CN103208097B - Filtering method is worked in coordination with in the principal component analysis of the multi-direction morphosis grouping of image - Google Patents

Abstract

Filtering method is worked in coordination with in the principal component analysis that the invention discloses the multi-direction morphosis grouping of a kind of image.The method carries out overlap partition to image, carries out multi-direction morphosis grouping, obtain smooth piece of group, randomized block group, direction edge block group and direction texture block group according to the variance of image block, gradient and singular value information to image block; Then the principal component analysis of feature self-adaptation is carried out to the image block of grouping, utilize hard-threshold shrinkage method to carry out filtering to image block conversion coefficient, and then carry out grouping reconstruct; Finally grouping reconstructed blocks is polymerized to view picture filtering image.The inventive method has taken into full account the multi-direction morphosis of image block and the non local affinity information of image, image filtering process has good structure and Acacia crassicarpaA characteristic, stress release treatment is very capable, can be widely used in the preprocessing process of image characteristics extraction and target detection.

Description

Filtering method is worked in coordination with in the principal component analysis of the multi-direction morphosis grouping of image

Technical field

The invention belongs to digital image processing techniques field, filtering method is worked in coordination with in the principal component analysis of the particularly multi-direction morphosis grouping of a kind of image.

Background technology

In the acquisition and transmitting procedure of digital picture, inevitably introduce various noise.The existence of noise, not only destroy the real information of image, but also have a strong impact on the visual effect of image, make to become difficulty to the post-processed of image as feature extraction and target detection etc., therefore filtering is carried out to noise image and just become a kind of effective image pre-processing method.Image filtering method can be divided into the method for spatial domain and frequency domain.Spatial domain method directly processes image pixel, as mean filter, the medium filtering and bilateral filtering etc. of classics.Frequency domain method first by image conversion to frequency field, then process conversion coefficient, last another mistake transforms to spatial domain to reach the object of filtering, as wavelet transformation, multi-scale geometric analysis and rarefaction representation etc.

Based on the importance of image filtering, a lot of scholar has carried out a large amount of correlative study work.The human hairs such as Wang Hongmei understand a kind of self-adapting method for filtering image (patent No.: 200610043000.5) keeping edge.First the method adopts coefficient correlation method marked pixels to be noise or edge, according to label information, utilize adaptive neighborhood Shrinkage Wavelet coefficient, in denoising simultaneously, there is certain edge hold facility, but do not make full use of non local similarity and the local otherness of picture material.The human hairs such as Wang Xigui understand a kind of image de-noising method (patent No.: 200610065679.8) of process of multi-template mixed filtering.The method first defines one group of wave filter, then to image block, considers the otherness of image block, according to its homogeneous degree, some wave filters are selected to carry out denoising to image block, can retaining part image detail information, but do not take into full account the non local similarity of image block.The human hairs such as Liu Fang understand a kind of non-local mean image de-noising method (number of patent application: 201110091450.2) of integrated structure information.The method first utilizes primal sketch to extract picture structure, image is divided into structural area and smooth area, then adopts non-local mean method to image denoising respectively.These methods improve the effect of image denoising to a certain extent, but all do not consider the morphosis that picture material has and multidirectional, thus cannot accomplish good equilibrium in noise remove and Hemifusus ternatanus.

Summary of the invention

The object of the invention is to provide the principal component analysis of the multi-direction morphosis grouping of a kind of image to work in coordination with filtering method, by the collaborative filtering of the image block of equidirectional same architectural characteristic, make full use of the non local affinity information of image, while filtering noise, better keep the detailed information such as the edge of image and texture, improve the visual effect of image denoising.

