CN106327459A - Visible light and infrared image fusion algorithm based on UDCT (Uniform Discrete Curvelet Transform) and PCNN (Pulse Coupled Neural Network) - Google Patents

Abstract

The invention discloses a visible light and infrared image fusion algorithm based on UDCT (Uniform Discrete Curvelet Transform) and a PCNN (Pulse Coupled Neural Network), which relates to the technical field of image processing and solves the technical problems that the fused image is not clear and details are unobvious as the similarity of low-frequency information of a to-be-fused image can not be judged in the prior art, and the detail richness degree of a source image can not be judged. The algorithm of the invention mainly comprises steps: (1) after source images of the visible light image and the infrared image are subjected to UDCT decomposition, UDCT subband coefficients with different scales in different directions can be obtained, and the UDCT subband coefficients comprise low-frequency and high-frequency UDCT coefficients; and (2) according to a specific rule, a different mode is adopted for each scale layer for fusion processing, a low-frequency coefficient fusion rule is adopted for the low-frequency coefficients, a high-frequency coefficient fusion rule is adopted for the high-frequency coefficients, and finally, the UDCT coefficient for each layer after fusion can be acquired; and a reconstructed image obtained after inverse transforma on the UDCT coefficient for each layer after fusion is the fused image.

Description

Visible ray based on UDCT and PCNN and infrared image blending algorithm

Technical field

The present invention relates to technical field of image processing, be specifically related to visible ray based on UDCT and PCNN and melt with infrared image Hop algorithm.

Background technology

Currently, utilizing the visible ray integration technology with infrared image is infrared image enhancement neck to promote infrared image quality The study hotspot in territory, is then the mainstream research direction in this focus based on multiple dimensioned integration technology.Based on multiple dimensioned Integration technology principle is as follows: the first step, by multi-scale transform, source figure is decomposed into a series of subimage, these subimage chis Degree difference, frequency and spatial character are different.Second step, makees to merge meter to the conversion coefficient of source images according to specific fusion rule Calculate.3rd step, will merge after conversion coefficient merged by the inverse transformation of multi-resolution decomposition after high quality graphic.Although Traditional image fusion technology based on multi-resolution decomposition, the integration technology of such as wavelet decomposition can promote to a certain extent Picture quality, but but there is structure complicated, the prominent shortcoming that data redudancy is high.

In order to solve the problems referred to above, the present invention uses based on uniform discrete warp wavelet (Uniform Discrete Curvelet Transform, UDCT) decompose source figure.Uniform discrete warp wavelet introducing Image Fusion is had a lot Advantage.First, original image carries out multi-direction, multiple dimensioned decomposition under UDCT assists, and caters to human world's perception information Mode, can obtain the minutia in original image in all directions, and then supply more efficiently reference for image co-registration Information.Secondly, the translation invariance of UDCT can effectively suppress the harmful effect produced fusion results due to registration accuracy difference.This Outward, UDCT coefficient redundancy is low and realizes simple, and the different coefficient of yardstick possesses spatial coherence, so at fusion rule Selection on can carry out fusion coefficients in mode the most flexibly.

In addition the core link during fusion rule is image co-registration, the quality of fusion rule is by shadow to a great extent Ring the quality of image co-registration.Due to recognize Pulse Coupled Neural Network (Pulse Coupled Neural Network, PCNN) there is pulse granting, capture and the feature of variable threshold value, be consistent with human visual system, so the present invention proposes one Plant and combine uniform discrete warp wavelet and the visible ray of Pulse Coupled Neural Network and infrared image fusion method.

Summary of the invention

For above-mentioned prior art, present invention aim at providing visible ray based on UDCT and PCNN to melt with infrared image Hop algorithm, solves prior art owing to can not judging the similarity of image low-frequency information to be fused, can not judging source images Details enrich degree and the technical problem such as the fusion image that causes is unintelligible, details is inconspicuous.

