CN103208097A - Principal component analysis collaborative filtering method for image multi-direction morphological structure grouping - Google Patents

Abstract

The invention discloses a principal component analysis collaborative filtering method for image multi-direction morphological structure grouping. The principal component analysis collaborative filtering method for the image multi-direction morphological structure grouping includes subjecting an image to overlapping partition, and subjecting image blocks to multi-direction morphological structure grouping according to variance, gradient and singular value information of the image blocks to obtain a smooth block group, a random block group, a direction edge block group and a direction texture block group; then subjecting the grouped image blocks to characteristic adaptive principal component analysis, utilizing hard threshold shrink method to subject image block transformation coefficients to filtering and then to grouping reconfiguration; and finally aggregating the grouping reconfiguration blocks into a full-size filtering image. The principal component analysis collaborative filtering method for the image multi-direction morphological structure grouping fully considers the multi-direction morphological structure of image blocks and non local similarity information of the image, has good structure and texture maintaining characteristics in image filtering process, is strong in noise elimination capacity, and is widely applicable to the image characteristic extraction and preprocessing process of target detection.

Description

The collaborative filtering method of principal component analysis of the multi-direction morphosis grouping of image

Technical field

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

Background technology

Digital picture obtain with transmission course in, inevitably can introduce various noises.The existence of noise, not only destroyed the real information of image, but also have a strong impact on the visual effect of image, make that therefore noise image is carried out filtering has just become a kind of effective image pre-processing method to the post-processed of image such as feature extraction and target detection etc. the difficulty that becomes.Image filtering method can be divided into the method for spatial domain and frequency domain.The spatial domain method is directly handled image pixel, as mean filter, medium filtering and the bilateral filtering etc. of classics.Frequency domain method transforms to frequency field with image earlier, then conversion coefficient is handled, and last another mistake transforms to the spatial domain to reach the purpose of filtering, as wavelet transformation, multi-scale geometric analysis and rarefaction representation etc.

Based on the importance of image filtering, a lot of scholars have carried out a large amount of correlative study work.People such as Wang Hongmei have invented a kind of self-adapting method for filtering image (patent No.: 200610043000.5) that keeps the edge.It is noise or edge that this method at first adopts coefficient correlation method marked pixels, according to label information, utilize adaptive neighborhood to shrink wavelet coefficient, in denoising simultaneously, have certain edge hold facility, but do not take full advantage of non local similarity and the local otherness of picture material.People such as Wang Xigui have invented a kind of image de-noising method (patent No.: 200610065679.8) of process of multi-template mixed filtering.This method is one group of wave filter of definition earlier, then to image block, has considered the otherness of image block, according to its even matter degree, select some wave filters that image block is carried out denoising, can the retaining part image detail information, but do not take into full account the non local similarity of image block.People such as Liu Fang have invented a kind of non-local mean image de-noising method (number of patent application: 201110091450.2) of integrated structure information.This method utilizes primal sketch to extract picture structure earlier, and image is divided into structural area peace skating area, adopts the non-local mean method to image denoising then respectively.These methods have improved the effect of image denoising to a certain extent, but all do not consider morphosis and multidirectional that picture material has, thereby can't accomplish equilibrium preferably aspect noise remove and the details maintenance.

Summary of the invention

The object of the invention is to provide the collaborative filtering method of principal component analysis of the multi-direction morphosis grouping of a kind of image, the collaborative filtering of the image block by equidirectional isostructure characteristic, take full advantage of the non local similarity information of image, in filtering noise, better keep detailed information such as edge of image and texture, improve the visual effect of image denoising.

The technical solution that realizes the object of the invention is: the collaborative filtering method of principal component analysis of the multi-direction morphosis grouping of a kind of image, and step is as follows:

Step 1: the input pixel size is the noise image of M * N, it is carried out overlapping piecemeal, and namely from the image upper left corner, extracting pixel size successively is the image block P of K * K, by the 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 images piece that obtains, according to variance, gradient and the singular value information of image block it is carried out multi-direction morphosis grouping, thereby can obtain smooth group, randomized block group, direction edge block group and direction texture block group;

Step 3: to the group image piece that obtains, carry out the principal component analysis of feature self-adaptation, thereby take suitable conversion, comprise principal component analysis conversion, row two-dimension principal component analysis conversion and the conversion of row two-dimension principal component analysis, obtain feature adaptive transformation coefficient;

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

Step 5: the filter factor to the group image piece that obtains, carry out the inverse transformation of the selected suitable conversion of step 3, thereby obtain the group image piece of filtering reconstruct;

Step 6: to the image block of all filtering reconstruct of obtaining, average polymerization, generate the full width filtering image, namely from the image upper left corner, at each pixel, seek the image block of all filtering reconstruct that comprise current pixel point, current pixel point to all images piece that finds averages, obtain the denoising result of current pixel point, by the each mobile location of pixels of row, thereby can obtain the full width filtering image.