The technical solution realizing the object of the invention is: filtering method is worked in coordination with in the principal component analysis of the multi-direction morphosis grouping of a kind of image, and step is as follows:

Step 1: input pixel size is the noise image of M × N, overlap partition is carried out to it, namely from the image upper left corner, extract the image block P that pixel size is K × K successively, by each mobile location of pixels of row, finally can obtain (M-K+1) × (N-K+1) individual size is the image block of K × K;

Step 2: to all image blocks obtained, carries out multi-direction morphosis grouping according to the variance of image block, gradient and singular value information to it, thus can obtain smooth piece of group, randomized block group, direction edge block group and direction texture block group;

Step 3: to the group image block obtained, carries out the principal component analysis of feature self-adaptation, thus takes suitable conversion, comprises principal component analysis conversion, the conversion of row two-dimension principal component analysis and the conversion of row two-dimension principal component analysis, obtains feature adaptive transformation coefficient;

Step 4: to the feature adaptive transformation coefficient of the group image block obtained, carry out threshold value contraction, threshold function table is $\mathrm{\&Psi;}\left(x\right)=\left\{\begin{array}{c}x,\left|x\right|\&GreaterEqual;\mathrm{HT}\\ 0,\left|x\right|<\mathrm{HT}\end{array}\right.,$ Thus obtain filter factor;

Step 5: to the filter factor of the group image block obtained, the inverse transformation of suitable transform selected by carry out step 3, thus the group image block obtaining filtering reconstruct;

Step 6: to the image block of all filtering reconstruct obtained, be averaged polymerization, generate full width filtering image, namely from the image upper left corner, for each pixel, the image block of all filtering reconstruct comprising current pixel point is found, the current pixel point of all image blocks found is averaged, obtain the denoising result of current pixel point, by each mobile location of pixels of row, thus full width filtering image can be obtained.

The present invention compared with prior art, its remarkable advantage: (1) The present invention gives a kind of group technology of image block, the method not only considers non local similarity and local aggregation between image, also use different shape structure and the directional difference of image block, give more effective and reasonable packet mode.(2) the present invention is to difference group image, considers the otherness of its content and feature, uses feature self-adaptation principal component analysis conversion to carry out collaborative filtering, thus can obtain better denoising effect.(3) the present invention well maintains the principal visual impression part of image simultaneously in denoising, particularly in edge and grain details, thus the tasks such as subsequent characteristics extraction and target detection is more easily carried out.

Below in conjunction with accompanying drawing, the present invention is described in further detail.

Accompanying drawing explanation

Fig. 1 is that filtering method process flow diagram is worked in coordination with in the principal component analysis that the present invention is based on the multi-direction morphosis grouping of image.

Fig. 2 is the multi-direction morphosis grouping process flow diagram of the image block in the present invention.

Fig. 3 is that filtering process flow diagram is worked in coordination with in the principal component analysis of group image block in the present invention.

Fig. 4 is standard testing image barbara of the present invention and adds the degraded image that Gauss's white noise variance is 20.

Fig. 5 is multi-direction morphosis group technology in the present invention to the cluster display figure of test pattern and degraded image.

Fig. 6 is a denoising Experimental comparison design sketch of the inventive method and classic method.

Fig. 7 is the Acacia crassicarpaA effect contrast figure during a denoising of the inventive method and classic method is tested.

Embodiment

Filtering method is worked in coordination with in the principal component analysis of the multi-direction morphosis grouping of image of the present invention, that a point block operations is carried out to image, consider morphosis and the different directions character of image block, variance, gradient and singular value information is utilized to carry out multi-direction morphosis grouping, use the principal component analysis of feature self-adaptation to carry out collaborative filtering to the image block of different grouping, filtered image block is polymerized to complete denoising image.Concrete steps are as follows:

Step 1: input pixel size is the noise image of M × N, overlap partition is carried out to it, concrete grammar is: from the image upper left corner, extract the image block P that pixel size is K × K successively, by each mobile location of pixels of row, finally can obtain (M-K+1) × (N-K+1) individual size is the image block of K × K.Wherein the span of K is [7,15].

Step 2: multi-direction morphosis grouping is carried out to all image blocks that step 1 obtains, namely to all image blocks obtained, according to the variance of image block, gradient and singular value information, multi-direction morphosis grouping is carried out to it, thus smooth piece of group, randomized block group, direction edge block group and direction texture block group can be obtained.Concrete grammar is:

The image block P of 2.1 pairs of each K × K sizes, is labeled as smooth piece or Non-smooth surface block.Mark criterion is: the variance of computed image block P $\mathrm{var}\left(P\right)=\frac{1}{K\&times;K}\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{K}{\mathrm{\&Sigma;}}}{(P_{i,j}-E\left(P\right))}^2,$ Wherein p _i,jfor image block P is at the pixel value at locus (i, j) place, if var (P)≤TH, then marking image block P is smooth piece, otherwise, be labeled as Non-smooth surface block.Wherein TH is threshold parameter, and value is K × σ ², σ ²for noise variance.

