CN105913383A - Image noise reduction method based on image block prior estimation mixed framework - Google Patents

Abstract

The invention discloses an image noise reduction method based on an image block prior estimation mixed framework, and specifically relates to prior estimation of image blocks, and an image noise reduction method. The prior estimation involves classifying the image blocks through detecting image structures and performing image noise reduction on the classified image blocks according to different image results by using an image block expected logarithm likelihood estimation and an image block coupling-three-dimensional filtering method. According to the invention, by use of the image noise reduction method based on the image block prior estimation mixed framework, noise reduction can be realized relevantly according to different image block structures, and experiment results show that the method, compared to other noise reduction methods, has the advantage of better-robustness noise reduction.

Description

Image denoising method based on image block prior estimate combination frame

Technical field

A kind of method that the present invention relates to image noise reduction, is specifically related to figure based on image block prior estimate combination frame As noise-reduction method.

Background technology

In recent years, the problem for image noise reduction proposes a lot of algorithm.One of which method is to carry out noise reduction based on block The noise reduction of image, such as image block expectation log-likelihood algorithm for estimating (EPLL) is realized by the redundancy of estimation space image block It is to utilize Gauss model to estimate clean image block from noisy image block；Another way is the space correlation by utilizing image block Property realizes image noise reduction, and representative algorithm is Block-matching and three-dimensional filtering (BM3D) algorithm, according to the phase between image block Carry out judging the correlation of image block like property.

For image noise reduction, prior estimate is also a kind of effective method, and rarefaction representation is the most effective in prior estimate A kind of method, the method mainly by noisy image block is trained study, utilize linear expression realize image block fall Make an uproar.Method mentioned above uses different image prior information to estimate natural image, but more multiple for distribution Miscellaneous image uses a certain method then can not realize good noise reduction, it is therefore desirable to the noise reduction algorithm of a kind of more robust To tackle different noisy images, and then improve noise reduction quality.

Summary of the invention

The present invention relates to a kind of image denoising method based on image block prior estimate combination frame.Concrete technical side Case is as follows:

A kind of image denoising method based on image block prior estimate combination frame, comprises the steps of

(1) image block prior estimate

(a), first noisy image block is divided into two kinds, two kinds of sizes are respectively 8 × 8 and 40 × 40；

(b), to size be 8 × 8 noisy image block use structure detection；

Utilize the noisy image block of average weight and mate through image block-image block the structure of three-dimensional filtering method noise reduction Intermediate image block, is divided into two classes according to picture block structure attribute by noisy image block: texture and the noisy image block at edge, smooth Noisy image block with details；

(c), to size be 40 × 40 noisy image block use smooth detection；

Utilize fast noise method of estimation that noisy image block is divided into noisy smooth block, noisy non-smooth block；

(d), utilize the classification of the noisy smooth block in step (c) and noisy non-smooth block by smooth in step (b) and The noisy image block of details is subdivided into smooth noisy image block, the noisy image block of details again；

The final image block obtaining three classes 8 × 8: texture and the noisy image block at edge, smooth noisy image block and details Noisy image block；

(2) image noise reduction

Image block coupling-three-dimensional filtering method is utilized to process edge and the noisy image block of texture and smooth noisy image block, The noisy image block of details is processed by image block expectation log-likelihood method of estimation, finally image block is carried out General Office Reason.

Above-mentioned steps (b) utilizes the noisy image block of average weight and mates through image block-three-dimensional filtering method fall The image block structure intermediate image block made an uproar, constructive formula is:

$y_i^{\′}={\mathrm{\αy}}_i+(1-\mathrm{\α}){\tilde{y}}_i^{BM}$

Wherein, α is weight coefficient, y_iIt is according to noise image Y i-th block, i=1,2 ... N, N are all image blocks Number, y '_iWithIt is relevant intermediate image block and noise-reduced image block.

