CN103500444A - Polarization image fusion method - Google Patents

Abstract

The invention discloses a polarization image fusion method. The polarization image fusion method combines a non-negative matrix factorization method with a pulse coupling neural network method. The non-negative matrix factorization method is adopted to factorize matrixes formed by polarization parameter images, two feature base vector images are acquired, and through a pulse coupling neural network and a section of frequency iteration, according to a corresponding fusion rule, polarization fusion images are acquired. Compared with other fusion methods, the polarization image fusion method is improved to a certain degree from the aspects of the visual effect and objective indicators, due to the fact that more scene polarization information is fused in the acquired result, disorderly backgrounds can be effectively restrained, the detail features of targets are prominent, and then the capacity for utilizing a polarization imaging detection means to recognize the targets can be improved.

Description

A kind of Polarization Image Fusion method

Technical field

The invention belongs to the information fusion field, relate to a kind of Polarization Image Fusion method, specifically, is a kind of Polarization Image Fusion method based on Non-negative Matrix Factorization and Pulse Coupled Neural Network.

Background technology

Polarization is one of inherent characteristic of optical radiation, and it is strong the supplementing to traditional imaging detection technology that polarization imaging is surveyed.The relevant target reflection that the Polarization Detection technology provides common detection means to obtain or the polarization information of radiant light, the polarization image obtained provides than common intensity image and the more abundant information about scene of spectrum picture, thereby is widely used in many fields such as geology detecting, materials classification, machine vision.

Polarized light generally take the form of elliptically polarized light.In actual applications, introduce the such one group of duplicate parameter of physics dimension of stokes parameter and mark the polarization state of levying polarized light.

Stokes parameter is through being often expressed as:

S＝[I，Q，U，V] ^T

Wherein, I means the total intensity of light wave, and Q means the intensity of linearly polarized light on horizontal direction, and U means the intensity of linearly polarized light on 45 ° of directions, and V means the circular polarization light intensity.Stokes vector can be easy to measure and be included in the random polarization light in partial poolarized light.Two other parameter is extremely important in the expression of polarization state, mean the ratio of complete polarized light intensity in whole light intensity with degree of polarization P, mean the angle of polarized light direction of vibration and reference direction (reference direction is got horizontal direction x axle) with polarization angle θ, be expressed from the next:

$\left\{\begin{array}{c}P=\frac{\sqrt{Q^2+U^2+V^2}}{I}\\ \mathrm{\θ}=\frac{1}{2}{\tan }^{-1}\frac{U}{Q}\end{array}\right.$

Each polarization parameter image, being redundancy, complementary aspect the expression polarization information, utilizes complementarity and redundancy between them, takes image interfusion method to solve these problems.By Polarization Image Fusion, can improve contrast and the sharpness of soft image; Can greatly compress natural background, detect man-made target simultaneously from complicated nature background.Therefore to the research of polarization information fusion method, there is very important researching value and certain construction value.

Document Shutao Li, Bin Yang, Jianwen Hu.Performance comparison of different multi-resolution transforms for image fusion[J] .Information Fusion, 2011,12 (2): in 74-84, openly utilize the method for wavelet transformation to carry out the fusion of image, the method is carried out wavelet transformation to original image, and it is decomposed on the different characteristic territory of different frequency range, sets up the small echo turriform of each image and decomposes; Then each decomposition layer is carried out respectively to fusion treatment, the different frequency component on each decomposition layer adopts different fusion rules to carry out fusion treatment, the wavelet pyramid after being merged; Finally to merging rear gained wavelet pyramid, carry out wavelet inverse transformation (carrying out Image Reconstruction), obtain fused images.If two dimensional image is carried out to the wavelet decomposition of N layer, (3N+1) individual different frequency bands is arranged the most at last, wherein comprise 3N high frequency band and a low-frequency band.This method has the following disadvantages:

(1) fused images obtained has more ground unrest, and improvement of visual effect is poor;

(2) do not possess translation invariance.This means that the displacement that input picture is very small will cause the very large variation of energy distribution between the discrete wavelet transform coefficients of different scale;

(3) limited directional selectivity.In common wavelet decomposition, each metric space can only be broken down into 3 limited directions: level, vertical, diagonal angle, its direction alternative is very limited.