If the image block P of 2.2 K × K sizes is Non-smooth surface blocks, it is labeled as randomized block or block structure again.Mark criterion is: by having the difference operator computed image block P of cyclic boundary condition at locus (i, j) horizontal and vertical direction difference: as 1 < i < K, during 1 < j < K, image block P is at the difference D of the horizontal direction of locus (i, j) _x(P _i,j)=P _{i+1, j}-P _i,jwith the difference D of vertical direction _y(P _i,j)=P _{i, j+1}-P _i,j; Work as i=K, during j=K, image boundary need take the difference D in cyclic boundary condition calculated level direction _x(P _k,j)=P _k,j-P _{k-1, j}with the difference D of vertical direction _y(P _i,K)=P _i,K-P _{i, K-1}, then the gradient of image block P at locus (i, j) place is G _i,j=[D _x(P _i,j), D _y(P _i,j)], by the gradient G of all locus, be arranged in order left to bottom right from image block by row, obtain g=[g ₁, g ₂..., g _{k × K}] ^t=[G _1,1, G _2,1..., G _k,K] ^t.SVD decomposition is carried out to g, obtains g=USV ^t.First row v in V ₁master corresponding to image block gradient is dominant direction, secondary series v ₂corresponding to the secondary direction that is dominant of image block gradient, the singular value of its correspondence is respectively s ₁and s ₂.Calculated direction is dominant and measures R=(s ₁-s ₂)/(s ₁+ s ₂), given threshold value R ^*if, R≤R ^*, then marking image block P is randomized block, otherwise, be labeled as block structure.Wherein R ^*span is [0.01,0.2].

If the image block P of 2.3 K × K sizes is block structures, it is labeled as multi-direction block structure again.Mark criterion is: to be dominant direction v according to the master of image block P ₁, the directivity d=arctan (v of computing block ₁(2)/v ₁(1)).Interval [0, π] is divided into W part, and it is interval which d belongs to, and is that direction structure block with regard to marking image block P.Wherein the span of W is [4,12].

If the image block P of 2.4 K × K sizes is direction structure blocks, it is labeled as again direction edge block or direction texture block.Mark criterion is: the structural strength of computed image block P every bit wherein Ω _iit is the neighborhood of the z × z size centered by current point i.Given threshold tau, if ST (i) > is τ, then thinks that i point is system point, is designated as M (i)=1, otherwise, be designated as M (i)=0.The number of all system points in statistical picture block P if M (P) < (K × K)/2, then marking image block P is direction edge block, otherwise, be labeled as direction texture block.Wherein to be 3, τ value be the value of z $0.01\&times;\max {\left\{\mathrm{ST}\left(i\right)\right\}}_1^{M\&times;N}.$

Step 3: the principal component analysis of feature self-adaptation is carried out to the group image block that step 2 obtains, to the group image block obtained, carry out the principal component analysis of feature self-adaptation, thus take suitable conversion, comprise principal component analysis (PCA) conversion, row two-dimension principal component analysis (R2DPCA) conversion and row two-dimension principal component analysis (C2DPCA) conversion, obtain feature adaptive transformation coefficient.Concrete grammar is: to L group image block P ₁, P ₂..., P _t, calculate covariance matrix Σ, Σ of its principal component analysis (PCA), row two-dimension principal component analysis (R2DPCA) and row two-dimension principal component analysis (C2DPCA) respectively _rand Σ _c:

$\mathrm{\&Sigma;}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}(p_i-\stackrel{\&OverBar;}{p}){(p_i-\stackrel{\&OverBar;}{p})}^T,$ Wherein $\stackrel{\&OverBar;}{p}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{p}_i,$ P _ifor by P _ivector is pulled into by row, ${\mathrm{\&Sigma;}}_r=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{(p_i-\stackrel{\&OverBar;}{p})}^T(p_i-\stackrel{\&OverBar;}{p}),$ ${\mathrm{\&Sigma;}}_c=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}(p_i-\stackrel{\&OverBar;}{p}){(p_i-\stackrel{\&OverBar;}{p})}^T,$ Wherein $\stackrel{\&OverBar;}{p}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{p}_i.$