Structure detection in above-mentioned steps (b) is that the proximity relations of the intermediate image block according to gained judges edge and texture Image block；

Object block y ' is selected by the similar contiguous block of k_i, utilize the value of correlation between contiguous blockCalculate object block, After by image block y_iIt is divided in different classes,

$y_i\⋐\left\{\begin{array}{cc}{S}^{e,st}& {\stackrel{\‾}{R}}_i\≥\mathrm{\ϵ}and\mathrm{var}\left(y_i^{\′}\right)\≥r\\ S^{d,sm}& otherwise\end{array}\right.$

Wherein, var (.) is the characteristic value of variance operator, ε and r is two threshold values, S^e,tRepresent the noisy of edge and texture Image block, S^d,mRepresent details and smooth noisy image block.

Fast noise method of estimation in above-mentioned steps (c) is:

Each image block of 40 × 40 is carried out noise variance estimate, by with given image noise variance σ² Compare, it is achieved to noisy smoothed image block and the separation of noisy non-smooth image block, detailed process is as follows:

$B_i=\left\{\begin{array}{cc}{B}_i^{uf}& {\mathrm{\σ}}_{B_i}^2-{\mathrm{\σ}}^2\≥\mathrm{\θ}\\ B_i^f& otherwise\end{array}\right.$

Wherein, B_iRepresent the image block of 40 × 40,It is the noise variance of image block, σ²It it is the picture noise side be given Difference,WithBeing noisy smoothed image block and noisy non-smooth image block respectively, θ is threshold value.

Above-mentioned steps (d) utilizes noisy smooth block and the classification of noisy non-smooth block of 40 × 40, by 8 × 8 smooth Again being subdivided into smooth noisy image block and the noisy image block of details with the noisy image block of details, the formula of concrete classification is:

$y_i\⋐\left\{\begin{array}{cc}{S}^{sm}& N_i^f>N_i^{uf}\\ S^d& otherwise\end{array}\right.$

Wherein,WithIt is noisy image y_iNoisy smoothed image block and noisy non-smooth image block.

Image block carries out following integrated treatment by above-mentioned steps (2):

$\tilde{Y}={\left({\mathrm{\Σ}}_{i=1}^N{R}_i^T{R}_i\right)}^{-1}{\mathrm{\Σ}}_{i=1}^N{R}_i^T{\tilde{y}}_i$

Wherein,It is the image block recovered, R_iIt is image block y_iSample matrix.

Beneficial effects of the present invention is as follows:

The prior estimate of the method is to be classified by image block, according to image block by detecting picture structure Different classification, use different noise-reduction methods that image block is carried out noise reduction.Then the image block recovered by distinct methods is entered Row merges, it is achieved the noise reduction of image.Experiment shows, the noise reduction of the method performance is more effective than some noise-reduction methods.

Accompanying drawing explanation

Fig. 1 is image noise reduction combination frame of the present invention；

Fig. 2 is the algorithm noise reduction picture utilizing BM3D, EPLL and the present invention.

Detailed description of the invention

Below in conjunction with the accompanying drawings, the image denoising method based on image block prior estimate combination frame of the present invention is made into one Step illustrates.

As it is shown in figure 1, a kind of image denoising method based on image block prior estimate combination frame, comprise the steps of

(1) image block prior estimate

(a), first noisy image block is divided into two kinds, two kinds of sizes are respectively 8 × 8 and 40 × 40；

(b), to size be 8 × 8 noisy image block use structure detection；

Utilize the noisy image block of average weight and mate through image block-image block the structure of three-dimensional filtering method noise reduction Intermediate image block,

Constructive formula is:

$y_i^{\′}={\mathrm{\αy}}_i+(1-\mathrm{\α}){\tilde{y}}_i^{BM}$

Wherein, α is weight coefficient, y_iIt is according to noise image Y i-th block, i=1,2 ... N, N are all image blocks Number, y '_iWithIt is relevant intermediate image block and noise-reduced image block.