The present invention compared with prior art, its remarkable advantage: (1) the present invention has provided a kind of group technology of image block, this method has not only been considered non local similarity and local aggregation between image, also utilize different shape structure and the directivity difference of image block, provided more effective and reasonable packet mode.(2) the present invention organizes image to difference, considers the otherness of its content and feature, and filtering is worked in coordination with in use characteristic self-adaptation principal component analysis conversion, thereby can obtain better denoising effect.(3) the present invention has well kept simultaneously the main visual experience part of image in denoising, particularly aspect edge and grain details, thereby make subsequent characteristics extract and target detection etc. task is easier carries out.

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

Description of drawings

Fig. 1 is the collaborative filtering method process flow diagram of 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 among the present invention.

Fig. 3 is the collaborative filtering process flow diagram of the principal component analysis of the group image piece among the present invention.

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

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

Fig. 6 is a denoising experiment contrast effect figure of the inventive method and classic method.

Fig. 7 is that the texture during a denoising of the inventive method and classic method is tested keeps the effect contrast figure.

Embodiment

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

Step 1: the input pixel size is the noise image of M * N, it is carried out overlapping piecemeal, concrete grammar is: from the image upper left corner, extracting pixel size successively is the image block P of K * K, by the 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: all images piece that step 1 is obtained carries out multi-direction morphosis grouping, i.e. all images piece to obtaining, variance, gradient and singular value information according to image block are carried out multi-direction morphosis grouping to it, thereby can obtain smooth group, randomized block group, direction edge block group and direction texture block group.Concrete grammar is:

2.1 to the image block P of each K * K size, it is labeled as smooth or non-smooth.The mark criterion is: the variance of computed image piece 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, j}For image block P in the locus (if var (P)≤TH, then marking image piece P is smooth for i, the pixel value of j) locating, otherwise, be labeled as non-smooth.Wherein TH is threshold parameter, and value is K * σ ², σ ²Be noise variance.

If 2.2 smooth of the image block P right and wrong of K * K size are labeled as randomized block or block structure again with it.The mark criterion is: (i, level j) and vertical direction difference: as 1＜i＜K, during 1＜j＜K, image block P is (i, the difference D of horizontal direction j) in the locus in the locus for the difference operator computed image piece P by having cyclic boundary condition _x(P _{I, j})=P _{I+1, j}-P _{I, j}Difference D with vertical direction _y(P _{I, j})=P _{I, j+1}-P _{I, j}Work as i=K, during j=K, image boundary need be taked the difference D of cyclic boundary condition calculated level direction _x(P _{K, j})=P _{K, j}-P _{K-1, j}Difference D with vertical direction _y(P _{I, K})=P _{I, K}-P _{I, K-1}, then (i, the gradient of j) locating is G to image block P in the locus _{I, j}=[D _x(P _{I, j}), D _y(P _{I, j})], with 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}] ^TG is carried out SVD decompose, obtain g=USV ^TAmong the V first is listed as v ₁Corresponding to the master of the image block gradient direction that is dominant, secondary series v ₂Corresponding to the inferior direction that is dominant of image block gradient, its corresponding singular value is respectively s ₁And s ₂Calculated direction is dominant and measures R=(s ₁-s ₂)/(s ₁+ s ₂), given threshold value R ^*If, R≤R ^*, then marking image piece P is randomized block, otherwise, be labeled as block structure.R wherein ^*Span is [0.01,0.2].