Covariance matrix is carried out Eigenvalues Decomposition respectively, and obtaining the corresponding eigenwert by descending sort is λ ₁>=λ ₂>=...>=λ _{k × K}, η ₁>=η ₂>=...>=η _kand ξ ₁>=ξ ₂>=...>=ξ _k.Calculate $\mathrm{SR}=\max \{{\mathrm{\&Sigma;}}_{i=1}^{K\&times;K}{\mathrm{\&lambda;}}_i^2,{\mathrm{\&Sigma;}}_{i=1}^K{\mathrm{\&eta;}}_i^2,{\mathrm{\&Sigma;}}_{i=1}^K{\mathrm{\&xi;}}_i^2\},$ If $\mathrm{SR}={\mathrm{\&Sigma;}}_{i=1}^{K\&times;K}{\mathrm{\&lambda;}}_i^2,$ Then PCA conversion is used to this group image block, if then R2DPCA conversion is used to this group image block, if then C2DPCA conversion is used to this group image block.

Step 4: to the feature adaptive transformation coefficient of the image block that step 3 obtains, carry out threshold value shrink process, threshold function table is $\mathrm{\&Psi;}\left(x\right)=\left\{\begin{array}{c}x,\left|x\right|\&GreaterEqual;\mathrm{HT}\\ 0,\left|x\right|<\mathrm{HT}\end{array}\right.,$ Thus obtain filter factor.Wherein the value of threshold value HT is 2.75 × σ ², σ ²for noise variance.

Step 5: to the filter factor of the group image block that step 4 obtains, the inverse transformation of suitable transform selected by carry out step 3, thus the group image block obtaining filtering reconstruct.

Step 6: to the image block of all filtering reconstruct obtained, be averaged polymerization, generates full width filtering image.Concrete grammar is: from the image upper left corner, for each pixel, find the image block of all filtering reconstruct comprising current pixel point, the current pixel point of all image blocks found is averaged, obtain the denoising result of current pixel point, by each mobile location of pixels of row, thus full width filtering image can be obtained.

Embodiment

With reference to Fig. 1, filtering method is worked in coordination with in the principal component analysis that the present invention is based on the multi-direction morphosis grouping of image, and concrete steps are as follows:

Step 1: piecemeal is carried out to the barbara noise image (as Fig. 4 .b) of 256 × 256 sizes.Piecemeal criterion is: from the image upper left corner, extracts the image block P of 11 × 11 sizes, by each mobile location of pixels of row, can obtain 246 × 246 image blocks.

Original true picture for the barbara image shown in Fig. 4 .a, size be 256 × 256, gray level is 256, is σ to its variance adding white Gaussian noise ²=20, obtain noise image as Fig. 4 .b.

Step 2: multi-direction morphosis grouping is carried out to all image blocks that step 1 obtains.As shown in Figure 2, concrete grammar is:

The image block P of 2.1 pairs of each 11 × 11 sizes, is labeled as smooth piece or Non-smooth surface block.Mark criterion is: the variance of computed image block P $\mathrm{var}\left(P\right)=\frac{1}{11\&times;11}\underset{i=1}{\overset{11}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{11}{\mathrm{\&Sigma;}}}{(P_{i,j}-E\left(P\right))}^2,$ Wherein p _i,jfor image block P is at the pixel value at locus (i, j) place, given threshold value TH=220, if var (P)≤TH, then marking image block P is smooth piece, otherwise, be labeled as Non-smooth surface block.

