field, be specifically related to a kind of image de-noising method based on structural similarity and total variation mixture model, be applicable to the noise remove of natural image.
Background technology
Image form and transmission course in because interference of noise causes quality to descend, this has had a strong impact on the correct understanding that people convey a message to image, so before image carried out subsequent treatment, must first carry out denoising to image.
The method of relevant image denoising has a lot, roughly can be divided into two big classes: spatial domain denoising method and transform domain denoising method.The spatial domain denoising method is directly pixel to be handled, and representational algorithm has mean filter algorithm and middle finger filtering algorithm.The different characteristic that the transform domain denoising method mainly utilizes useful signal and noise signal to show in transform domain is removed noise effectively.Representational algorithm has based on the denoise algorithm of Fourier transform with based on the denoise algorithm of wavelet transformation.
At present, the airspace filter method has had very big development, many new methods occurred, as based on the method for fuzzy mathematics, based on method of partial differential equation etc.Particularly based on the method for partial differential equation, each image processing field such as cut apart at image denoising, image and obtained great success, become image handle with analyze in important tool and research focus.And wherein most typical representative is by ROF, and Osher and Fatemi propose the ROF model based on TV.The upsurge of image transaction module researchs such as this model started image denoising based on TV, cut apart, and become image and handle one of research focus.
Though the ROF model can keep edge of image preferably when carrying out denoising, two significant disadvantages are arranged also: the one, be easy to generate " staircase effect ", the 2nd, bad to detailed information such as the texture maintenance of image.At these problems, many scholars improve the ROF model, but these improvement great majority are the regular terms at the ROF model, and a loyal improved research is very limited for model.In fact the loyal item of ROF model uses L
2Modeling is carried out to " vibration " composition (comprising texture, noise) of image in the space, and uses L
2Tolerance is portrayed.In fact L
2Norm has only reflected the difference between image list pixel, and it is structural to have ignored image space.Yet natural image is highly structural, and the pixel that relatively approaches in the very strong relevance, particularly spatial domain is namely arranged between pixel, and this relevance is containing the important information of object structures in the visual scene.Therefore, use L
2Norm is as loyal of denoising model, and denoising result can not be preferably be consistent with the visual characteristic of human eye, thereby has reduced the visual effect of recovering image.
Summary of the invention
In view of the deficiencies in the prior art, the present invention aims to provide a kind of image denoising new method based on structural similarity and total variation mixture model.Specifically, the angle structural information that structural similarity is formed from image is defined as and is independent of brightness, contrast, and reflects the structure attribute of object in the image.It is the combination of brightness, contrast and three different factors of structure with distortion modeling, and with the estimation of average as brightness, the estimation that standard deviation is spent as a comparison, covariance is as the estimation of structural similarity.The present invention introduces structural similarity and replaces L in loyal of image denoising model
2Norm can make image keep original structure information better in the denoising process, improves the visual effect of recovering image.
To achieve these goals, the technical solution used in the present invention is as follows:
A kind of image de-noising method based on structural similarity and total variation mixture model has target image, said method comprising the steps of:
(1) design functional E (u);
(2) in described functional E (u), introduce auxiliary variable, obtain its equivalent form of value E
*(u);
(3) utilization replaces alternative manner with described functional E
*(u) be converted into two submodels, model 1 and model 2;
(4) use gradient descent method, model 2 to use the chambolle projecting method to carry out numerical solution, and obtain the discrete mathematics model to described model 1;
(5) input noise image f;
(6) utilize described discrete mathematics model that image f is carried out the iteration denoising;
When (7) reaching the iteration stopping condition, the image after the output denoising.
What need further specify is that described functional E (u) is E (u)=∫
Ω(1-SSIM (f, u)) dxdy+ λ ∫
Ω| ▽ u|dxdy, wherein, described ∫
Ω| ▽ u|dxdy is regular terms, in order to represent the bound term of level and smooth of output image, described ∫
Ω(1-SSIM (f, u)) dxdy is loyal, strengthens in order to the contrast of representing the initial observation image of output image u f, λ is regular parameter, in order to balance regular terms and loyal item, u (x, y) be in the Ω of image support territory, coordinate position is (x, the grey scale pixel value of y) locating.
What need further specify is described new functional E
*(u) be:
Wherein, the auxiliary variable of v for introducing, μ is the parameter of penalty, is used for guaranteeing that u and v are fully approaching.
