CN100433046C - Image blind separation based on sparse change - Google Patents

Abstract

First, using Contourlet transform, the method carries out multidirectional multiscale sparse decomposition for received signal of mixed image; in Contourlet transformed domain, using discriminating criterion of sparsity to select group of sub image with best sparsity; then, using traditional quick analysis method of independent components of fixed point to carry out blind separation for selected group of sub image so as to obtain separation matrix; finally, using the separation matrix to carrying out separation for received signal of mixed image, the method picks up each independent components in mixed image so as to reach purpose of separating images from blind sources. Raising precision for separating images from blind sources, the invention is applicable to radio communication system, sonar, and radar system as well as audio, acoustics and medicinal signal processing.

Description

A kind of Image Blind source separation method based on sparse conversion

Technical field

The present invention relates to a kind of image denoising method, particularly a kind of Image Blind source separation method based on sparse conversion.The important use potentiality are all arranged in the Flame Image Process in military field or non-military field.

Background technology

Usually, image its obtain or transmission course in all can be subjected to the pollution of other signals, for follow-up further processing, the necessary separating treatment of carrying out.The purpose of separation of images is exactly each independent signal component that extracts as much as possible in the received signal, to improve the quality of image.At present, image denoising method mainly is divided into traditional filtering method and blind source separation method, and is wherein the most representative with blind source separation method.

Blind source separation method is under information source S and the equal condition of unknown of signal transmission feature, only carries out the separation of these separate source signals by the mixed signal X that receives.Now, main blind source separation method mainly contains independent component analysis method (ICA), the maximum entropy method (Infomax) based on gradient decline at random, the natural gradient learning method (NGA) based on high-order statistic and adopts the quick ICA method (FastICA) of negentropy criterion, i.e. fix-point method (Fixed-point).Though these methods are obtaining effect preferably aspect the separation of blind source,, they also are not best.

Studies show that the sparse property of input signal influences the performance of blind source separation method to a great extent.Sparse more when input signal, Image Blind source separating effect is good more.So, arise at the historic moment based on the Image Blind source separation method of wavelet transformation, improved separating effect to a great extent.But, the two-dimentional separable wavelets conversion that is formed by tensor product by the one dimension small echo can only represent effectively that the unusual information of one dimension promptly puts unusual information, and two dimension or the unusual information of higher-dimension in the image can not be described effectively, as important informations such as line, profiles, thereby restricted performance based on the Image Blind source separation method of wavelet transformation.Profile wavelet transformation Contourlet is as a kind of new signal analysis instrument, solved wavelet transformation and can not effectively represent the two dimension or the shortcoming of higher-dimension singularity more, can exactly the edge in the image have been captured in the subband of different scale, different frequency, different directions.It not only has the multiple dimensioned characteristic of wavelet transformation, also has directivity and anisotropy that wavelet transformation does not have, therefore can be advantageously applied in the Flame Image Process.

Summary of the invention

The objective of the invention is to deficiency at the existence of conventional images blind source separation method method, a kind of Image Blind source separation method based on sparse conversion has been proposed, this method is obtained separation matrix in profile wavelet transformation Contourlet, utilize this separation matrix to come the vision-mix signal that receives is separated, extract each isolated component in the vision-mix, reach the purpose that separate in the Image Blind source.

In order to achieve the above object, the present invention adopts following technical proposals:

A kind of Image Blind source separation method based on sparse conversion.It is characterized in that at first utilizing profile wavelet transformation Contourlet that the vision-mix signal that receives is carried out multiple dimensioned, multidirectional Sparse Decomposition, and utilize sparse property discrimination standard to choose the best subimage group of sparse property in profile wavelet transformation Contourlet territory; Utilize traditional quick fixed point independent component analysis method that the subimage group of choosing is carried out blind separation then, obtain separation matrix; At last, utilize this separation matrix to come the vision-mix signal that receives is separated, extract each isolated component in the vision-mix, reach the purpose that separate in the Image Blind source.