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

2.4 if the image block P of K * K size is the direction structure piece, it is labeled as direction edge block or direction texture block again.The mark criterion is: the structural strength of computed image piece P every bit

Ω wherein _iIt is the neighborhood of the z * z size centered by current some i.Given threshold tau if ST (i)＞τ thinks that then the i point is system point, is designated as M (i)=1, otherwise, be designated as M (i)=0.The number of all system points among the statistical picture piece P

If M (P)＜(K * K)/2, then marking image piece P is the direction edge block, otherwise, be labeled as the direction texture block.Wherein the value of z is that 3, τ value is $0.01\&times;\max {\left\{\mathrm{ST}\left(i\right)\right\}}_1^{M\&times;N}.$

Step 3: the group image piece that step 2 is obtained carries out the principal component analysis of feature self-adaptation, to the group image piece that obtains, carry out the principal component analysis of feature self-adaptation, thereby 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 piece P ₁, P ₂..., P _T, calculate covariance matrix Σ, the 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 with P _iPull into vector 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 eigenwert respectively decompose, 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 _KCalculate $\mathrm{SR}=\max \{{\mathrm{\&Sigma;}}_{i=1}^{K\&times;K}{\mathrm{\&lambda;}}_{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="HDA00002789220800021.png"\; state="\{\{state\}\}">i</figure-callout>}}^2,{\mathrm{\&Sigma;}}_{i=1}^K{\mathrm{\&eta;}}_{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="HDA00002789220800021.png"\; state="\{\{state\}\}">i</figure-callout>}}^2,{\mathrm{\&Sigma;}}_{i=1}^K{\mathrm{\&xi;}}_i^2\},$ If $\mathrm{SR}={\mathrm{\&Sigma;}}_{i=1}^{K\&times;K}{\mathrm{\&lambda;}}_{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="HDA00002789220800021.png"\; state="\{\{state\}\}">i</figure-callout>}}^2,$ Then this group image block is used the PCA conversion, if Then this group image block is used the R2DPCA conversion, if

Then this group image block is used the C2DPCA conversion.

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

Step 5: the filter factor of the group image piece that step 4 is obtained, carry out the inverse transformation of the selected suitable conversion of step 3, thereby obtain the group image piece of filtering reconstruct.

Step 6: the image block to all filtering reconstruct of obtaining, average polymerization, generate the full width filtering image.Concrete grammar is: from the image upper left corner, at each pixel, searching comprises the image block of all filtering reconstruct of current pixel point, current pixel point to all images piece that finds averages, obtain the denoising result of current pixel point, by the each mobile location of pixels of row, thereby can obtain the full width filtering image.

Embodiment

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

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

Original true picture is the barbara image shown in Fig. 4 .a, and size is 256 * 256, and gray level is 256, is σ to its variance that adds white Gaussian noise ²=20, obtain noise image such as Fig. 4 .b.

Step 2: all images piece that step 1 is obtained carries out multi-direction morphosis grouping.As shown in Figure 2, concrete grammar is:

2.1 to the image block P of each 11 * 11 size, it is labeled as smooth or non-smooth.The mark criterion is: the variance of computed image piece 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, j}For image block P in the locus (i, the pixel value of j) locating, given threshold value TH=220, if var (P)≤TH, then marking image piece P is smooth, otherwise, be labeled as non-smooth.

If 2.2 smooth of the image block P right and wrong of 11 * 11 sizes are labeled as randomized block or block structure again with it.The mark criterion is: (i, level j) and vertical direction difference: when 1＜i＜11,1＜j＜11 o'clock, image block P is (i, the difference D of horizontal direction j) in the locus in the locus for the difference operator computed image piece P by having cyclic boundary condition _x(P _{I, j})=P _{I+1, j}-P _{I, j}Difference D with vertical direction _y(P _{I, j})=P _{I, j+1}-P _{I, j}Work as i=11, during j=11, image boundary need be taked the difference D of cyclic boundary condition calculated level direction _x(P _{K, j})=P _{K, j}-P _{K-1, j}Difference D with vertical direction _y(P _{I, K})=P _{I, K}-P _{I, K-1}, then (i, the gradient of j) locating is G to image block P in the locus _{I, j}=[D _x(P _{I, j}), D _y(P _{I, j})], with 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] ^TG is carried out SVD decompose, obtain g=USV ^TAmong the V first is listed as v ₁Corresponding to the master of the image block gradient direction that is dominant, secondary series v ₂Corresponding to the inferior direction that is dominant of image block gradient, its corresponding singular value is respectively s ₁And s ₂Calculated 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 piece P is randomized block, otherwise, be labeled as block structure.