For reaching above-mentioned purpose, the technical solution used in the present invention is as follows:

Visible ray based on UDCT and PCNN and infrared image blending algorithm, comprise the steps,

Step 1, acquisition visible light source image and infrared radiation source image, be respectively divided visible light source image and infrared radiation source image be Visible images block and infrared image block；

Step 2, visible light source image and infrared radiation source image carrying out uniform discrete warp wavelet, every width source images obtains respectively Obtain one group of different scale and the low frequency sub-band coefficient of different directions and high-frequency sub-band coefficient；

Step 3, calculating visible images block and the difference of Gaussian eigenvalue of infrared image block, further according to difference of Gaussian feature Value, is selected high frequency fusion rule and is merged high-frequency sub-band coefficient by Pulse Coupled Neural Network, it is thus achieved that high frequency merges Coefficient；

Step 4, calculating visible images block and the energy similarity measure eigenvalue of infrared image block, further according to energy similarity measure Eigenvalue and difference of Gaussian eigenvalue, select low frequency fusion rule and merge low frequency sub-band coefficient, it is thus achieved that low frequency merges Coefficient；

Step 5, according to low frequency fusion coefficients and high frequency fusion coefficients, it is thus achieved that fusion coefficients, further according to fusion coefficients, utilize Multi-scale transform inverse transformation carries out image reconstruction, it is thus achieved that fusion image.

In said method, described step 3, comprise the steps,

Step 3.1, calculating visible images block and the difference of Gaussian eigenvalue F of infrared image block_DoG,

$F_{DoG}=g_2-g_1=G_{{\mathrm{\σ}}_2}*I-G_{{\mathrm{\σ}}_1}*I=(G_{{\mathrm{\σ}}_2}-G_{{\mathrm{\σ}}_1})*I$

Wherein, Represent that parameter is σ respectively₁And σ₂Two Gaussian functions Number, I is visible images block or infrared image block；

Step 3.2, according to difference of Gaussian eigenvalue, using equation below as high frequency fusion rule,

$C_{j,l}^F(i,j)=\left\{\begin{array}{cc}{C}_{j,l}^A(i,j),& D_A(i,j)>D_B(i,j)\\ C_{j,l}^B(i,j),& D_A(i,j)<D_B(i,j)\end{array}\right.$

Wherein, with two-dimensional image coordinate (i, j) centered by, D_A(i,j)、D_B(i j) is respectively visible images block and infrared The difference of Gaussian eigenvalue of image block, X takes F, A or B,Corresponding X value is respectively high frequency fusion coefficients, visible light source figure High-frequency sub-band coefficient or the high-frequency sub-band coefficient of infrared radiation source image B as A；

Step 3.3, by Pulse Coupled Neural Network, high-frequency sub-band coefficient is merged, it is thus achieved that high frequency fusion coefficients.

In said method, described step 4, comprise the steps,

Step 4.1, by equation below, calculate visible images block and the energy similarity measure eigenvalue S of infrared image block (i,j)

$S(i,j)=\frac{\min \{E\left(F_A\right(i,j\left)\right),E\left(F_B\right(i,j\left)\right)\}}{max\{E\left(F_A\right(i,j\left)\right),E\left(F_B\right(i,j\left)\right)\}}$

Wherein, F_A(i,j),F_B(i, j) be respectively visible images block and infrared image block, with two-dimensional image coordinate (i, j) Centered by, min{E (F_A(i,j)),E(F_B(i, j)) } it is visible images block and the minimum value function of infrared image block energy, max{E(F_A(i,j)),E(F_B(i, j)) } it is visible images block and the max function of infrared image block energy；

Step 4.2, default similarity threshold, it is judged that (whether i, j) more than similarity threshold for energy similarity measure eigenvalue S；

If (i, j) more than similarity threshold, chooses average rule and merges low frequency step 4.2.1 energy similarity measure eigenvalue S Sub-band coefficients；

If (i, j) less than or equal to similarity threshold, chooses energy the most more to step 4.2.2 energy similarity measure eigenvalue S Big visible images block or infrared image block, and it is composed the biggest weighting；

Step 4.3, combine difference of Gaussian eigenvalue, by updating equation below

$C_{j_0}^F(i,j)={\mathrm{\&alpha;}}_1{C}_{j_0}^A(i,j)+{\mathrm{\&alpha;}}_2{C}_{j_0}^B(i,j)$

For fusion formula

$C_{j0}^F(i,j)\left\{\begin{array}{cc}\frac{C_{j0}^A(i,j)+C_{j0}^B(i,j)}{2},& S(i,j)>T\\ C_{j0}^A(i,j),& D_{j0}^A(i,j)>D_{j0}^B(i,j),S(i,j)<T\\ C_{j0}^B(i,j),& D_{j0}^A(i,j)<D_{j0}^B(i,j),S(i,j)<T\end{array}\right.$

And threshold value T is set, thus obtain low frequency fusion coefficientsWherein, α₁,α₂For weight coefficient, X takes F, A or B, With two-dimensional image coordinate (i, j) centered by,Corresponding X value be respectively low frequency fusion coefficients, visible light source image A low Frequently sub-band coefficients or the low frequency sub-band coefficient of infrared radiation source image B.