Need to prove that described model 1 is respectively with model 2:
Need to prove, described use gradient descent method solving model 1, corresponding Eulerian equation is:
Need to prove, described
For:
Wherein, x, y represent respectively to make the column vector with same dimension from the image block that the same spatial location of image f and u is extracted, the sum of M presentation video piece, and m represents the number of pixels of topography's piece, 1 each element of expression all is 1 column vector.
Need to prove described A
1, A
2, B
1, B
2Be respectively
A
1=2μ
xμ
y+C
1,A
2=2σ
xy+C
2,
Wherein, μ
x, μ
yThe average of representing x and y respectively, σ
x, σ
yRepresent x and y standard deviation respectively, σ
XyThe covariance of expression x and y, C
1, C
2Be constant.
Need to prove described μ
x, μ
y, σ
x, σ
y, σ
Xy, C
1, C
2Be respectively:
Wherein, x
i, y
i(i=1 2...N) is respectively the element of x and y, and N is the dimension of x and y, ω
iBe the weighting coefficient of i element and satisfy
K
1=0.01, K
2=0.03, L=255.
Need to prove that with chambolle projecting method solving model (2), the iterative formula that obtains is in the described step (3)
Wherein,
τ represents the iteration interval time parameter.
Beneficial effect of the present invention is:
1, because the present invention introduces structural similarity in loyal of denoising model, to replace original L
2Norm, the visual effect of image improves so that image can keep original structural information preferably in the denoising process in institute;
2, the present invention utilizes alternately process of iteration in the model solution process, and two variable mutual restriction on calculating influence each other like this, alternately calculate, and makes reconstructed image develop to more excellent direction, till obtaining a better image;
3, the iteration stopping principle is to recover image u at L among the present invention
2Fully near initial observation image f, make this model have structural similarity and L concurrently like this under the norm meaning
2Norm advantage separately, the denoising effect of image is more excellent.
Description of drawings
Fig. 1 is the process flow diagram of image de-noising method of the present invention;
Fig. 2 is the original test sample figure of image de-noising method of the present invention;
Fig. 3 is the exemplary plot behind the interpolation white Gaussian noise of image de-noising method of the present invention;
Fig. 4 is ROF model denoising exemplary plot in the prior art;
Fig. 5 is the denoising exemplary plot of image de-noising method of the present invention.
Embodiment
The invention will be further described below in conjunction with accompanying drawing.
As shown in Figure 1, the present invention is a kind of image de-noising method based on structural similarity and total variation mixture model, has target image, said method comprising the steps of:
Step 1, design functional E (u);
Need to prove that for convenience of description, present embodiment is at a width of cloth digital picture, (x y) is illustrated in the Ω of image support territory u, and coordinate is (x, the grey scale pixel value of y) locating.
Wherein, ▽ u presentation video u (x, gradient fields y): ▽ u=(u
x, u
y);
▽ u has reflected near the situation of change any point in the image, and gradient magnitude is represented the speed that changes, the direction that the direction indication of gradient changes.
SSIM (f, u) the structural similarity degree of presentation video f and u also is structural similarity:
Wherein, x, y represent respectively to make the column vector with same dimension from the image block that the same spatial location of f and u is extracted, and M represents the sum of topography's piece.SSIM (x y) is defined as:
Wherein, μ
x, μ
yThe average of representing x and y respectively, σ
x, σ
yRepresent x and y standard deviation respectively, σ
XyThe covariance of expression x and y, C
1, C
2Be constant.Their concrete computing formula is as follows:
Wherein, x
i, y
i(i=1 2...N) is respectively the element of x and y, and N is the dimension of x and y, ω
iBe the weighting coefficient of i element and satisfy
K
1=0.01, K
2=0.03, L=255.
From the definition of structural similarity as can be seen, SSIM (f, value u) between-1 to 1, and SSIM (f, value u) is the bigger the better.
Take all factors into consideration the flatness of image and the maximum of original structure information is kept, structural similarity maximum at target image u and noise image f, perhaps (1-SSIM (f, u)) under the least meaning, make the regular terms minimum of image smoothing item, namely seek to strengthen image u, make functional E (u) minimum: E (u)=∫
Ω(1-SSIM (f, u)) dxdy+ λ ∫
Ω| ▽ u|dxdy;
Wherein,
▽ u represents the gradient of noise image u,
With
Be respectively image at the single order partial derivative of x and y direction.
