CN103295187A - Mixed-noise-resisting blind image source separating method based on feedback mechanism - Google Patents

Abstract

The invention relates to a mixed-noise-resisting blind image source separating method based on a feedback mechanism. The method comprises the steps of utilizing wavelet transform to enable mixed signals to be sparse, carrying out linear cluster on sparse wavelet coefficients to estimate a hybrid matrix of a system, carrying out primary separation according to mixed images, then respectively calculating mean values of separating branches, taking out the most value to be outputted, utilizing a 0-setting feedback method to remove signals of the branch out of original mixed signals, carrying out next blind separation on the residual mixed signals in the method, and repeatedly carrying out the process until only a noise branch is left, namely completely separating all mixed signals. The method can effectively and blindly separate image mixing where Gaussian white noise participates. Compared with classis FastICA, the method can achieve higher separation accuracy.

Description

Blind image source separation method based on the anti-mixed noise of feedback mechanism

Technical field:

The invention belongs to Digital Image Processing and blind signal processing crossing domain, is a kind of blind image source separation method of the anti-mixed Gaussian white noise based on feedback mechanism.

Background technology:

Effectively denoising is that signal is handled one of greatest problem that the boundary runs into, coming from the blind separation of images (BSS of " cocktail party ", blind source separation) in the technology, most of researcher has only considered removal channel additive noise (with reference to documents 1).And in fact, in the tributary signal of source, not only have conventional signal branch, and also there is the noise source branch road, they have participated in the mixing of system together, and model is as shown in Figure 1.In existing blind image source separation algorithm, to the situation that has noise source branch road participation system to mix, do not see effective solution.To sum up, realize that the blind image source separation that contains the noise mixing has great importance and actual value.

The blind image source separation method of main flow all is based on independent component analysis (ICA, independent component analysis) and Sparse Component Analysis (SCA, sparse component analysis) algorithm, there are the following problems in the mixing situation that has noise to participate in:

1) generally can remove additive noise preferably based on the algorithm of ICA, but require between the source signal to be independently and to satisfy non-Gauss, this image blend to there being noise to participate in is difficult to guarantee only to have the Gauss that is of 1 branch road, so causes separating resulting undesirable (with reference to documents 2);

2) the linear clustering algorithm based on SCA is not having noise intervention and mixing source to satisfy under the condition of sparse type, good separation, but one have noise to participate in, and causes the reduction of sparse property, cause separating effect sharply to descend, finally correct separation source image (with reference to documents 3).

Because the mixing of multiple noise is close to Gaussian distribution, and white Gaussian noise has maximum noise entropy under identical energy, disturbs the most serious to signal.Based on above problem, be that the situation that white Gaussian noise participates in mixing has been done relevant treatment to 1 source branch road emphatically, in conjunction with feedback and the method for extracting one by one, solved the blind image source separation problem of mixed noise targetedly.

Documents 1:Fadili J M, Starck J L, Bobin J, et al.Image decomposition and separation using sparse representations:an overview[J] .Proceedings of the IEEE, 2010,98 (6): 983-994.

Documents 2: Lu Xiaoguang, Han Ping, Wu Renbiao, Liu Ruihua. based on the SAR image de-noising method [J] of two-dimensional wavelet transformation and independent component analysis, electronics and information journal, 2008,30 (5): 1052-1055.

Documents 3: Yu Xianchuan, Cao Tingting, Hu Dan, Zhang Libao is for Sha. based on the blind image separation method [J] of wavelet transformation and sparse component analysis, Beijing University of Post ﹠ Telecommunication's journal, 2010,33 (2): 58-63.

Summary of the invention:

Invented a kind of blind image source separation method of the anti-mixed Gaussian white noise based on feedback mechanism: implement one-level Haar integer wavelet transformation to containing the vision-mix signal that mixes white Gaussian noise, obtain the diagonal components coefficient of rarefaction, wavelet coefficient to rarefaction carries out linear clustering, thereby estimate the hybrid matrix of system, and separate first according to vision-mix, calculate the average of separating branch road then respectively, get the maximum output; By put 0, the method for feedback removes this road signal from former mixed signal, and carry out blind separation again, the normalized correlation coefficient until between isolated each tributary signal equals 1, namely separates the signal that all participate in mixing fully.Flow process as shown in Figure 2.