The concrete steps of above-mentioned Image Blind source separation method are as follows:

1. initialization setting.Set the direction Number of Decomposition L in the middle Laplce tower decomposition LP number of plies K of profile wavelet transformation Contourlet and every layer _k

2. vision-mix X1 and the X2 that receives carried out multiple dimensioned, multidirectional profile wavelet transformation Contourlet Sparse Decomposition respectively, promptly

$[{X1}_{\mathrm{lf}},{X1}_{\mathrm{hf}}^{\left(\mathrm{1,1}\right)},\·\·\·,{X1}_{\mathrm{hf}}^{(1,L_1)},{X1}_{\mathrm{hf}}^{\left(\mathrm{2,1}\right)},\·\·\·,{X1}_{\mathrm{hf}}^{(K,L_k)}]=T\left(X1\right),---\left(1\right)$

$[{X2}_{\mathrm{lf}},{X2}_{\mathrm{hf}}^{\left(\mathrm{1,1}\right)},\·\·\·,{X2}_{\mathrm{hf}}^{(1,L_1)},{X2}_{\mathrm{hf}}^{\left(\mathrm{2,1}\right)},\·\·\·,{X2}_{\mathrm{hf}}^{(K,L_k)}]=T\left(X2\right),---\left(2\right)$

Wherein T () is profile wavelet transformation Contourlet, thereby obtains a width of cloth low frequency subgraph as Xi _LfWith a series of high frequency subimage Xi with different resolution _Hf ^{(k, l)}, wherein k ∈ (1, K) and l ∈ (1, L _k) indicating that subimage is positioned at the l direction of the tower decomposition of k layer Laplce LP, i represents 1 or 2;

3. according to sparse property criterion, choose the high frequency subimage group X1 behind the profile wavelet transformation Contourlet _Hf ^{(k, l)}And X2 _Hf ^{(k, l)}In the most sparse subimage group, be designated as X1 _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}This method is carried out the judgement of sparse property according to the star chart distribution and the clustering method of subimage group;

4. to the high frequency subimage group X1 that 3. obtains in the step _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}, adopt traditional natural gradient learning method NGA to carry out blind source and separate, obtain separation matrix W, promptly

$W=\mathrm{NGA}({X1}_{\mathrm{hf}}^{(\mathrm{ksel},\mathrm{lsel})},{X2}_{\mathrm{hf}}^{(\mathrm{ksel},\mathrm{lsel})}),---\left(3\right)$

Wherein, NGA () represents natural gradient learning method NGA;

5. utilize the W that 4. obtains in the step to separate the mixed signal that receives, obtaining isolated component Y1 and Y2 has

$\left[\begin{array}{c}Y1\\ Y2\end{array}\right]=W\left[\begin{array}{c}X1\\ X2\end{array}\right].---\left(4\right)$

Separating resulting Y1 that obtains and Y2 are the estimation of the original signal of separating;

Above-mentioned sparse property criterion be based on choose the star chart of subimage group distribute and clustering method carries out.Concrete estimating step is:

(a) order $Z_{k,l}={[{X1}_{\mathrm{hf}}^{(k,l)},{X2}_{\mathrm{hf}}^{(k,l)}]}^T,$ Wherein k ∈ (1, K) and l ∈ (1, L _k);

(b) remove less coefficient component in the signal, to eliminate The noise;

(c) all data points are projected on the unit sphere, i.e. Z _{K, l}=Z _{K, l}/ || Z _{K, l}||;

(d) all signaling points are moved on to positive hemisphere face: if first coordinate of data point $z_{k,l}^1<0,$ Z _{K, l}=-Z _{K, l}Otherwise, Z _{K, l}=Z _{K, l}

(e) determine poly-axle and poly-axle center by clustering algorithm;

(f) calculate all data points to gathering the distance and the D of axle recently from self _{K, l}, and weigh sparse property, D with this _{K, l}More little, sparse more, the poly-axle in its star chart is just clear more;

(g) to all Z _k, (for k=1 ..., N _L) calculating D _{K, l}, seek its minimum value, order $(\mathrm{ksel},\mathrm{lsel})=\underset{k,l}{\arg \min }\left(D_{k,l}\right);$

(h) therefore, the most sparse subimage group is X1 _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}

The inventive method has following conspicuous outstanding substantive distinguishing features and remarkable advantage compared with prior art:

Image Blind source separation method based on sparse conversion provided by the invention is at first to utilize profile wavelet transformation Contourlet that the vision-mix signal that receives is carried out multiple dimensioned, multidirectional Sparse Decomposition, and utilizes sparse property discrimination standard to choose the best subimage group of sparse property in profile wavelet transformation Contourlet territory; Utilize traditional quick fixed point independent component analysis method that the subimage group of choosing is carried out blind separation then, obtain separation matrix; At last, utilize this separation matrix to come the vision-mix signal that receives is separated, extract each isolated component in the vision-mix, reach the purpose that separate in the Image Blind source.