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

2.4 if the image block P of 11 * 11 sizes is direction structure pieces, it is labeled as direction edge block or direction texture block again.The mark criterion is: the structural strength of computed image piece P every bit

Ω wherein _iIt is the neighborhood of 3 * 3 sizes centered by current some i.Given threshold value

If ST (i)＞τ thinks that then the i point is system point, be designated as M (i)=1, otherwise, be designated as M (i)=0.The number of all system points among the statistical picture piece P If M (P)＜(11 * 11)/2, then marking image piece P is the direction edge block, otherwise, be labeled as the direction texture block.Thereby, can be with all image blocks by directivity and structural 14 classes that are divided into.

As shown in Figure 3, integrating step 3, step 4 and step 5 have provided the collaborative filtering process flow diagram of principal component analysis.

Step 3: the group image piece that step 2 is obtained carries out the principal component analysis of feature self-adaptation.Concrete grammar is: to L group image piece P ₁, P ₂..., P _T, calculate covariance matrix Σ, the 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 with P _iPull into vector 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 eigenwert respectively decompose, 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 _KCalculate $\mathrm{SR}=\max \{{\mathrm{\&Sigma;}}_{i=1}^{11\&times;11}{\mathrm{\&lambda;}}_{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="HDA00002789220800021.png"\; state="\{\{state\}\}">i</figure-callout>}}^2,{\mathrm{\&Sigma;}}_{i=1}^{11}{\mathrm{\&eta;}}_{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="HDA00002789220800021.png"\; state="\{\{state\}\}">i</figure-callout>}}^2,{\mathrm{\&Sigma;}}_{i=1}^{11}{\mathrm{\&xi;}}_i^2\},$ If $\mathrm{SR}={\mathrm{\&Sigma;}}_{i=1}^{11\&times;11}{\mathrm{\&lambda;}}_{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="HDA00002789220800021.png"\; state="\{\{state\}\}">i</figure-callout>}}^2,$ Then this group image block is used the PCA conversion, if

Then this group image block is used the R2DPCA conversion, if

Then this group image block is used the C2DPCA conversion.

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

Step 5: the filter factor of the group image piece that step 4 is obtained, carry out the inverse transformation of the selected suitable conversion of step 3, thereby obtain the group image piece of filtering reconstruct.

Step 6: the group image piece of all filtering reconstruct is polymerized to complete image.Concrete grammar is: from the image upper left corner, at each pixel, searching comprises the image block of all filtering reconstruct of current pixel point, current pixel point to all images piece that finds averages, obtain the denoising result of current pixel point, move a location of pixels by row are each, thereby can obtain the filtering image (as Fig. 6 .e) of 256 * 256 last sizes.

Below in conjunction with Fig. 4, Fig. 5, Fig. 6 and shown in Figure 7 illustrates the practicality of the inventive method by an image denoising example and effect assessment thereof.

In this embodiment, shown in Fig. 4 .a, experimental image is a width of cloth 256 * 256 standard testing image clearly.Shown in Fig. 4 .b, be the degraded image of 20 white Gaussian noise for the adding variance of correspondence.

The contrast algorithm comprises in the experiment: based on the denoising method (BLS-GSM) of small echo, and 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 D CPU3.00GHz, 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 more big, and SSIM is more close to 1, and the denoising effect of expression algorithm is more good.

At picture rich in detail shown in Figure 4 and degraded image, Fig. 5 .a and 5.b have provided the cluster design sketch of the multi-direction morphosis grouping of picture rich in detail and degraded image respectively, and wherein the same color representative belongs to picture material on the same group.Fig. 6 has provided the denoising result image of above-mentioned contrast algorithm and the inventive method, and Fig. 7 has provided the local detail enlarged drawing corresponding to Fig. 6.

Can see that degraded image is subjected to noise very serious, particularly at texture parts such as tablecloths, the vision that has a strong impact on people is understood.As Fig. 6 .a(BLS-GSM algorithm), the 6.c(K-SVD algorithm), the 6.d(GP-PCA algorithm) and the 7.a(BLS-GSM algorithm), 7.c(K-SVD algorithm), 7.d(GP-PCA algorithm), these three kinds of methods can be good at removing noise, but have also removed grain details in large quantities simultaneously.And shown in Fig. 6 .b and 7.b, though NL Means method can keep most grain details, remain with much noise equally, do not reach the purpose of effective denoising.The inventive method can be good at balanced noise remove and texture and keeps shown in Fig. 6 .e and 7.e, and the most clear in the details contrast algorithm of denoising image, visual quality is the highest.

Table 1 has provided the performance index situation of this paper inventive method and contrast algorithm.