Description of drawings:

The blind source of Fig. 1 disjunctive model figure

Fig. 2 is based on the blind image source separation process figure of the anti-mixed noise of feedback mechanism

Fig. 3 tests the test source of employing

Fig. 4 standard testing image and noise image be mixing resultant at random

Fig. 5 FastICA separating resulting

Fig. 6 feeds back the SCA separating resulting

Fig. 7 standard testing image and noise image be mixing resultant at random

Fig. 8 feeds back the SCA separating resulting

Embodiment:

1) rarefaction.Vision-mix X to m width of cloth same size carries out the one-level integer wavelet transformation, chooses the diagonal components matrix of coefficients, and vectorial as the row of a matrix respectively by the row expansion, and this matrix is the matrix to the vision-mix rarefaction;

2) zero-suppress row and direction is unitized.To each row X of the matrix after the rarefaction _j(j=1,2 ..., T), if

Satisfy X _Ij=0,

J row deletion with X; If X _1j＜0, X then _j=-X _jProcessing obtains new mixed signal X ';

3) linear clustering.Any 2 column vector X ' for X ' _iAnd X ' _j, if

X ' then _iAnd X ' _jConllinear is established X ' _i∈ θ (k), X ' _j∈ θ (k), by this all column vector linear clusterings are obtained θ | θ (k), k=1,2 ..., T};

4) estimated mixing matrix A.Get the maximum preceding m class of cluster element among the θ, calculate the arithmetic mean of each class, obtain the cluster centre of corresponding class, the cluster centre matrix A _SCABe the hybrid matrix A of estimation;

5) direction is recovered.Therefore what adopt when finding the solution hybrid matrix is the method for conllinear, can have the problem of vector direction negate, so, ask the average of the source signal that of separation, then constant if greater than 0, if less than 0, then negate.

6) ask source signal S.According to the hybrid matrix A and the S=A that estimate ^-1X isolates source signal S;

7) output.If m=1, algorithm finishes; Otherwise, the maximum branch road of output average;

8) form new mixed signal.1 tunnel of perfection separation is set to complete 0 signal, by acquired hybrid matrix A _SCAReach perfect other tributary signal that separates and feed back to system, the new mixed signal that obtains so again is not for containing the signal of perfect separation;

9) choose m=m-1 road mixed signal and forward 1 to arbitrarily.

Simulation result:

The test image signal (256*256pixel) of Fig. 3 is adopted in experiment, and mixed noise intensity is 40dBW.The accurate test pattern of 2 road signs+1 road noise image is mixed at random, then it is carried out blind source and separate, and advanced the contrast experiment with the FastICA blind source separation method of classics.At random mixing resultant as shown in Figure 4, the separating resulting of FastICA as shown in Figure 5, the experimental result of this patent proposition method is (Shu Chu separate picture one by one) as shown in Figure 6.Calculate separation accuracy by normalized correlation coefficient (NCC Normalized correlation coefficients), formula is as follows:

$\mathrm{NCC}=\frac{{\mathrm{\Σ}}_{i,j}|X(i,j)-\stackrel{\‾}{X}|\×|Y(i,j)-\stackrel{\‾}{Y}|}{\sqrt{{\mathrm{\Σ}}_{i,j}{|X(i,j)-\stackrel{\‾}{X}|}^2}\×\sqrt{{\mathrm{\Σ}}_{i,j}{|Y(i,j)-\stackrel{\‾}{Y}|}^2}}$

Wherein, X (i, j) with Y (i, j) be respectively two width of cloth image X and Y (i, grey scale pixel value j),

With Be the pixel grey scale mean value of corresponding two width of cloth images.NCC is more big, illustrates that degree of correlation is more big between two width of cloth images, and is namely more similar.

Normalized correlation coefficient (NCC Normalized correlation coefficients) between two kinds of method separating resultings and original signal is as shown in table 1.By Fig. 5, Fig. 6 and table 1, on subjective naked-eye observation and objective indicator NCC, all can find, based on the method for feedback mechanism separation signal effectively not only, and can isolate noise effectively; And FastICA can not separate effectively.