Document P.J.Burt and E.H.Adelson, " The Laplacian pyramid as a compact image code ", IEEE Trans.Comm.31 (4), disclose and utilized the method for Laplacian Pyramid Transform to merge image in 532-540 (1983).At first the method carries out Gassian low-pass filter and the interlacing down-sampling every row to original input picture, obtains the ground floor of gaussian pyramid; To the filtering of ground floor Image Low-passed and down-sampling, obtain the second layer of gaussian pyramid again; Repeat above process, form the gaussian pyramid of each image.And then this tomographic image of gaussian pyramid that calculates each image tomographic image high with it image poor after interpolation is amplified, obtain the laplacian pyramid of each image.Laplacian pyramid has represented the edge details of every one-level image, therefore by comparing the laplacian pyramid of two width image respective stages, with regard to having, outstanding separately image detail can be utilized fusion rule choose in fused images, make like this quantity of information of fused images abundant as far as possible, reach the purpose of fusion.This method, in conversion process, because noise has been introduced in the operations such as quantification or threshold value, has reduced sharpness and the signal to noise ratio (S/N ratio) of fused images, has affected syncretizing effect.

Summary of the invention

In order to address the above problem, the invention provides a kind of Polarization Image Fusion method that Non-negative Matrix Factorization method and Pulse Coupled Neural Network method are combined, can effectively give prominence to the minutia of object, improve the ability of utilizing polarization imaging detection means recognition object.

Polarization Image Fusion method of the present invention, realize by following step:

Step 1: the polarization parameter view data that obtains the target polarization image;

Utilize polarized imaging system to carry out polarization imaging to target, the angle that obtains the light transmission shaft of polaroid in polarized imaging system and horizontal direction is respectively the output intensity polarization image I of 0 °, 45 °, 90 °, 135 ° four polarization directions ₀, I ₄₅, I ₉₀, I ₁₃₅; And then obtain the polarization parameter view data, comprise stokes parameter I, Q, the U of polarized light, and the parameter image of degree of polarization P and polarization angle θ.

Step 2: obtain two width feature bases images;

By polarization parameter view data I, Q, U, P and the θ obtained in step 1, form the data vector matrix V of the capable q row of p _{p * q};

$V_{p\×q}={\left[\begin{array}{c}I\\ Q\\ U\\ P\\ \mathrm{\θ}\end{array}\right]}^T=\left[\begin{array}{ccccc}{I}_{11}& Q_{11}& U_{11}& P_{11}& \mathrm{\θ}\\ I_{12}& Q_{12}& U_{12}& P_{12}& \mathrm{\θ}\\ .& .& .& .& .\\ .& .& .& .& .\\ .& .& .& .& .\\ I_{M\×N}& Q_{M\×N}& U_{M\×N}& P_{M\×N}& {\mathrm{\θ}}_{M\×N}\end{array}\right]---\left(1\right)$

Wherein, M and N are the polarization parameter data line number and columns.

To the data vector matrix, V carries out Non-negative Matrix Factorization, by data vector matrix V approximate factorization, is the non-negative base vector matrix D with the capable r row of p _{p * r}with the non-negative coefficient matrix H with the capable q row of r _{r * q}product; And make r=2; Have:

$\underset{D,H}{\min }f(D,H)\≡\frac{1}{2}\underset{i=1}{\overset{p}{\mathrm{\Σ}}}\underset{j=1}{\overset{q}{\mathrm{\Σ}}}{[V_{\mathrm{ij}}-{\left(\mathrm{DH}\right)}_{\mathrm{ij}}]}^2$ (2)

$s.t.\left\{\begin{array}{c}{D}_{\mathrm{ia}}\≥0\\ H_{\mathrm{bj}}\≥0\end{array}\right.,\∀i,a,j$

In formula (2), V _ijpixel for the capable j row of i in matrix V; I=1,2,3 ..., p; J=1,2,3 ..., q; D _iarepresent the pixel of the capable a row of i in matrix D, H _ajrepresent the pixel of the capable j row of a in matrix H; A=1,2.