If the image block P of 2.2 11 × 11 sizes is Non-smooth surface blocks, it is labeled as randomized block or block structure again.Mark criterion is: by having the difference operator computed image block P of cyclic boundary condition at locus (i, j) horizontal and vertical direction difference: as 1 < i < 11, during 1 < j < 11, image block P is at the difference D of the horizontal direction of locus (i, j) _x(P _i,j)=P _{i+1, j}-P _i,jwith the difference D of vertical direction _y(P _i,j)=P _{i, j+1}-P _i,j; Work as i=11, during j=11, image boundary need take the difference D in cyclic boundary condition calculated level direction _x(P _k,j)=P _k,j-P _{k-1, j}with the difference D of vertical direction _y(P _i,K)=P _i,K-P _{i, K-1}, then the gradient of image block P at locus (i, j) place is G _i,j=[D _x(P _i,j), D _y(P _i,j)], by the gradient G of all locus, be arranged in order left to bottom right from image block by row, obtain g=[g ₁, g ₂..., g _{11 × 11}] ^t=[G _1,1, G _2,1..., G _11,11] ^t.SVD decomposition is carried out to g, obtains g=USV ^t.First row v in V ₁master corresponding to image block gradient is dominant direction, secondary series v ₂corresponding to the secondary direction that is dominant of image block gradient, the singular value of its correspondence is respectively s ₁and s ₂.Calculated direction is dominant and measures R=(s ₁-s ₂)/0s ₁+ s ₂), given threshold value R ^*=0.05, if R≤R ^*, then marking image block P is randomized block, otherwise, be labeled as block structure.

If the image block P of 2.3 11 × 11 sizes is block structures, it is labeled as multi-direction block structure again.Mark criterion is: to be dominant direction v according to the master of image block P ₁, the directivity d=arctan (v of computing block ₁(2)/v ₁(1)).Interval [0, π] is divided into 6 parts, and it is interval which d belongs to, and is that direction structure block with regard to marking image block P.

If the image block P of 2.4 11 × 11 sizes is direction structure blocks, it is labeled as again direction edge block or direction texture block.Mark criterion is: the structural strength of computed image block P every bit wherein Ω _iit is the neighborhood of 3 × 3 sizes centered by current point i.Given threshold value if ST (i) > is τ, then thinks that i point is system point, be designated as M (i)=1, otherwise, be designated as M (i)=0.The number of all system points in statistical picture block P if M (P) < (11 × 11)/2, then marking image block P is direction edge block, otherwise, be labeled as direction texture block.Thus, can by all image blocks by directivity be structurally divided into 14 classes.

As shown in Figure 3, integrating step 3, step 4 and step 5, give principal component analysis and work in coordination with filtering process flow diagram.

Step 3: the principal component analysis of feature self-adaptation is carried out to the group image block that step 2 obtains.Concrete grammar is: to L group image block P ₁, P ₂..., P _t, calculate covariance matrix Σ, Σ of its principal component analysis (PCA), row two-dimension principal component analysis (R2DPCA) and row two-dimension principal component analysis (C2DPCA) respectively _rand Σ _c:

$\mathrm{\&Sigma;}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}(p_i-\stackrel{\&OverBar;}{p}){(p_i-\stackrel{\&OverBar;}{p})}^T,$ Wherein $\stackrel{\&OverBar;}{p}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{p}_i,$ P _ifor by P _ivector is pulled into by row, ${\mathrm{\&Sigma;}}_r=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{(p_i-\stackrel{\&OverBar;}{p})}^T(p_i-\stackrel{\&OverBar;}{p}),$ ${\mathrm{\&Sigma;}}_c=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}(p_i-\stackrel{\&OverBar;}{p}){(p_i-\stackrel{\&OverBar;}{p})}^T,$ Wherein $\stackrel{\&OverBar;}{p}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{p}_i.$

Covariance matrix is carried out Eigenvalues Decomposition respectively, and obtaining the corresponding eigenwert by descending sort is λ ₁>=λ ₂>=...>=λ _{k × K}, η ₁>=η ₂>=...>=η _kand ξ ₁>=ξ ₂>=...>=ξ _k.Calculate $\mathrm{SR}=\max \{{\mathrm{\&Sigma;}}_{i=1}^{11\&times;11}{\mathrm{\&lambda;}}_i^2,{\mathrm{\&Sigma;}}_{i=1}^{11}{\mathrm{\&eta;}}_i^2,{\mathrm{\&Sigma;}}_{i=1}^{11}{\mathrm{\&xi;}}_i^2\},$ If $\mathrm{SR}={\mathrm{\&Sigma;}}_{i=1}^{11\&times;11}{\mathrm{\&lambda;}}_i^2,$ Then PCA conversion is used to this group image block, if then R2DPCA conversion is used to this group image block, if then C2DPCA conversion is used to this group image block.