Concrete characteristics and advantage are:

(1) at two or higher-dimension singularity in the presentation video of the shortcoming of wavelet transformation in the most representative existing wavelet field threshold value noise-reduction method-effectively, Contourlet is applied in the image noise reduction with the profile wavelet transformation, carry out multiple dimensioned, multi-direction decomposition, for follow-up noise reduction process provides sparse iamge description coefficient.

(2) deficiency that exists at conventional images blind source separation method method has proposed the Image Blind source separation method based on sparse conversion.

(3) signal that receives is utilized profile wavelet transformation Contourlet carry out Sparse Decomposition, under sparse condition, carry out blind source and separate, improve the separation accuracy in Image Blind source, improve the effect of separate picture.

(4) star chart distribution and the clustering method according to the subimage group carries out the judgement of sparse property, chooses the best subimage group of sparse property, obtains separation matrix W accurately.

Provided by the inventionly can improve the separation accuracy in Image Blind source, reach desirable separation of images effect based on the Image Blind source separation method of sparse conversion.In radio communications system, sonar and radar system, audio frequency and the acoustics in military field or non-military field and medical signals are handled, the important use potentiality are arranged all.

Description of drawings

Fig. 1 is the Image Blind source separation method block diagram based on sparse conversion of one embodiment of the invention.

Fig. 2 is the blind source of Fig. 1 example separating resulting photo figure.Among the figure, (a) and (b) two width of cloth vision-mix for receiving (c) and (d) be the separating resulting based on the Image Blind source separation method of wavelet transformation, (e) and (f) are the separating resulting of employing the inventive method.

Embodiment

A preferred embodiment of the present invention is auspicious in conjunction with the accompanying drawings state as follows:

This is based on the Image Blind source separation method of sparse conversion, as shown in Figure 1.At first utilize profile wavelet transformation Contourlet that the vision-mix signal that receives is carried out multiple dimensioned, multidirectional Sparse Decomposition, and utilize sparse property discrimination standard to choose the best subimage group of sparse property in profile wavelet transformation Contourlet territory; Utilize traditional quick fixed point independent component analysis method that the subimage group of choosing is carried out blind separation then, obtain separation matrix; At last, utilize this separation matrix to come the vision-mix signal that receives is separated, extract each isolated component in the vision-mix, reach the purpose that separate in the Image Blind source.

Concrete steps are:

1. initialization setting.The tower decomposition of the middle Laplce LP that sets profile wavelet transformation Contourlet decomposes the direction Number of Decomposition L in number of plies K and every layer _k

2. vision-mix X1 and the X2 that receives carried out multiple dimensioned, multidirectional profile wavelet transformation Contourlet Sparse Decomposition respectively, promptly

$[{X1}_{\mathrm{lf}},{X1}_{\mathrm{hf}}^{\left(\mathrm{1,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X1}_{\mathrm{hf}}^{(1,L_1)},{X1}_{\mathrm{hf}}^{\left(\mathrm{2,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X1}_{\mathrm{hf}}^{(K,L_k)}]=T\left(X1\right),$

$[{X2}_{\mathrm{lf}},{X2}_{\mathrm{hf}}^{\left(\mathrm{1,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X2}_{\mathrm{hf}}^{(1,L_1)},{X2}_{\mathrm{hf}}^{\left(\mathrm{2,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X2}_{\mathrm{hf}}^{(K,L_k)}]=T\left(X2\right),$

Wherein T () is profile wavelet transformation Contourlet, thereby obtains a width of cloth low frequency subgraph as Xi _LfWith a series of high frequency subimage Xi with different resolution _Hf ^{(k, l)}, wherein k ∈ (1, K) and l ∈ (1, L _k) indicating that subimage is positioned at the l direction of the tower decomposition of k layer Laplce LP, i represents 1 or 2;