According to picture block structure attribute noisy image block is divided into two classes: texture and the noisy image block at edge, smooth and The noisy image block of details；

The proximity relations of the intermediate image block according to gained judges edge and texture image block；

Object block y ' is selected by the similar contiguous block of k_i, utilize the value of correlation between contiguous blockCalculate object block, After by image block y_iIt is divided in different classes,

$y_i\⋐\left\{\begin{array}{cc}{S}^{e,st}& {\stackrel{\‾}{R}}_i\≥\mathrm{\ϵ}and\mathrm{var}\left(y_i^{\′}\right)\≥r\\ S^{d,sm}& otherwise\end{array}\right.$

Wherein, var (.) is the characteristic value of variance operator, ε and r is two threshold values, S^e,tRepresent the noisy of edge and texture Image block, S^d,mRepresent details and smooth noisy image block.

(c), to size be 40 × 40 noisy image block use smooth detection；

Utilize fast noise method of estimation that noisy image block is divided into noisy smooth block, noisy non-smooth block；

In order to the image block of 40 × 40 being carried out smooth detection, need the method (J. utilizing fast noise to estimate"Fast Noise Variance Estimation,"Comput.Vis.Image Und.,vol.4,no.2, Pp.300 302,1996) each image block of 40 × 40 is carried out noise variance estimate, by making an uproar with given image Sound variances sigma²Compare, it is achieved to noisy smoothed image block and the separation of noisy non-smooth image block, detailed process is as follows:

$B_i=\left\{\begin{array}{cc}{B}_i^{uf}& {\mathrm{\σ}}_{B_i}^2-{\mathrm{\σ}}^2\≥\mathrm{\θ}\\ B_i^f& otherwise\end{array}\right.$

Wherein, B_iRepresent the image block of 40 × 40,It is the noise variance of image block, σ²It it is the picture noise side be given Difference,WithBeing noisy smoothed image block and noisy non-smooth image block respectively, θ is threshold value.

(d), utilize 40 × 40 noisy smooth block and the classification of noisy non-smooth block, containing of the smooth and details by 8 × 8 Image block of making an uproar is subdivided into smooth noisy image block and the noisy image block of details again, and the formula of concrete classification is:

$y_i\⋐\left\{\begin{array}{cc}{S}^{sm}& N_i^f>N_i^{uf}\\ S^d& otherwise\end{array}\right.$

Wherein,WithIt is noisy image y_iNoisy smoothed image block and noisy non-smooth image block.

The final image block obtaining three classes 8 × 8: texture and the noisy image block at edge, smooth noisy image block and details Noisy image block；

(2) image noise reduction

Image block coupling-three-dimensional filtering method is utilized to process edge and the noisy image block of texture and smooth noisy image block, The noisy image block of details is processed by image block expectation log-likelihood method of estimation, finally image block is carried out General Office Reason.

$\tilde{Y}={\left({\mathrm{\Σ}}_{i=1}^N{R}_i^T{R}_i\right)}^{-1}{\mathrm{\Σ}}_{i=1}^N{R}_i^T{\tilde{y}}_i$

Wherein,It is the image block recovered, R_iIt is image block y_iSample matrix.

Finally give the noise reduction of image as shown in Figure 2.Three groups of pictures of Fig. 2 are clean figure the most successively Sheet, noisy picture, utilize the noise reduction of the algorithm proposed in the picture of BM3D noise reduction, the picture utilizing EPLL noise reduction and the present invention Picture.Experiment shows, the noise reduction of present invention performance is more effective than some noise-reduction methods.

The above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For Yuan, on the premise of without departing from the technology of the present invention principle, it is also possible to make some improvement and deformation, these improve and deformation Also should be regarded as protection scope of the present invention.

Claims (6)

1. an image denoising method based on image block prior estimate combination frame, is characterized in that, comprise the steps of

(1) image block prior estimate

(a), first noisy image block is divided into two kinds, two kinds of sizes are respectively 8 × 8 and 40 × 40；

(b), to size be 8 × 8 noisy image block use structure detection；

Utilize the noisy image block of average weight and mate through image block-the image block structure of three-dimensional filtering method noise reduction in the middle of Image block, is divided into two classes according to picture block structure attribute by noisy image block: texture and the noisy image block at edge, smooth and thin The noisy image block of joint；