A, random initializtion matrix D and H.

Matrix D and H are upgraded in b, employing alternately multiplication update rule, have:

$\left\{\begin{array}{c}{H}_{\mathrm{aj}}\←H_{\mathrm{aj}}\frac{{\left(D^{T}V\right)}_{\mathrm{aj}}}{{\left(D^T\mathrm{DH}\right)}_{\mathrm{aj}}}\\ D_{\mathrm{ia}}\←D_{\mathrm{ia}}\frac{{\left({\mathrm{VH}}^T\right)}_{\mathrm{ia}}}{{\left({\mathrm{DHH}}^T\right)}_{\mathrm{ia}}}\end{array}\right.---\left(3\right)$

Obtain matrix D=[D ₁, D ₂]; Matrix D has comprised the approximate complete information that participates in whole polarization parameter images of fusion, can be for the approximate reproduction of object polarization information; D ₁for the first row pixel of matrix D, as the first image array; D ₂for the secondary series pixel of matrix D, as the second image array.

E, acquisition D ₁with D ₂sharpness g and variance C:

$g=\frac{1}{\mathrm{MN}}\underset{i^{\′}=1}{\overset{M}{\mathrm{\Σ}}}\underset{j^{\′}=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{\frac{{\mathrm{\Δf}}_x^2{(i^{\′},j^{\′})}^2+{\mathrm{\Δf}}_y^2{(i^{\′},j^{\′})}^2}{2}}---\left(4\right)$

$C=\frac{1}{\mathrm{MN}}\underset{i^{\′}=1}{\overset{M}{\mathrm{\Σ}}}\underset{j^{\′}=1}{\overset{N}{\mathrm{\Σ}}}{[I_{i^{\′}{j}^{\′}}-\stackrel{\‾}{I}]}^2---\left(5\right)$

In formula (4), Δ f _x, Δ f _ydifference representative image matrix x, the difference of y direction; I ' and j ' mean respectively i ' row and the j ' row pixel in the polarization parameter view data, i '=1,2,3 ..., M, j '=1,2,3 ..., N.

In formula (5), I _{i ' j '}gray-scale value for the i ' row, the j ' row pixel in image array;

for the average gray value of image array, choose D ₁and D ₂the feature bases of the image array that middle sharpness and variance are larger is as the standard base vector, feature bases to another image array carries out the histogram specification processing, finally obtains feature bases image A and feature bases image B after two width are processed.

Step 3: obtain the polarization image merged;

The two width feature bases images that obtain in step 2 are sent into respectively in two identical PCNN networks, utilized the PCNN model formation to carry out iteration N time, obtain respectively two feature bases image the i ' row j ' and be listed as neuronic output Y _{i ' j '}; Subsequently, according to formula R _{i ' j '}(n)=R _{i ' j '}(n-1)+Y _{i ' j '}(n), two feature bases image the i ' row j ' of iterative computation are listed as neuronic igniting number of times R _{i ' j '}; N is iterations.

Obtain the matching degree Z of the PCNN pulse number of two feature bases images:

$Z_{i^{\′}{j}^{\′}}\left(n\max \right)=\frac{2\×R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)\×R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)}{{\left[R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)\right]}^2+{\left[R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)\right]}^2}---\left(6\right)$

In formula (6), Z _{i ' j '}for feature bases image the i ' row, j ' is listed as neuronic matching degree; Nmax is maximum iteration time; for feature bases image A the i ' row, j ' is listed as neuronic igniting number of times;

for feature bases image B the i ' row, j ' is listed as neuronic igniting number of times.