Compared with prior art, beneficial effects of the present invention:

The similarity of fusion method coefficient of utilization block analysis low-frequency information of the present invention, if information is similar, then uses simple Averagely, the most just use the Differential Characteristics of Gaussian function to select the details relatively horn of plenty of which source images, and be selected as Fusion image；

In the fusion rule of high-frequency signal, the inventive method has abandoned traditional direct take high-frequency band pass sub-band coefficients The method of absolute value；By introducing Pulse Coupled Neural Network, neighborhood characteristic in Fire mapping image is more significant, i.e. neighborhood is equal The sub-band coefficients of the bigger pixel of value leads to directional subband coefficient as the band of fusion image respective pixel；By drawing in high-frequency signal Entering Pulse Coupled Neural Network, this fusion method can make full use of the characteristic information in infrared image and visible images；We Method can preferably catch visible images and the main body of infrared image and detailed information, and the vision promoting fusion image represents power；

Fusion image indoor setting thing stereovision is strong, and noise is less, substantially eliminates gibbs blocky effect, has preferably Visual effect；Relative to method based on broad sense random walk, the present invention retains the part edge information of image, makes contrast have Improved.

Accompanying drawing explanation

Fig. 1 is the image interfusion method block diagram of the present invention；

Fig. 2 is the infrared and visible images group figure used, and (a) (b) is infrared image, and it is right that (c) (d) is respectively (a) (b) The visible ray figure answered；

Fig. 3 be various method to STOWAGE PLAN as the comparable group figure of fusion results, (a) infrared image (b) visible images (c) base Result (d) result (e) based on Contourlet method fusion method (f) based on UDCT in wavelet method based on broad sense with The result of fusion method (g) the inventive method of machine migration；

Fig. 4 is that STOWAGE PLAN is as the partial enlarged drawing of the fusion results rectangle tab area of visible ray and infrared image pair, figure A () (g) is the partial enlarged drawing of rectangular area in Fig. 3 (a) (g)；

Fig. 5 is the various method comparable group figure to woods image co-registration result, (a) infrared image (b) visible images (c) Result (d) based on wavelet method result (e) based on Contourlet method fusion method (f) based on UDCT is based on broad sense The result of result (g) the inventive method of random walk.

Detailed description of the invention

All features disclosed in this specification, or disclosed all methods or during step, except mutually exclusive Feature and/or step beyond, all can combine by any way.

The present invention will be further described below in conjunction with the accompanying drawings:

Embodiment 1

In view of the premium properties of UDCT conversion, it is applied to herein in the infrared fusion method with visible images.This Algorithm can be summarized as following three steps:

(1) after the source images of visible images and infrared image is carried out respectively UDCT conversion decomposition, available different chis The UDCT sub-band coefficients of degree different directions.It comprises the UDCT coefficient of low frequency and high frequency.

(2) according to specific rule, each scale layer is carried out fusion treatment in different ways, i.e. low frequency coefficient is adopted Use low frequency coefficient fusion rule, and use high frequency coefficient fusion rule to process high frequency coefficient, finally can obtain each layer after fusion UDCT coefficient.

(3) the reconstruct image that each layer UDCT coefficient after inverse transformation is merged obtains is fusion image.

The fusion rule that algorithm is used

Selecting fusion rule directly to determine the quality of image co-registration, therefore, fusion rule is in the process of image co-registration Among extremely important, it is also the focus in image studies simultaneously.Under normal circumstances, have different from HFS in view of low frequency Characteristic, therefore to these two parts use different fusion rules process.