Step 4: to the feature adaptive transformation coefficient of the image block that step 3 obtains, carry out threshold value shrink process, threshold function table is $\mathrm{\&Psi;}\left(x\right)=\left\{\begin{array}{c}x,\left|x\right|\&GreaterEqual;\mathrm{HT}\\ 0,\left|x\right|<\mathrm{HT}\end{array}\right.,$ Thus obtain filter factor.Wherein threshold value HT=55.

Step 5: to the filter factor of the group image block that step 4 obtains, the inverse transformation of suitable transform selected by carry out step 3, thus the group image block obtaining filtering reconstruct.

Step 6: the group image block all filtering reconstructed is polymerized to complete image.Concrete grammar is: from the image upper left corner, for each pixel, find the image block of all filtering reconstruct comprising current pixel point, the current pixel point of all image blocks found is averaged, obtain the denoising result of current pixel point, by each mobile location of pixels of row, thus the filtering image (as Fig. 6 .e) of 256 × 256 last sizes can be obtained.

Shown in Fig. 4, Fig. 5, Fig. 6 and Fig. 7, the practicality of the inventive method is described by an image denoising example and effect assessment thereof.

In this embodiment, as shown in Fig. 4 .a, experimental image be a width clearly 256 × 256 standard testing image.As shown in Fig. 4 .b, add for correspondence the degraded image that variance is the white Gaussian noise of 20.

Contrast algorithm in experiment to comprise: based on the denoising method (BLS-GSM) of small echo, non-local mean method (NL Means), based on rarefaction representation dictionary learning method (K-SVD), and block-based principal component method (GP-PCA).

This programme embodiment is in MATLAB7.1 platform simulation the Realization of Simulation, and computing environment is Intel Pentium DCPU3.00GHz, the PC of internal memory 2G.In order to verify the validity of the inventive method, take the performance index of Y-PSNR (PSNR) and structural similarity (SSIM) qualitative assessment denoise algorithm.Generally speaking, PSNR is larger, and SSIM, more close to 1, represents that the denoising effect of algorithm is better.

For the picture rich in detail shown in Fig. 4 and degraded image, Fig. 5 .a and 5.b sets forth the Clustering Effect figure of the multi-direction morphosis grouping of picture rich in detail and degraded image, and wherein same color representative belongs to organizing picture material.Fig. 6 gives the denoising result image of above-mentioned contrast algorithm and the inventive method, and Fig. 7 gives the blown-up partial detail corresponding to Fig. 6.

Can see, degraded image is very serious by noise, particularly at texture parts such as tablecloths, has a strong impact on the visual analysis of people.As Fig. 6 .a(BLS-GSM algorithm), 6.c(K-SVD algorithm), 6.d(GP-PCA algorithm) and 7.a(BLS-GSM algorithm), 7.c(K-SVD algorithm), 7.d(GP-PCA algorithm) shown in, these three kinds of methods can be good at removing noise, but also eliminate grain details in large quantities simultaneously.And as shown in Fig. 6 .b and 7.b, although NL Means method can retain most grain details, remain with much noise equally, do not reach the object of effective denoising.The inventive method is as shown in Fig. 6 .e and 7.e, and can be good at equilibrium noise and remove and Acacia crassicarpaA, the most clear in the Detail contrast algorithm of denoising image, visual quality is the highest.

Table 1 gives the performance index situation of inventive method and contrast algorithm herein.

From denoising picture quality objective evaluation index, the objective evaluation index of the inventive method denoising picture quality is further improved.Such as, PSNR improves more than 0.5dB, SSIM and improves more than 0.02.

Table 1: different denoise algorithm performance index comparative result

Method	PSNR(dB)	SSIM
			Degraded image	22.13	0.4704
BLS-GSM algorithm	29.42	0.8279
			NL Means algorithm	29.76	0.8377
K-SVD algorithm	29.63	0.8379
			GP-PCA algorithm	29.68	0.8350
The inventive method	30.31	0.8627

In sum, no matter from visual perception or from objective evaluation index, the inventive method all achieves good effect, and denoising result is obvious, and grain details keeps significantly, having good application prospect and value.