Finally, by following formula, the corresponding neuron in two width two width feature bases images is chosen the image after being merged:

In formula (7), I ' _{i ' j '}the gray-scale value that means the polarization image the i ' row the j ' row pixel of fusion; R _thit is 0～1 constant; R in the present invention _thbe preferably 0.95; A _{i ' j '}for feature bases image A the i ' row the j ' row neuron; B _{i ' j '}for the i ' row, the j ' row neuron in the feature bases image B;

$G_A=\frac{R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)}{R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)+R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)};$

$G_B=\frac{R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)}{R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)+R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)}.$

The invention has the advantages that:

1, Polarization Image Fusion method of the present invention all improves than additive method on visual effect and objective indicator;

2, Polarization Image Fusion method of the present invention, fusion results has incorporated more scene polarization information, can effectively suppress mixed and disorderly background, and the minutia of outstanding object, can improve the ability of utilizing polarization imaging detection means recognition object;

3, Polarization Image Fusion method of the present invention, applicability is strong, can be applied to multiple occasion.

The accompanying drawing explanation

Fig. 1 is Polarization Image Fusion method flow diagram of the present invention;

Fig. 2 is for adopting the fusion results figure to target image;

Fig. 3 is for adopting the fusion results figure to target image;

Fig. 4 is for adopting the fusion results figure of the inventive method to target image.

Embodiment

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

Polarization Image Fusion method of the present invention, as shown in Figure 1, realize by following step:

Step 1: the polarization parameter view data that obtains the target polarization image;

Utilize polarized imaging system to carry out polarization imaging to target, the angle that obtains the light transmission shaft of polaroid in polarized imaging system and horizontal direction is respectively the output intensity polarization image I of 0 °, 45 °, 90 °, 135 ° four polarization directions ₀, I ₄₅, I ₉₀, I ₁₃₅; Can obtain the polarization parameter view data thus, comprise stokes parameter I, Q, U, the V (it is 0 that common engineering makes V in surveying and calculating) of polarized light, and the parameter image of degree of polarization P and polarization angle θ;

$\left\{\begin{array}{c}I=\frac{1}{2}(I_0+I_{45}+I_{90}+I_{135})\\ Q=I_0-I_{90}\\ U=I_{45}-I_{135}\\ V=0\end{array}\right.---\left(1\right)$

$\left\{\begin{array}{c}P=\frac{\sqrt{Q^2+U^2}}{I}\\ \mathrm{\θ}=\frac{1}{2}{\tan }^{-1}\frac{U}{Q}\end{array}\right.---\left(2\right)$

Step 2: obtain two width feature bases images;

By polarization parameter view data I, Q, U, P and the θ obtained in step 1, form the data vector matrix V of p line number q columns _{p * q}, its each row represent a polarization parameter, so n=5.

$V_{p\×q}={\left[\begin{array}{c}I\\ Q\\ U\\ P\\ \mathrm{\θ}\end{array}\right]}^T=\left[\begin{array}{ccccc}{I}_{11}& Q_{11}& U_{11}& P_{11}& \mathrm{\θ}\\ I_{12}& Q_{12}& U_{12}& P_{12}& \mathrm{\θ}\\ .& .& .& .& .\\ .& .& .& .& .\\ .& .& .& .& .\\ I_{M\×N}& Q_{M\×N}& U_{M\×N}& P_{M\×N}& {\mathrm{\θ}}_{M\×N}\end{array}\right]---\left(3\right)$

Wherein, M and N are the polarization parameter data line number and columns;

To the data vector matrix, V carries out Non-negative Matrix Factorization, by data vector matrix V approximate factorization, is the non-negative base vector matrix D with the capable r row of p _{p * r}with the non-negative coefficient matrix H with the capable q row of r _{r * q}product; Have:

$\underset{D,H}{\min }f(D,H)\≡\frac{1}{2}\underset{i=1}{\overset{p}{\mathrm{\Σ}}}\underset{j=1}{\overset{q}{\mathrm{\Σ}}}{[V_{\mathrm{ij}}-{\left(\mathrm{DH}\right)}_{\mathrm{ij}}]}^2$

$s.t.\left\{\begin{array}{c}{D}_{\mathrm{ia}}\≥0\\ H_{\mathrm{bj}}\≥0\end{array}\right.,\∀i,a,j---\left(4\right)$