Low frequency fusion rule

Low frequency signal in image mainly main information and image detail information that some are less by image forms.For In fusion image, embody infrared image and the main information of visible images, traditional method be application weighted average or Low frequency part is merged by the method for person's simple average:

$C_{j_0}^F(i,j)={\mathrm{\&alpha;}}_1{C}_{j_0}^A(i,j)+{\mathrm{\&alpha;}}_2{C}_{j_0}^B(i,j)$

Wherein, α₁,α₂For weight coefficient, work as α₁=α₂It it is simple average when=0.5.

But if the low-frequency information between source images differs greatly, the fusion results obtained by simple average will less Ideal, fuzzyyer features.In order to overcome this problem, first this algorithm analyzes the low-frequency information judging source images: if Source images low-frequency information to be fused is the most similar, then use simple average mode；Consider that image co-registration is intended to obtain the most clear The fusion image that clear and details is the most various.So, if source images low-frequency information to be fused is not much like, just choose details Information is fusion image compared with the source images of horn of plenty.

For realizing above-mentioned target, be necessary for solve following two problem:

1) how to judge that the low-frequency information merging source images is dissimilar？

2) how to judge that the details of which source images is enriched？

For first problem, the most how to judge that the low-frequency information merging source images is dissimilar, following methods solution can be used Certainly.

The information that the information comprised due to image (coefficient) block represents far more than pixel (coefficient), and be less susceptible to make an uproar The impact of sound, so the similarity of coefficient of utilization block analysis low-frequency information.For the difference between quick design factor block, can be straight Connecing the energy of design factor block, then the difference between analysing energy calculates similarity.This algorithm uses the following degree of approximation (also Be referred to as similarity) function S (.) judge merge source images low-frequency information the most similar:

$S(i,j)=\frac{min\{E\left(F_A\right(i,j\left)\right),E\left(F_B\right(i,j\left)\right)\}}{max\{E\left(F_A\right(i,j\left)\right),E\left(F_B\right(i,j\left)\right)\}}$

Wherein, F_A(i,j),F_B(i, j) be respectively source images A, in B with (i, j) centered by a coefficient block, coefficient block Be dimensioned to 3 × 3；E (.) is the energy of design factor block.S (i, span j) can be substantially analyzed from above formula For [0,1].(i, value j) is the highest, and the low-frequency information of source images A, B is the most similar for S；(i, value j) is the lowest, and source images A, B's is low for S Frequently information gap is the biggest.If the low-frequency information of source images A, B is similar, i.e. S (.) is more than threshold value T set in advance, can adopt Merge by average rule；If the low-frequency information similarity of source images is low, i.e. S (.) is less than a threshold value set in advance T, will choose and have the coefficient that energy is bigger.This is because this coefficient comprises more details, and fusion image expectation comprises more Details.

For Second Problem, how to judge the details more horn of plenty of which source images, the difference of Gaussian function can be used Details is measured by feature (Difference of Gaussian, DOG), and DoG characteristic energy is the biggest, illustrates that details is the richest Richness, more should be chosen as fusion image.DOG is the difference of Gaussian function, it be two different parameters gaussian filtering after Image subtraction.

Gaussian filter function is expressed as:

$G_{\mathrm{\&sigma;}}(x,y)=\frac{1}{\sqrt{2{\mathrm{\&pi;\&sigma;}}^2}}\exp (-\frac{x^2+y^2}{2{\mathrm{\&sigma;}}^2})$

The gaussian filtering of two different parameters of image I is:

$g_1=G_{{\mathrm{\&sigma;}}_1}*I$

$g_2=G_{{\mathrm{\&sigma;}}_2}*I$

DoG character representation is

$F_{DoG}=g_2-g_1=G_{{\mathrm{\&sigma;}}_2}*I-G_{{\mathrm{\&sigma;}}_1}*I=(G_{{\mathrm{\&sigma;}}_2}-G_{{\mathrm{\&sigma;}}_1})*I$

Wherein,Represent that parameter is σ respectively₁And σ₂Two Gaussian functions, I is pending image；g₂,g₁Respectively ForTo the image after pending image gaussian filtering.F_DoGFor the DoG feature obtained.