Table 1 normalized correlation coefficient (NCC)

Situation to the accurate test pattern of 3 road signs+1 road noise image has been done related experiment.At random mixing resultant as shown in Figure 7, separating resulting as shown in Figure 8, the NCC between separating resulting and corresponding test pattern is 1, and is as shown in table 2.Therefore, also can isolate each tributary signal and noise effectively, reliably based on the SCA blind source separation method of feedback mechanism to this situation.

The NCC (the accurate test pattern of 3 road signs+1 road noise image) of table 2 feedback SCA

Claims (7)

1. based on the blind image source separation method of the anti-mixed noise of feedback mechanism, it is characterized in that: the mixing gray level image X that mixes the white Gaussian noise noise that contains to m width of cloth same size does one-level haar wavelet transformation respectively, m the diagonal components matrix that obtains is converted into m vector by row respectively, and it is vectorial as the row of matrix X ' respectively, column vector to matrix X ' is carried out linear clustering, estimated mixing matrix A _SCA, utilize A _SCA ^-1X obtains the source images signal S ' of separation first, calculates the average of each separate picture signal S ' then, the branch road of output average maximum; Simultaneously with this tributary signal zero setting, and unite other and separate tributary signal by calculating A _SCAS ' removes this branch road from former mixed signal, form new vision-mix signal B, and B uses said method again to the vision-mix signal, extracts each branch road picture signal one by one, realizes that the suitable fixed blind image source of anti-mixed Gaussian white noise is separated.

2. the blind image source separation method of the anti-mixed noise based on feedback mechanism as claimed in claim 1, wherein the m width of cloth contains the vision-mix of mixing noise and is characterised in that: to 2 these white noise images of width of cloth standard testing image+1 panel height of same size linear hybrid at random, obtain the m=3 width of cloth and contain the vision-mix of mixing noise; Perhaps to 3 these white noise images of width of cloth standard testing image+1 panel height of same size linear hybrid at random, obtain the m=4 width of cloth and contain the vision-mix of mixing noise.

3. the blind image source separation method of the anti-mixed noise based on feedback mechanism as claimed in claim 1, wherein linear clustering is characterised in that: by judging vectorial X ' _iWith vectorial X ' _jBetween the included angle cosine value whether equal 1, that is: Judge vectorial X ' _iWith vectorial X ' _jConllinear whether is if conllinear then is classified as same item.

4. the blind image source separation method of the anti-mixed noise based on feedback mechanism as claimed in claim 1, wherein estimated mixing matrix A _SCABe characterised in that: choose the m class that the class interior element is maximum in the linear cluster result, by asking the arithmetic mean of every class, obtain corresponding cluster centre, m cluster centre namely formed the hybrid matrix A of estimation _SCA

5. the blind image source separation method of the anti-mixed noise based on feedback mechanism as claimed in claim 1 wherein feeds back removal and is characterised in that: calculates the average of each separation signal, the branch road of output average maximum; With this tributary signal zero setting, and unite other separation tributary signal and pass through A _SCAThe method of S ' is removed this branch road from former mixed signal, form new vision-mix signal B.

6. the blind image source separation method of the anti-mixed noise based on feedback mechanism as claimed in claim 1, wherein extract one by one and be characterised in that: distinguish the signal that separates the participation mixing whether fully by the normalized correlation coefficient (NCC) that calculates between each separation signal, if NCC＞0.99, then think only remaining one road signal in the mixed signal, separate and finish, otherwise continue to use said method until NCC＞0.99, the computing method of NCC are shown below:

$\mathrm{NCC}=\frac{{\mathrm{\Σ}}_{i,j}|X(i,j)-\stackrel{\‾}{X}|\×|Y(i,j)-\stackrel{\‾}{Y}|}{\sqrt{{\mathrm{\Σ}}_{i,j}{|X(i,j)-\stackrel{\‾}{X}|}^2}\×\sqrt{{\mathrm{\Σ}}_{i,j}{|Y(i,j)-\stackrel{\‾}{Y}|}^2}}$

In the formula, X (i, j) with Y (i, j) be respectively two width of cloth image X and Y (i, grey scale pixel value j),

With Be the pixel grey scale mean value of corresponding two width of cloth images.

7. the blind image source separation method of the anti-mixed noise based on feedback mechanism as claimed in claim 1, wherein suitable fixed blind image source is separated and is characterised in that: the number that participates in the image source of mixing equals the number of vision-mix.

Images