In formula (4), V _ijpixel for the capable j row of i in matrix V; I=1,2,3 ..., p; J=1,2,3 ..., q; D _iarepresent the pixel of the capable a row of i in matrix D, H _ajrepresent the pixel of the capable j row of a in matrix H; A=1,2,3 ..., r; For matrix D and matrix H, if select the dimension r of base vector to be less than p or q, the matrix D obtained so and matrix H can be less than the data vector matrix V, be equivalent to thus mean original polarization data vector with relatively few base vector, the matrix D obtained has certain linear independence and sparse property, thereby makes matrix D have suitable ability to express to feature and the structure of original polarization data; As long as suitably choose the dimension r of base vector, just can obtain most of characteristic information of original polarization data; Get r=2 in the present invention;

A, random initializtion matrix D and H;

Matrix D and H are upgraded in b, employing alternately multiplication update rule, have:

$\left\{\begin{array}{c}{H}_{\mathrm{aj}}\←H_{\mathrm{aj}}\frac{{\left(D^{T}V\right)}_{\mathrm{aj}}}{{\left(D^T\mathrm{DH}\right)}_{\mathrm{aj}}}\\ D_{\mathrm{ia}}\←D_{\mathrm{ia}}\frac{{\left({\mathrm{VH}}^T\right)}_{\mathrm{ia}}}{{\left({\mathrm{DHH}}^T\right)}_{\mathrm{ia}}}\end{array}\right.---\left(5\right)$

Obtain matrix D=[D ₁, D ₂]; Matrix D has comprised the approximate complete information that participates in whole polarization parameter images of fusion, can be for the approximate reproduction of object polarization information; D ₁for the first row pixel of matrix D, as the first image array; D ₂for the secondary series pixel of matrix D, as the second image array;

E, acquisition D ₁with D ₂sharpness g and variance C:

$g=\frac{1}{\mathrm{MN}}\underset{i^{\′}=1}{\overset{M}{\mathrm{\Σ}}}\underset{j^{\′}=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{\frac{\mathrm{\Δ}{f}_x^2{(i^{\′},j^{\′})}^2+\mathrm{\Δ}{f}_y^2{(i^{\′},j^{\′})}^2}{2}}---\left(6\right)$

$C=\frac{1}{\mathrm{MN}}\underset{i^{\′}=1}{\overset{M}{\mathrm{\Σ}}}\underset{j^{\′}=1}{\overset{N}{\mathrm{\Σ}}}{[I_{i^{\′}{j}^{\′}}-\stackrel{\‾}{I}]}^2---\left(7\right)$

In formula (6), Δ f _x, Δ f _ydifference representative image matrix x, the difference of y direction; I ' and j ' mean respectively i ' row and the j ' row pixel in the polarization parameter view data, i '=1,2,3 ..., M, j '=1,2,3 ..., N.

In formula (7), I _{i ' j '}gray-scale value for the i ' row, the j ' row pixel in image array;

for the average gray value of image array, choose D ₁and D ₂the feature bases of the image array that middle sharpness and variance are larger is as the standard base vector, feature bases to another image array carries out the histogram specification processing, finally obtains feature bases image A and feature bases image B after two width are processed.

Step 3: obtain the polarization image merged.

The two width feature bases images that obtain in step 2 are sent into respectively in two identical PCNN networks, are utilized the PCNN model formation to carry out iteration N time, be specially:

$\left\{\begin{array}{c}{F}_{i^{\′}{j}^{\′}}\left(n\right)=I_{i^{\′}{j}^{\′}}\left(n\right)\\ L_{i^{\′}{j}^{\′}}\left(n\right)=\exp (-{\mathrm{\α}}_L)L_{i^{\′}{j}^{\′}}(n-1)+V_L\underset{k,l}{\mathrm{\Σ}}{W}_{i^{\′}{j}^{\′}\mathrm{kl}}{Y}_{\mathrm{kl}}(n-1)\\ U_{i^{\′}{j}^{\′}}\left(n\right)=F_{i^{\′}{j}^{\′}}\left(n\right)(1+\mathrm{\β}{L}_{i^{\′}{j}^{\′}}\left(n\right))\\ Y_{i^{\′}{j}^{\′}}\left(n\right)=\left\{\begin{array}{cc}1,& U_{i^{\′}{j}^{\′}}\left(n\right)>{\mathrm{\θ}}_{i^{\′}{j}^{\′}}\left(n\right)\\ 0,& U_{i^{\′}{j}^{\′}}\left(n\right)\≤{\mathrm{\θ}}_{i^{\′}{j}^{\′}}\left(n\right)\end{array}\right.\\ {\mathrm{\θ}}_{i^{\′}{j}^{\′}}\left(n\right)=\exp (-{\mathrm{\α}}_{\mathrm{\θ}}){\mathrm{\θ}}_{i^{\′}{j}^{\′}}(n-1)+V_{\mathrm{\θ}}{Y}_{i^{\′}{j}^{\′}}\left(n\right)\end{array}\right.---\left(8\right)$

Wherein, n is iterations; F _{i ' j '}for feature bases image the i ' row, j ' is listed as neuronic feed back input; L _{i ' j '}for feature bases image the i ' row j ' is listed as neuronic link input; I _{i ' j '}for the i ' row j ' in image array is listed as neuronic gray-scale value; U _{i ' j '}for feature bases image the i ' row the j ' row inside neurons state.Y _{i ' j '}for feature bases image the i ' row, j ' is listed as neuronic output.W _{i ' j ' KL}for L _{i ' j '}(n) Y in _kl(n-1) weighting coefficient matrix; Y _klfor L _{i ' j '}neuron output in neighborhood (k, 1 means neuron and the scope be connected on every side, generally gets 3 * 3 or 5 * 5); θ _{i ' j '}for feature bases image the i ' row, j ' is listed as neuronic dynamic threshold; α _lfor link input the time ask attenuation constant; α _θtime constant for the variable threshold value function.V _lfor link input constant; V _θfor the threshold value amplification coefficient; β is link modulation constant.When carrying out interative computation, if U at every turn _{i ' j '}be greater than threshold value θ _{i ' j '}y greatly, _{i ' j '}value is 1, now claims the i ' row, the j ' row neuron firing; Otherwise, Y _{i ' j '}value is that 0, the i ' row, the j ' row neuron misfires.

Subsequently, according to formula R _{i ' j '}(n)=R _{i ' j '}(n-1)+Y _{i ' j '}(n), in two feature bases images of iterative computation, the i ' row j ' is listed as neuronic igniting number of times R _{i ' j '};

Through type (9) obtains the matching degree Z of the PCNN pulse number of two width feature bases images:

$Z_{i^{\′}{j}^{\′}}\left(n\max \right)=\frac{2\×{R_A}_{i^{\′}{j}^{\′}}\left(n\max \right)\×{R_B}_{i^{\′}{j}^{\′}}\left(n\max \right)}{{\left[{R_A}_{i^{\′}{j}^{\′}}\left(n\max \right)\right]}^2+{\left[{R_B}_{i^{\′}{j}^{\′}}\left(n\max \right)\right]}^2}---\left(9\right)$

In formula (9), Z _{i ' j '}for feature bases image i ' row, j ' is listed as neuronic matching degree; Nmax is maximum iteration time;

for feature bases image A the i ' row, j ' is listed as neuronic igniting number of times; for feature bases image B the i ' row, j ' is listed as neuronic igniting number of times.

Finally, through type (10) is chosen the corresponding neuron in two width feature bases images, the image after being merged:

In formula (10), I ' _{i ' j '}the gray-scale value that means the polarization image the i ' row the j ' row pixel of fusion; R _thconstant for O～1; R in the present invention _thbe preferably 0.95; A _{i ' j '}for feature bases image A the i ' row the j ' row neuron; B _{i ' j '}

For the i ' row, the j ' row neuron in the feature bases image B. $G_A=\frac{{R_A}_{i^{\′}{j}^{\′}}\left(n\max \right)}{{R_A}_{i^{\′}{j}^{\′}}\left(n\max \right)+{R_B}_{i^{\′}{j}^{\′}}\left(n\max \right)};$