3. according to sparse property criterion, choose the high frequency subimage group X1 behind the profile wavelet transformation Contourlet _Hf ^{(k, l)}And X2 _Hf ^{(k, l)}In the most sparse subimage group, be designated as X1 _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}This method is carried out the judgement of sparse property according to the star chart distribution and the clustering method of subimage group, and concrete grammar is as follows:

A. order $Z_{k,l}={[{X1}_{\mathrm{hf}}^{(k,l)},{X2}_{\mathrm{hf}}^{(k,l)}]}^T,$ Wherein k ∈ (1, K) and l ∈ (1, L _k);

B. remove coefficient component less in the signal, to eliminate The noise;

C. all data points are projected on the unit sphere, i.e. Z _{K, l}=Z _{K, l}/ || Z _{K, l}||;

D. all signaling points are moved on to positive hemisphere face: if first coordinate of data point $z_{k,l}^1<0,$ Z _{K, l}=-Z _{K, l}Otherwise, Z _{K, l}=Z _{K, l}

E. determine poly-axle and poly-axle center by clustering algorithm;

F. calculate all data points to gathering the distance and the D of axle recently from self _{K, l}, and weigh sparse property, D with this _{K, l}More little, sparse more, the poly-axle in its star chart is just clear more;

G. to all Z _k, (for k=1 ..., N _L) calculating D _{K, l}, seek its minimum value, order $(\mathrm{ksel},\mathrm{lsel})=\underset{k,l}{\arg \min }\left(D_{k,l}\right);$

H. therefore, the most sparse subimage group is X1 _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}

4. to the high frequency subimage group X1 that 3. obtains in the step _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}, adopt traditional natural gradient learning method NGA to carry out blind source and separate, obtain separation matrix W, promptly

$W=\mathrm{NGA}({X1}_{\mathrm{hf}}^{(\mathrm{ksel},\mathrm{lsel})},{X2}_{\mathrm{hf}}^{(\mathrm{ksel},\mathrm{lsel})}),$

Wherein, NGA () represents natural gradient learning method NGA;

5. utilize the W that 4. obtains in the step to separate the mixed signal that receives, obtaining isolated component Y1 and Y2 has

$\left[\begin{array}{c}Y1\\ Y2\end{array}\right]=W\left[\begin{array}{c}X1\\ X2\end{array}\right].$

Separating resulting Y1 that obtains and Y2 are the estimation of the original signal of separating;

As can be seen from Figure 2, compare best Image Blind source separation method at present based on wavelet transformation, this Image Blind source separation method can separate the independent image component in the received signal better, further improves the separation accuracy in Image Blind source, improves the effect of separate picture.

Table 1 has provided the objective evaluation index of Image Blind of the present invention source separating resulting.

Adopt Y-PSNR (PSNR) to weigh the quality of noise reduction image in the table, and then estimated the quality of Image Blind source separation method of the present invention.

Also can draw same conclusion from table, this Image Blind source separation method can separate the independent image component in the received signal better, further improves the separation accuracy in Image Blind source, improves the effect of separate picture.

In a word, no matter be from the human eye vision effect, still from the objective evaluation index, show that all the inventive method has higher separation accuracy, better separating effect.

The objective evaluation index of table 1 separating effect

	Separate picture 1	Separate picture 2
			Image Blind source separation method based on wavelet transformation	37.2593	58.2365
The inventive method	48.5709	51.0370

Claims (3)

1, a kind of Image Blind source separation method based on sparse conversion, it is characterized in that at first utilizing profile wavelet transformation Contourlet that the vision-mix signal that receives is carried out multiple dimensioned, multidirectional Sparse Decomposition, and utilize sparse property discrimination standard to choose the best subimage group of sparse property in profile wavelet transformation Contourlet territory; Utilize traditional quick fixed point independent component analysis method that the subimage group of choosing is carried out blind separation then, obtain separation matrix; At last, utilize this separation matrix to come the vision-mix signal that receives is separated, extract each isolated component in the vision-mix, reach the purpose that separate in the Image Blind source.