The information represented far more than pixel (coefficient) due to the information of image (coefficient) block representative.Therefore, when calculating energy Calculate in units of coefficient block.Multiple image block will be divided into by source images, then calculate the DoG feature of image block, choose spy The coefficient of the image levying the image block of maximum is the coefficient after merging.Region dissimilar to source images low-frequency information, merges rule Then it is expressed as:

$C_{j,l}^F(i,j)=\left\{\begin{array}{cc}{C}_{j,l}^A(i,j),& D_A(i,j)>D_B(i,j)\\ C_{j,l}^B(i,j),& D_A(i,j)<D_B(i,j)\end{array}\right.$

Wherein, D_A(i is j) so that (i, j) center carry out the energy of the filtered image block of DoG to image.

In sum, the coefficient fusion rule of low frequency is embodied as:

$C_{j0}^F(i,j)\left\{\begin{array}{cc}\frac{C_{j0}^A(i,j)+C_{j0}^B(i,j)}{2},& S(i,j)>T\\ C_{j0}^A(i,j),& D_{j0}^A(i,j)>D_{j0}^B(i,j),S(i,j)<T\\ C_{j0}^B(i,j),& D_{j0}^A(i,j)<D_{j0}^B(i,j),S(i,j)<T\end{array}\right.$

Wherein, T is a threshold value set in advance.D is for carry out the filtered image block of DoG to image, so that (i, in j) being The energy of the heart.

High frequency fusion rule

Fusion rule for high fdrequency component, it is common practice to the big value that takes absolute value high-frequency band pass sub-band coefficients is carried out Fusion treatment, because in general, coefficient is the biggest, and to represent high-frequency information the abundantest:

$C_{j,l}^F(i,j)=\left\{\begin{array}{cc}{C}_{j,l}^A(i,j),& C_A(i,j)>C_B(i,j)\\ C_{j,l}^B(i,j),& C_A(i,j)<C_B(i,j)\end{array}\right.$

The algorithm of the present invention obtains UDCT conversion coefficient, then as PCNN by source images carries out UDCT conversion The input of neuron, inputs into PCNN neutral net.I.e. UDCT conversion coefficient and PCNN input neuron exist therebetween by One corresponding relation.Meanwhile, each neuron is linked to each other with neighbouring several neurons in link field, there are two kinds of states Neuron exports, namely igniting and non-ignition.Thus, the igniting corresponding with the summation of a neuron firing number of times distribution Matrix is generated as.

The each sub-band coefficients obtained by UDCT inputs PCNN respectively, obtains each self-corresponding point via the iteration of same number of times Fire mapping graph.Compare the ignition times at Fire mapping image respective pixel again, so that it may whether the target judging this position is high Frequently details area.Thus when image co-registration, the UDCT conversion coefficient of source images is inputted PCNN, by each neuron firing The Fire mapping images that number generates just can efficiently extract in conversion coefficient figure the high-frequency informations such as the edge of correspondence, texture.Further, The neuron firing number of times that a certain conversion coefficient point is corresponding is the most, then in explanatory diagram picture, the information of this point is the abundantest.

The band of fusion image respective pixel leads to directional subband coefficient, and can directly to take neighborhood characteristic in Fire mapping image more significant, The i.e. sub-band coefficients of the bigger pixel of neighboring mean value.

If the transform coefficient bits of source images A, B is set to, (i, j) corresponding PCNN neuron firing number of times is expressed as P_A (i, j), P_B(i, j), high frequency fusion rule is:

$C_{j,l}^F(i,j)=\left\{\begin{array}{cc}{C}_{j,l}^A(i,j),& P_A(i,j)>P_B(i,j)\\ C_{j,l}^B(i,j),& P_A(i,j)<P_B(i,j)\end{array}\right.$

Based on UDCT conversion and the blending algorithm flow process of PCNN

Image co-registration schematic diagram based on UDCT and PCNN is as shown in Figure 1.The basic step merged is:

Converted by UDCT, decompose source images A and B, it is thus achieved that the UDCT sub-band coefficients that yardstick is different and direction is different.It UDCT coefficient comprises thick yardstick and top thin yardstick.Except top, other each thin scale layer all forms in different directions UDCT coefficient；Use different fusion codes that each scale layer is carried out fusion treatment.

Embodiment 2

For verifying the effectiveness of algorithm, experiment uses the infrared and visible ray figure of the two groups of same scene registrated Picture.Experimental image is as shown in Figure 2.The method that the present invention proposes becomes with method based on wavelet transformation with based on un-downsampling wavelet transform The method changed, method based on contourlet transformation, method based on UDCT conversion and side based on broad sense random walk Method compares.In experiment, what wavelet transformation and undecimated wavelet transform used is all 3 layers of decomposition, and uses Harr small echo Basic function, uses greatest coefficient rule to merge high frequency coefficient, uses average rule to merge low frequency coefficient.