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

Table 1: different denoise algorithm performance index comparative results

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

In sum, no matter from visual perception still from the objective evaluation index, the inventive method has all obtained effect preferably, denoising result is obvious, grain details keeps significantly having good application prospects and value.

Claims (6)

1. 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: the input pixel size is the noise image of M * N, it is carried out overlapping piecemeal, and namely from the image upper left corner, extracting pixel size successively is the image block P of K * K, by the 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 images piece that obtains, according to variance, gradient and the singular value information of image block it is carried out multi-direction morphosis grouping, thereby can obtain smooth group, randomized block group, direction edge block group and direction texture block group;

Step 3: to the group image piece that obtains, carry out the principal component analysis of feature self-adaptation, thereby take suitable conversion, comprise principal component analysis conversion, row two-dimension principal component analysis conversion and the conversion of row two-dimension principal component analysis, obtain feature adaptive transformation coefficient;

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

Step 5: the filter factor to the group image piece that obtains, carry out the inverse transformation of the selected suitable conversion of step 3, thereby obtain the group image piece of filtering reconstruct;

Step 6: to the image block of all filtering reconstruct of obtaining, average polymerization, generate the full width filtering image, namely from the image upper left corner, at each pixel, seek the image block of all filtering reconstruct that comprise current pixel point, current pixel point to all images piece that finds averages, obtain the denoising result of current pixel point, by the each mobile location of pixels of row, thereby can obtain the full width filtering image.

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

2.1 to the image block P of each K * K size, it is labeled as smooth or non-smooth, the mark criterion is: the variance of computed image piece 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, j}For image block P in the locus (i, the pixel value of j) locating, given threshold value TH, if var (P)≤TH, then marking image piece P is smooth, otherwise, be labeled as non-smooth;

If 2.2 smooth of the image block P right and wrong of K * K size, it is labeled as randomized block or block structure again, the mark criterion is: the difference operator computed image piece P by having cyclic boundary condition is at locus (i, j) level and vertical direction difference: as 1＜i＜K, during 1＜j＜K, image block P is (i, the difference D of horizontal direction j) in the locus _x(P _{I, j})=P _{I+1, j}-P _{I, j}Difference D with vertical direction _y(P _{I, j})=P _{I, j+1}-P _{I, j}Work as i=K, during j=K, image boundary need be taked the difference D of cyclic boundary condition calculated level direction _x(P _{K, j})=P _{K, j}-P _{K-1, j}Difference D with vertical direction _y(P _{I, K})=P _{I, K}-P _{I, K-1}, then (i, the gradient of j) locating is G to image block P in the locus _{I, j}=[D _x(P _{I, j}), D _y(P _{I, j})], with 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, g is carried out SVD decompose, obtain g=USV ^T, first among the V is listed as v ₁Corresponding to the master of the image block gradient direction that is dominant, secondary series v ₂Corresponding to the inferior direction that is dominant of image block gradient, its corresponding singular value 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 piece P is randomized block, otherwise, be labeled as block structure;

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

2.4 if the image block P of K * K size is the direction structure piece, it is labeled as direction edge block or direction texture block again, the mark criterion is: the structural strength of computed image piece P every bit

Ω wherein _iBe the neighborhood of the z * z size centered by current some i, given threshold tau if ST (i)＞τ thinks that then the i point is system point, is designated as M (i)=1, otherwise, be designated as M (i)=0, the number of all system points among the statistical picture piece P

If M (P)＜(K * K)/2, then marking image piece P is the direction edge block, otherwise, be labeled as the direction texture block.

3. according to the collaborative filtering method of principal component analysis of the multi-direction morphosis grouping of the described image of claim 2, it is characterized in that the TH value is K * σ ², R ^*Span be [0.01,0.2], the span of W is [4,12], the value of z is that the value of 3, τ is

σ wherein ²Be noise variance, ST (i) is the structural strength that i is ordered.

4. according to the collaborative filtering method of principal component analysis of the multi-direction morphosis grouping of the described image of claim 1, it is characterized in that the concrete grammar that the group image piece of step 3 carries out the principal component analysis of feature self-adaptation is: to L group image piece P ₁, P ₂..., P _T, calculate covariance matrix Σ, the 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 with P _iPull into vector 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 eigenwert respectively decompose, 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 this group image block is used the PCA conversion, if

Then this group image block is used the R2DPCA conversion, if

Then this group image block is used the C2DPCA conversion.

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

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

Images