(c), to size be 40 × 40 noisy image block use smooth detection；

Utilize fast noise method of estimation that noisy image block is divided into noisy smooth block, noisy non-smooth block；

(d), utilize the classification of the noisy smooth block in step (c) and noisy non-smooth block by the smooth and details in step (b) Noisy image block be again subdivided into smooth noisy image block, the noisy image block of details；

The final image block obtaining three classes 8 × 8: texture and the noisy image block at edge, smooth noisy image block and details are noisy Image block；

(2) image noise reduction

Image block coupling-three-dimensional filtering method is utilized to process edge and the noisy image block of texture and smooth noisy image block, details Noisy image block is processed by image block expectation log-likelihood method of estimation, finally image block is carried out integrated treatment.

Image denoising method based on image block prior estimate combination frame the most according to claim 1, is characterized in that, institute State the noisy image block utilizing average weight in step (b) and mate through image block-the image block of three-dimensional filtering method noise reduction Structure intermediate image block, constructive formula is:

$y_i^{\′}={\mathrm{\αy}}_i+(1-\mathrm{\α}){\tilde{y}}_i^{BM}$

Wherein, α is weight coefficient, y_iIt is according to noise image Y i-th block, i=1,2 ... N, N are the numbers of all image blocks, y′_iWithIt is relevant intermediate image block and noise-reduced image block.

Image denoising method based on image block prior estimate combination frame the most according to claim 1, is characterized in that, institute State the proximity relations that the structure detection in step (b) is intermediate image block according to gained and judge edge and texture image block；

Object block y ' is selected by the similar contiguous block of k_i, utilize the value of correlation between contiguous blockCalculate object block, finally will figure As block y_iIt is divided in different classes,

$y_i\⋐\left\{\begin{array}{cc}{S}^{e,st}& {\stackrel{\‾}{R}}_i\≥\mathrm{\ϵ}and\mathrm{var}\left(y_i^{\′}\right)\≥r\\ S^{d,sm}& otherwise\end{array}\right.$

Wherein, var (.) is the characteristic value of variance operator, ε and r is two threshold values, S^e,tRepresent the noisy image of edge and texture Block, S^d,mRepresent details and smooth noisy image block.

Image denoising method based on image block prior estimate combination frame the most according to claim 1, is characterized in that, institute The fast noise method of estimation stated in step (c) is:

Each image block of 40 × 40 is carried out noise variance estimate, by with given image noise variance σ²Carry out Relatively, it is achieved to noisy smoothed image block and the separation of noisy non-smooth image block, detailed process is as follows:

$B_i=\left\{\begin{array}{cc}{B}_i^{uf}& {\mathrm{\σ}}_{B_i}^2-{\mathrm{\σ}}^2\≥\mathrm{\θ}\\ B_i^f& otherwise\end{array}\right.$

Wherein, B_iRepresent the image block of 40 × 40,It is the noise variance of image block, σ²It is the image noise variance be given, WithBeing noisy smoothed image block and noisy non-smooth image block respectively, θ is threshold value.

Image denoising method based on image block prior estimate combination frame the most according to claim 1, is characterized in that, institute State noisy smooth block and the classification of noisy non-smooth block utilizing 40 × 40 in step (d), containing of the smooth and details by 8 × 8 Image block of making an uproar is subdivided into smooth noisy image block and the noisy image block of details again, and the formula of concrete classification is:

$y_i\⋐\left\{\begin{array}{cc}{S}^{sm}& N_i^f>N_i^{uf}\\ S^d& otherwise\end{array}\right.$

Wherein,WithIt is noisy image y_iNoisy smoothed image block and noisy non-smooth image block.

Image denoising method based on image block prior estimate combination frame the most according to claim 1, is characterized in that, institute State and image block carried out following integrated treatment by step (2):

$\tilde{Y}={\left({\mathrm{\Σ}}_{i=1}^N{R}_i^T{R}_i\right)}^{-1}{\mathrm{\Σ}}_{i=1}^N{R}_i^T{\tilde{y}}_i$

Wherein,It is the image block recovered, R_iIt is image block y_iSample matrix.