$G_B=\frac{{R_B}_{i^{\′}{j}^{\′}}\left(n\max \right)}{{R_A}_{i^{\′}{j}^{\′}}\left(n\max \right)+{R_B}_{i^{\′}{j}^{\′}}\left(n\max \right)}.$

As shown in Figure 2, Figure 3, Figure 4, be respectively and adopt existing small wave converting method, laplacian pyramid method and the inventive method fusion results figure to same target image; Can find out that the fused images that adopts the inventive method to obtain all improves than additive method on visual effect and objective indicator, the result obtained is owing to having incorporated more scene polarization information, therefore can effectively suppress mixed and disorderly background, the minutia of outstanding target, thus the ability of utilizing polarization imaging detection means identification target can be improved.

Claims (2)

1. a Polarization Image Fusion method is characterized in that: by following step, realize:

Step 1: the polarization parameter view data that obtains the target polarization image;

Utilize polarized imaging system to carry out polarization imaging to target, the angle that obtains the light transmission shaft of polaroid in polarized imaging system and horizontal direction is respectively the output intensity polarization image I of 0 °, 45 °, 90 °, 135 ° four polarization directions ₀, I ₄₅, I ₉₀, I ₁₃₅; And then obtain the polarization parameter view data, comprise stokes parameter I, Q, the U of polarized light, and the parameter image of degree of polarization P and polarization angle θ;

Step 2: obtain two width feature bases images;

By polarization parameter view data I, Q, U, P and the θ obtained in step 1, form the data vector matrix V of the capable q row of p _{p * q};

$V_{p\×q}={\left[\begin{array}{c}I\\ Q\\ U\\ P\\ \mathrm{\θ}\end{array}\right]}^T=\left[\begin{array}{ccccc}{I}_{11}& Q_{11}& U_{11}& P_{11}& \mathrm{\θ}\\ I_{12}& Q_{12}& U_{12}& P_{12}& \mathrm{\θ}\\ \·& \·& \·& \·& \·\\ \·& \·& \·& \·& \·\\ \·& \·& \·& \·& \·\\ I_{M\×N}& Q_{M\×N}& U_{M\×N}& P_{M\×N}& {\mathrm{\θ}}_{M\×N}\end{array}\right]---\left(1\right)$

Wherein, M and N are the polarization parameter data line number and columns;

To the data vector matrix, V carries out Non-negative Matrix Factorization, by data vector matrix V approximate factorization, is the non-negative base vector matrix D with the capable r row of p _{p * r}with the non-negative coefficient matrix H with the capable q row of r _{r * q}product; And make r=2; Have:

$\underset{D,H}{\min }f(D,H)\≡\frac{1}{2}\underset{i=1}{\overset{p}{\mathrm{\Σ}}}\underset{j=1}{\overset{q}{\mathrm{\Σ}}}{[V_{\mathrm{ij}}-{\left(\mathrm{DH}\right)}_{\mathrm{ij}}]}^2$ (2)

$s.t.\{\begin{array}{c}{D}_{\mathrm{ia}}\≥0\\ H_{\mathrm{bj}}\≥0\end{array},\∀i,\mathrm{\α},j$

In formula (2), V _ijpixel for the capable j row of i in matrix V; I=1,2,3 ..., p; J=1,2,3 ..., q; D _iarepresent the pixel of the capable a row of i in matrix D, H _ajrepresent the pixel of the capable j row of a in matrix H; A=1,2;

A, random initializtion matrix D and H;

Matrix D and H are upgraded in b, employing alternately multiplication update rule, have:

$\left\{\begin{array}{c}{H}_{\mathrm{aj}}\←H_{\mathrm{aj}}\frac{{\left(D^{T}V\right)}_{\mathrm{aj}}}{{\left(D^T\mathrm{DH}\right)}_{\mathrm{aj}}}\\ D_{\mathrm{ia}}\←D_{\mathrm{ia}}\frac{{\left({\mathrm{VH}}^T\right)}_{\mathrm{ia}}}{{\left({\mathrm{DHH}}^T\right)}_{\mathrm{ia}}}\end{array}\right.---\left(3\right)$