Claims (6)

1. a filtering method is worked in coordination with in the principal component analysis of the multi-direction morphosis grouping of image, it is characterized in that step is as follows:

Step 1: input pixel size is the noise image of M × N, overlap partition is carried out to it, namely from the image upper left corner, extract the image block P that pixel size is K × K successively, by each mobile location of pixels of row, finally can obtain (M-K+1) × (N-K+1) individual size is the image block of K × K;

Step 2: to all image blocks obtained, carries out multi-direction morphosis grouping according to the variance of image block, gradient and singular value information to it, thus can obtain smooth piece of group, randomized block group, direction edge block group and direction texture block group;

Step 3: to the group image block obtained, carries out the principal component analysis of feature self-adaptation, thus takes suitable conversion, comprises principal component analysis conversion, the conversion of row two-dimension principal component analysis and the conversion of row two-dimension principal component analysis, obtains feature adaptive transformation coefficient;

Step 4: to the feature adaptive transformation coefficient of the group image block obtained, carry out threshold value contraction, threshold function table is $\mathrm{\&Psi;}\left(x\right)=\left\{\begin{array}{c}x,\left|x\right|\&GreaterEqual;\mathrm{HT}\\ 0,\left|x\right|<\mathrm{HT}\end{array}\right.$ Wherein HT is threshold value, thus obtains filter factor;

Step 5: to the filter factor of the group image block obtained, the inverse transformation of suitable transform selected by carry out step 3, thus the group image block obtaining filtering reconstruct;

Step 6: to the image block of all filtering reconstruct obtained, be averaged polymerization, generate full width filtering image, namely from the image upper left corner, for each pixel, the image block of all filtering reconstruct comprising current pixel point is found, the current pixel point of all image blocks found is averaged, obtain the denoising result of current pixel point, by each mobile location of pixels of row, thus full width filtering image can be obtained.

2. work in coordination with filtering method according to the principal component analysis of the multi-direction morphosis grouping of image according to claim 1, it is characterized in that, the concrete grammar of the multi-direction morphosis grouping of image block of step 2 is:

The image block P of 2.1 pairs of each K × K sizes, is labeled as smooth piece or Non-smooth surface block, and mark criterion is: the variance of computed image block P $\mathrm{var}\left(P\right)=\frac{1}{K\&times;K}\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{K}{\mathrm{\&Sigma;}}}{(P_{i,j}-E\left(P\right))}^2,$ Wherein $E\left(P\right)=\frac{1}{K\&times;K}\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{K}{\mathrm{\&Sigma;}}}{P}_{i,j},$ P _i,jfor image block P is at the pixel value at locus (i, j) place, given threshold value TH, if var (P)≤TH, then marking image block P is smooth piece, otherwise, be labeled as Non-smooth surface block;

If the image block P of 2.2 K × K sizes is Non-smooth surface blocks, it is labeled as randomized block or block structure again, mark criterion is: by having the difference operator computed image block P of cyclic boundary condition at locus (i, j) horizontal and vertical direction difference: work as 1<i<K, during 1<j<K, image block P is at the difference D of the horizontal direction of locus (i, j) _x(P _i,j)=P _{i+1, j}-P _i,jwith the difference D of vertical direction _y(P _i,j)=P _{i, j+1}-P _i,j; Work as i=K, during j=K, image boundary need take the difference D in cyclic boundary condition calculated level direction _x(P _k,j)=P _k,j-P _{k-1, j}with the difference D of vertical direction _y(P _i,K)=P _i,K-P _{i, K-1}, then the gradient of image block P at locus (i, j) place is G _i,j=[D _x(P _i,j), D _y(P _i,j)], by the gradient G of all locus, be arranged in order left to bottom right from image block by row, obtain g=[g ₁, g ₂..., g _{k × K}] ^t=[G _1,1, G _2,1..., G _k,K] ^t, SVD decomposition is carried out to g, obtains g=USV ^t, the first row v in V ₁master corresponding to image block gradient is dominant direction, secondary series v ₂corresponding to the secondary direction that is dominant of image block gradient, the singular value of its correspondence is respectively s ₁and s ₂, calculated direction is dominant and measures R=(s ₁-s ₂)/(s ₁+ s ₂), given threshold value R ^*if, R≤R ^*, then marking image block P is randomized block, otherwise, be labeled as block structure;