2, the Image Blind source separation method based on sparse conversion according to claim 1 is characterized in that the concrete operations step is:

1.. the initialization setting, set the direction Number of Decomposition L in the middle Laplce tower decomposition LP number of plies K of profile wavelet transformation Contourlet and every layer _k

2.. vision-mix X1 and the X2 that receives carried out multiple dimensioned, multidirectional profile wavelet transformation Contourlet Sparse Decomposition respectively, promptly

$[{X1}_{\mathrm{lf}},{X1}_{\mathrm{hf}}^{\left(\mathrm{1,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X1}_{\mathrm{hf}}^{(1,L_1)},{X1}_{\mathrm{hf}}^{\left(\mathrm{2,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X1}_{\mathrm{hf}}^{(K,L_k)}]=T\left(X1\right),$

$[{X2}_{\mathrm{lf}},{X2}_{\mathrm{hf}}^{\left(\mathrm{1,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X2}_{\mathrm{hf}}^{(1,L_1)},{X2}_{\mathrm{hf}}^{\left(\mathrm{2,1}\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{X2}_{\mathrm{hf}}^{(K,L_k)}]=T\left(X2\right),$

Wherein T () is profile wavelet transformation Contourlet, thereby obtains a width of cloth low frequency subgraph as Xi _LfWith a series of high frequency subimage Xi with different resolution _Hf ^{(k, l)}, wherein k ∈ (1, K) and l ∈ (1, L _k) indicating that subimage is positioned at the l direction of the tower decomposition of k layer Laplce LP, i represents 1 or 2;

3.. according to sparse property criterion, choose the high frequency subimage group X1 behind the profile wavelet transformation Contourlet _Hf ^{(k, l)}And X2 _Hf ^{(k, l)}In the most sparse subimage group, be designated as X1 _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}, this method is carried out the judgement of sparse property according to the star chart distribution and the clustering method of subimage group;

4.. to the high frequency subimage group X1 that 3. obtains in the step _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}, adopt traditional natural gradient learning method NGA to carry out blind source and separate, obtain separation matrix W, promptly

$W=\mathrm{NGA}({X1}_{\mathrm{hf}}^{(\mathrm{ksel},\mathrm{lsel})},{X2}_{\mathrm{hf}}^{(\mathrm{ksel},\mathrm{lsel})}),$

Wherein, NGA () represents natural gradient learning method NGA;

5.. utilize the W that 4. obtains in the step to separate the mixed signal that receives, obtaining isolated component Y1 and Y2 has

$\left[\begin{array}{c}Y1\\ Y2\end{array}\right]=W\left[\begin{array}{c}X1\\ X2\end{array}\right],$

Separating resulting Y1 that obtains and Y2 are the estimation of the original signal of separating.

3, the Image Blind source separation method based on sparse conversion according to claim 2 is characterized in that during described step is 3. that star chart according to the subimage group distributes and step that clustering method carries out the judgement of sparse property is:

(a) order $Z_{k,l}={[{X1}_{\mathrm{hf}}^{(k,l)},{X2}_{\mathrm{hf}}^{(k,l)}]}^T,$ Wherein k ∈ (1, K) and l ∈ (1, L _k);

(b) remove less coefficient component in the signal, to eliminate The noise;

(c) all data points are projected on the unit sphere, i.e. Z _k, l=Z _{K, l}/ || Z _{K, l}||;

(d) all signaling points are moved on to positive hemisphere face: if first coordinate of data point $z_{k,l}^1<0,$ Z _{K, l}=-Z _{K, l}Otherwise, Z _{K, l}=Z _{K, l}

(e) determine poly-axle and poly-axle center by clustering algorithm;

(f) calculate all data points to gathering the distance and the D of axle recently from self _{K, l}, and weigh sparse property, D with this _{K, l}More little, sparse more, the poly-axle in its star chart is just clear more;

(g) to all Z _k, (for k=1 ..., N _L) calculating D _{K, l}, seek its minimum value, order $(\mathrm{ksel},\mathrm{lsel})=\underset{k,l}{\arg \min }\left(D_{k,l}\right);$

(h) therefore, the most sparse subimage group is X1 _Hf ^{(ksel, lsel)}And X2 _Hf ^{(ksel, lsel)}

Images