Above-mentioned functional comprises two parts, regular terms and loyal item, wherein, ∫
Ω| ▽ u|dxdy is regular terms, the bound term that the expression output image is level and smooth, ∫
Ω(1-SSIM (f, u)) dxdy is loyal of the noise image f of output image u.This loyalty item requires image after the denoising and input picture that similarity on structure and the content should be arranged, and regular parameter λ plays important equilibrium activity between regular terms and loyal.
Step 2 is introduced auxiliary variable v in functional E (u), obtain its equivalent form of value E
*(u).
Wherein, μ is the parameter of penalty, is used for guaranteeing that u and v are fully approaching.
The minimal value of functional E (u) found the solution change into functional E of equal valuely
*Finding the solution (u).
Step 3 is utilized alternately iterative strategy E
*(u) be following two simple submodels, model 1 and model 2:
Wherein, n is the iterations of variable.
Step 4 is used gradient descent method and chambolle projecting method respectively to model 1 and model 2
Carry out numerical solution, obtain the discrete mathematics model.
Wherein use the Eulerian equation of gradient descent method solving model 1 correspondence to be:
Wherein, x, y represent respectively to make the column vector with same dimension from the image block that the same spatial location of image f and u is extracted, the sum of M presentation video piece, and m represents the number of pixels of topography's piece, 1 each element of expression all is 1 column vector, and
A
1=2μ
xμ
y+C
1,A
2=2σ
xy+C
2,
Wherein, μ
x, μ
yThe average of representing x and y respectively, σ
x, σ
yRepresent x and y standard deviation respectively, σ
XyThe covariance of expression x and y, C
1, C
2Be constant.Their computing formula is respectively:
C
1=(K
1×L)
2,C
2=(K
2×L)
2;
Wherein, x
i, y
i(i=1 2...N) is respectively the element of x and y, and N is the dimension of x and y, ω
iBe the weighting coefficient of i element and satisfy
K
1=0.01, K
2=0.03, L=255.
Fixed variable v, according to the Eulerian equation of model 1, utilize the gradient descent method can obtain mathematical model:
Wherein, t represents the time.
And then derive its discrete mathematical model:
Namely
Wherein, τ represents the iteration interval time parameter, and n represents the iterations parameter.
Utilize chambolle projecting method solving model (2), the solution formula that obtains is:
Utilize alternately iterative idea, obtained the iterative formula of u and v by (3) and (4).
Step 5 is imported the original denoising image f that treats;
Step 6 utilizes the discrete mathematics model that image f is carried out the iteration denoising, obtains denoising image u one time;
Step 7 is proceeded the iteration denoising to image u, up to reaching the iteration stopping condition, the image u after the output denoising.
Finish an iteration denoising of this method by above-mentioned steps (1) to step (6).Need to prove that along with the increase of iterations, it is smooth that image more and more is tending towards.But iterations is too much, and image is too smooth, causes image blurring.Therefore, set the iteration stopping condition:
Wherein, ε is very little constant.
Further understand technical scheme of the present invention and effect by following experiment.
Experiment condition and content
Experiment condition: test employed input picture as shown in Figure 2, wherein, figure (2a) is test pattern House, and figure (2b) is test pattern Pepper, and Fig. 3 is that figure (2a) adding standard deviation is 20 noise image.
Experiment content: under above-mentioned experiment condition, select the comparison that experimentizes of existing ROF denoise algorithm and the inventive method for use.The objective evaluation index of denoising result is weighed with Y-PSNR PSNR and structural similarity SSIM.
Experiment 1 is carried out denoising with existing ROF method to figure (3), and what numerical solution adopted is the algorithm of chambolle projection, wherein, and regular parameter λ=18, iteration interval time parameter τ=0.005.As shown in Figure 4, the method possesses certain denoising ability, but it is undesirable to recover image visual effect, can not keep the original structure information of image well, and be easy to generate staircase effect.
Experiment 2 is carried out denoising, the parameter setting with the inventive method to figure (3): regular parameter λ=0.08, iteration interval time parameter τ=0.01, stop condition ε=10
-3As shown in Figure 5, the inventive method has significant noise inhibiting ability, and the image after the denoising seems smoother, nature, and the structural information of image keeps ground better, details is also more clear, and than existing ROF method, the visual effect of recovering image is greatly improved.