Obtain matrix D=[D ₁, D ₂]; Matrix D has comprised the approximate complete information that participates in whole polarization parameter images of fusion, can be for the approximate reproduction of object polarization information; D ₁for the first row pixel of matrix D, as the first image array; D ₂for the secondary series pixel of matrix D, as the second image array;

E, acquisition D ₁with D ₂sharpness g and variance C:

$g=\frac{1}{\mathrm{MN}}\underset{i^{\′}=1}{\overset{M}{\mathrm{\Σ}}}\underset{j^{\′}=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{\frac{\mathrm{\Δ}{f}_x^2{(i^{\′},j^{\′})}^2+\mathrm{\Δ}{f}_y^2{(i^{\′},j^{\′})}^2}{2}}---\left(4\right)$

$C=\frac{1}{\mathrm{MN}}\underset{i^{\′}=1}{\overset{M}{\mathrm{\Σ}}}\underset{j^{\′}=1}{\overset{N}{\mathrm{\Σ}}}{[I_{i^{\′}{j}^{\′}}-\stackrel{\‾}{I}]}^2---\left(5\right)$

In formula (4), Δ f _x, Δ f _ydifference representative image matrix x, the difference of y direction; I ' and j ' mean respectively i ' row and the j ' row pixel in the polarization parameter view data, i '=1,2,3 ..., M, j '=1,2,3 ..., N:

In formula (5), I _{i ' j '}gray-scale value for the i ' row, the j ' row pixel in image array;

for the average gray value of image array, choose D ₁and D ₂the feature bases of the image array that middle sharpness and variance are larger is as the standard base vector, feature bases to another image array carries out the histogram specification processing, finally obtains feature bases image A and feature bases image B after two width are processed;

Step 3: obtain the polarization image merged;

The two width feature bases images that obtain in step 2 are sent into respectively in two identical PCNN networks, utilized the PCNN model formation to carry out iteration N time, obtain respectively two feature bases image the i ' row j ' and be listed as neuronic output Y _{i ' j '}; Subsequently, according to formula R _{i ' j '}(n)=R _{i ' j '}(n-1)+Y _{i ' j '}(n), two feature bases image the i ' row j ' of iterative computation are listed as neuronic igniting number of times R _{i ' j '}; N is iterations;

Obtain the matching degree z of the PCNN pulse number of two feature bases images:

$Z_{i^{\′}{j}^{\′}}\left(n\max \right)=\frac{2\×R_{A{i}^{\′}{j}^{\′}}\left(n\max \right)\×R_{B{i}^{\′}{j}^{\′}}\left(n\max \right)}{{\[R_{A{i}^{\′}{j}^{\′}}\left(n\max \right)\]}^2+{\left[R_{B{i}^{\′}{j}^{\′}}\left(n\max \right)\right]}^2}---\left(6\right)$

In formula (6), Z _{i ' j '}for feature bases image the i ' row, j ' is listed as neuronic matching degree; Nmax is maximum iteration time; R _{i ' j '}for the neuronic igniting number of times of feature bases image A the i ' row ' be listed as;

for the neuronic igniting number of times of feature bases image B the i ' row ' be listed as;

Finally, by following formula, the corresponding neuron in two width two width feature bases images is chosen the image after being merged:

In formula (7), I ' _{i ' j '}the gray-scale value that means the polarization image the i ' row the j ' row pixel of fusion; R _thit is 0～1 constant; R in the present invention _thbe preferably 0.95; A _{i ' j '}for feature bases image A the i ' row the j ' row neuron; B _{i ' j '}for the i ' row, the j ' row neuron in the feature bases image B.

$G_A=\frac{R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)}{R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)+R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)};$

$G_B=\frac{R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)}{R_{A_{i^{\′}{j}^{\′}}}\left(n\max \right)+R_{B_{i^{\′}{j}^{\′}}}\left(n\max \right)}.$

2. a kind of Polarization Image Fusion method as claimed in claim 1 is characterized in that: in step 3, and described R _thbe preferably 0.95.

Images