If the image block P of 2.3 K × K sizes is block structures, it is labeled as multi-direction block structure again, mark criterion is: to be dominant direction v according to the master of image block P ₁, the directivity d=arctan (v of computing block ₁(2)/v ₁(1)), interval [0, π] is divided into W part, it is interval which d belongs to, and is that direction structure block with regard to marking image block P;

If the image block P of 2.4 K × K sizes is direction structure blocks, it is labeled as again direction edge block or direction texture block, mark criterion is: the structural strength of computed image block P every bit wherein Ω _ibe the neighborhood of the z × z size centered by current point i, given threshold tau, if ST (i) > is τ, then think that i point is system point, be designated as M (i)=1, otherwise, be designated as M (i)=0, the number of all system points in statistical picture block P if M (P) < (K × K)/2, then marking image block P is direction edge block, otherwise, be labeled as direction texture block.

3. work in coordination with filtering method according to the principal component analysis of the multi-direction morphosis grouping of image according to claim 2, it is characterized in that, TH value is K × σ ², R ^*span be [0.01,0.2], the span of W is [4,12], and to be the value of 3, τ be for the value of z wherein σ ²for noise variance, ST (i) is the structural strength of i-th.

4. work in coordination with filtering method according to the principal component analysis of image according to claim 1 multi-direction morphosis grouping, it is characterized in that, the concrete grammar that the group image block of step 3 carries out the principal component analysis of feature self-adaptation is: to L group image block P ₁, P ₂..., P _t, calculate covariance matrix Σ, Σ of its principal component analysis, row two-dimension principal component analysis and row two-dimension principal component analysis respectively _rand Σ _c:

$\mathrm{\&Sigma;}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}(p_i-\stackrel{\&OverBar;}{p}){(p_i-\stackrel{\&OverBar;}{p})}^T,$ Wherein $\stackrel{\&OverBar;}{p}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{p}_i,$ P _ifor by P _ivector is pulled into by row,

${\mathrm{\&Sigma;}}_r=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{(P_i-\stackrel{\&OverBar;}{P})}^T(P_i-\stackrel{\&OverBar;}{P}),$ ${\mathrm{\&Sigma;}}_c=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}(P_i-\stackrel{\&OverBar;}{P}){(P_i-\stackrel{\&OverBar;}{P})}^T,$ Wherein $\stackrel{\&OverBar;}{P}=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{P}_i,$

Covariance matrix is carried out Eigenvalues Decomposition respectively, and obtaining the corresponding eigenwert by descending sort is λ ₁>=λ ₂>=...>=λ _{k × K}, η ₁>=η ₂>=...>=η _kand ξ ₁>=ξ ₂>=...>=ξ _k, calculate $\mathrm{SR}=\max \{{\mathrm{\&Sigma;}}_{i=1}^{K\&times;K}{\mathrm{\&lambda;}}_i^2,{\mathrm{\&Sigma;}}_{i=1}^K{\mathrm{\&eta;}}_i^2,{\mathrm{\&Sigma;}}_{i=1}^K{\mathrm{\&xi;}}_i^2\},$ If $\mathrm{SR}={\mathrm{\&Sigma;}}_{i=1}^{K\&times;K}{\mathrm{\&lambda;}}_i^2,$ Then PCA conversion is used to this group image block, if then R2DPCA conversion is used to this group image block, if then C2DPCA conversion is used to this group image block.

5. work in coordination with filtering method according to the principal component analysis of the multi-direction morphosis grouping of image according to claim 1, it is characterized in that, the span of the K of step 1 is [7,15].

6. work in coordination with filtering method according to the principal component analysis of the multi-direction morphosis grouping of image according to claim 1, it is characterized in that, the value of the threshold value HT of step 4 is 2.75 × σ ², wherein σ ²for noise variance.