Experiment 3, it is 20,30,40,50 Gauss's additive white noises that image among figure (2a) and the figure (2b) is added noise criteria difference σ respectively, with PSNR and the SSIM evaluation index as denoising effect, existing ROF denoising method and method of the present invention are compared, and PSNR value and the SSIM value of two kinds of method denoising results are listed in table 1 and table 2 respectively.
The PSNR of two kinds of denoise algorithm (dB) value relatively in the table 1
The SSIM value of two kinds of denoise algorithm relatively in the table 2
Can draw denoising result of the present invention according to table 1 and table 2 and on PSNR and SSIM evaluation index, all make moderate progress, especially be significantly improved in the SSIM value.Table 1 and table 2 have objectively shown superiority of the present invention.
For a person skilled in the art, can make other various corresponding changes and distortion according to technical scheme described above and design, and these all changes and distortion should belong within the protection domain of claim of the present invention all.
Claims (9)
1. the image de-noising method based on structural similarity and total variation mixture model has target image, it is characterized in that, said method comprising the steps of:
(1) design functional E (u);
(2) in described functional E (u), introduce auxiliary variable, obtain its equivalent form of value E
*(u);
(3) utilize the alternately described functional E of alternative mannerization
*(u) be two submodels, model 1 and model 2;
(4) use gradient descent method, model 2 to use the chambolle projecting method to carry out numerical solution, and obtain the discrete mathematics model to described model 1;
(5) input noise image f;
(6) utilize described discrete mathematics model that image f is carried out the iteration denoising;
When (7) reaching the iteration stopping condition, the image after the output denoising.
2. image de-noising method according to claim 1 is characterized in that, described functional E (u) is E (u)=∫
Ω(1-SSIM (f, u)) dxdy+ λ ∫
Ω| ▽ u|dxdy, wherein, described ∫
Ω| ▽ u|dxdy is regular terms, in order to represent the bound term of level and smooth of output image, described ∫
Ω(1-SSIM (f, u)) dxdy is loyal, strengthens in order to the contrast of representing the initial observation image of output image u f, λ is regular parameter, in order to balance regular terms and loyal item, u (x, y) be in the Ω of image support territory, coordinate position is (x, the grey scale pixel value of y) locating.
3. image de-noising method according to claim 1 is characterized in that, described new functional E
*(u) be
Wherein, the auxiliary variable of v for introducing, μ is the parameter of penalty, is used for guaranteeing fully approaching of u and v.
4. image denoising new method according to claim 1 is characterized in that, described model 1 is respectively with model 2:
5. image de-noising method according to claim 1 is characterized in that, described use gradient descent method solving model 1, and corresponding Eulerian equation is:
6. image de-noising method according to claim 5 is characterized in that, and is described
For:
Wherein, x, y represent respectively to make the column vector with same dimension from the image block that the same spatial location of image f and u is extracted, the sum of M presentation video piece, and m represents the number of pixels of topography's piece, 1 each element of expression all is 1 column vector.
7. image de-noising method according to claim 6 is characterized in that, described A
1, A
2, B
1, B
2Be respectively
A
1=2μ
xμ
y+C
1,A
2=2σ
xy+C
2,
Wherein, μ
x, μ
yThe average of representing x and y respectively, σ
x, σ
yRepresent x and y standard deviation respectively, σ
XyThe covariance of expression x and y, C
1, C
2Be constant.
8. image de-noising method according to claim 7 is characterized in that, described μ
x, μ
y, σ
x, σ
y, σ
Xy, C
1, C
2Be respectively:
9. image de-noising method according to claim 1 is characterized in that, with chambolle projecting method solving model (2), the solution formula that obtains is in the described step (3)
Wherein,
τ represents the iteration interval time parameter.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310285213.9A CN103310424B (en) | 2013-07-08 | 2013-07-08 | A kind of image de-noising method based on structural similarity Yu total variation hybrid model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310285213.9A CN103310424B (en) | 2013-07-08 | 2013-07-08 | A kind of image de-noising method based on structural similarity Yu total variation hybrid model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103310424A true CN103310424A (en) | 2013-09-18 |
CN103310424B CN103310424B (en) | 2016-12-28 |
Family
ID=49135602
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201310285213.9A Expired - Fee Related CN103310424B (en) | 2013-07-08 | 2013-07-08 | A kind of image de-noising method based on structural similarity Yu total variation hybrid model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103310424B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103955893A (en) * | 2014-04-11 | 2014-07-30 | 西安理工大学 | Image denoising method based on separable total variation model |
CN106204461A (en) * | 2015-05-04 | 2016-12-07 | 南京邮电大学 | Compound regularized image denoising method in conjunction with non local priori |
CN108122205A (en) * | 2016-11-29 | 2018-06-05 | 北京东软医疗设备有限公司 | The denoising method and device of ultrasonoscopy |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102663689A (en) * | 2012-03-22 | 2012-09-12 | 西安电子科技大学 | SAR image speckle suppression based on area division and non-local total variation |
-
2013
- 2013-07-08 CN CN201310285213.9A patent/CN103310424B/en not_active Expired - Fee Related
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102663689A (en) * | 2012-03-22 | 2012-09-12 | 西安电子科技大学 | SAR image speckle suppression based on area division and non-local total variation |
Non-Patent Citations (3)
Title |
---|
ANTONIN CHAMBOLLE: "An Algorithm for Total Variation Minimization and Applications", 《JOURNAL OF MATHEMATICAL IMAGING AND VISION》 * |
王静等: "基于分裂Bregman方法的全变差图像去模糊", 《电子学报》 * |
王静等: "基于分裂Bregman方法的全变差图像去模糊", 《电子学报》, vol. 40, no. 8, 31 August 2012 (2012-08-31) * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103955893A (en) * | 2014-04-11 | 2014-07-30 | 西安理工大学 | Image denoising method based on separable total variation model |
CN103955893B (en) * | 2014-04-11 | 2017-02-01 | 西安理工大学 | Image denoising method based on separable total variation model |
CN106204461A (en) * | 2015-05-04 | 2016-12-07 | 南京邮电大学 | Compound regularized image denoising method in conjunction with non local priori |
CN106204461B (en) * | 2015-05-04 | 2019-03-05 | 南京邮电大学 | In conjunction with the compound regularized image denoising method of non local priori |
CN108122205A (en) * | 2016-11-29 | 2018-06-05 | 北京东软医疗设备有限公司 | The denoising method and device of ultrasonoscopy |
Also Published As
Publication number | Publication date |
---|---|
CN103310424B (en) | 2016-12-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Xu et al. | An improved anisotropic diffusion filter with semi-adaptive threshold for edge preservation | |
CN107016642B (en) | Method and apparatus for resolution up-scaling of noisy input images | |
Yang et al. | Non-local means theory based Perona–Malik model for image denosing | |
CN103116873B (en) | Image denoising method | |
CN104156918B (en) | SAR image noise suppression method based on joint sparse representation and residual fusion | |
CN101980284A (en) | Two-scale sparse representation-based color image noise reduction method | |
CN108932699B (en) | Three-dimensional matching harmonic filtering image denoising method based on transform domain | |
CN103020918A (en) | Shape-adaptive neighborhood mean value based non-local mean value denoising method | |
CN105869133A (en) | Image sharpening method based on non-causal fractional order subdifferential | |
CN106204502A (en) | Based on mixing rank L0regularization fuzzy core method of estimation | |
CN103208104A (en) | Non-local theory-based image denoising method | |
Liu et al. | SGTD: Structure gradient and texture decorrelating regularization for image decomposition | |
CN104200434B (en) | Non-local mean image denoising method based on noise variance estimation | |
CN103310424A (en) | Image denoising method based on structural similarity and total variation hybrid model | |
CN115082336A (en) | SAR image speckle suppression method based on machine learning | |
Ouyang et al. | Research on DENOISINg of cryo-em images based on deep learning | |
Wang et al. | Difference curvature driven anisotropic diffusion for image denoising using Laplacian kernel | |
Raghuvanshi et al. | Analysing image denoising using non local means algorithm | |
CN103310429B (en) | Image enhancement method based on hidden Markov tree (HMT) model in directionlet domain | |
CN106327440B (en) | Picture breakdown filtering method containing non-local data fidelity term | |
Kozhemiakin et al. | Efficiency analysis for 3D filtering of multichannel images | |
Shakenov | Algorithms of background suppression in the problem of detection of point targets in images | |
Zheng et al. | Regularization parameter selection for total variation model based on local spectral response | |
Zachevsky et al. | Combining long-range dependencies with phase information in Natural Stochastic Texture enhancement | |
Tu et al. | A Sobel-TV based hybrid model for robust image denoising |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20161228 Termination date: 20210708 |