The different fusion method of application is to the experimental result of STOWAGE PLAN picture in Fig. 2 as shown in Figure 3.Fig. 3 (a) and Fig. 3 (b) is respectively The infrared image of ship and visible images；Fig. 3 (c) is result based on wavelet method；Fig. 3 (d) is based on Contourlet side The result of method；Fig. 3 (e) is fusion results based on UDCT；Fig. 3 (f) is result based on broad sense random walk method；Fig. 3 (g) The result of method is proposed for the present invention.From subjective vision effect, various blending algorithms all can be protected to varying degrees Stay infrared image and the main information in visible images.

The various method of Fig. 3 to STOWAGE PLAN as the comparison of fusion results

By relatively each fusion image it is found that the scenery of fusion results based on small echo lacks stereovision, and make an uproar Sound is relatively big, there is obvious gibbs blocky effect；Fusion results based on contourlet transformation and fusion based on small echo Result effect is similar to.Examine and can find that fusion results based on contourlet transformation is slightly better than fusion based on small echo knot Really, this is owing to contourlet transformation has ratio wavelet transformation better performance.The result of fusion method based on UDCT is excellent In the first two result, this is owing to UDCT conversion is better than wavelet transformation and contourlet transformation.And based on broad sense random walk Method several method earlier above more novel, its fused image is the most smooth, and whole structure is better than front several method.Can see Go out the syncretizing effect stereovision of method that the present invention proposes higher, there is more preferable visual effect, be substantially better than on the whole based on The fusion method of small echo, fusion method based on Contourlet and fusion method based on UDCT.And swim at random based on broad sense Although the result of the method walked seems more smooth than the result of context of methods, but smear the part edge information of the image that goes out, Its contrast is made to decrease.

The comparison to woods image co-registration result of Fig. 5 various method

In order to observe the visual effect of inventive algorithm further, rectangle tab area in Fig. 3 is amplified, amplifies After image as shown in Figure 4.It can be seen that be clearly present some mosaics based on small echo and method based on Contourlet, And gibbs blocky effect is the most serious.Fusion method based on UDCT can overcome certain blocky effect, but exists Certain noise.And method entirety of based on broad sense random walk is the most smooth, lost part marginal information.And calculation herein Method can preferably overcome these problems, has relatively sharp visual effect.This is mainly the most not only with of good performance UDCT converts, and uses PCNN to carry out the fusion of high frequency coefficient, and the syncretizing effect therefore obtained is more satisfactory.

Fig. 5 is the comparison to woods image co-registration result of the various method, examines and can draw the conclusion identical with Fig. 3. Algorithm obtains best visual effect compared to other method i.e. herein.

The objective evaluation Indexes Comparison of table 1 algorithm overall performance test

In order to analyze experimental result objectively, it is employed herein QSSIM, mutual information and tri-objective appraisals of QAB/F and refers to Number analyzes fusion results.Mutual information is the reflection to the quantity of information that fusion image obtains from source images, and value shows the most greatly to melt The effect closed is the best；QAB/F is that the edge to fusion image enriches degree and reflects, its span is 0 to 1, and value is more Show that greatly the effect merged is the best.QSSIM is picture structure phase knowledge and magnanimity index, represents the relatedness of neighbor, its value Scope is-1 to 1, and value is the biggest represents that the signal merged more is known each other.As shown in table 1 to the objective evaluation of distinct methods fusion results.

Although from table 1 it follows that in QSSIM index the method for the present invention with compared with method be respectively arranged with quality, But from other two indexs: fusion method based on small echo differs with fusion method performance based on Contourlet Not quite, performance is the most relatively low comparatively speaking.Method performance based on UDCT be substantially better than fusion method based on small echo with based on The fusion method of Contourlet.Relatively new method based on broad sense random walk is again much better than front several method.And this Though inventive method is slightly poorer than method based on broad sense random walk in this index of mutual information, but in edge richness Being substantially better than it, this is consistent with subjective vision effect.

The MI merging contrast experiment is compared by the extra five groups of images of table 2

The QAB/F merging contrast experiment is compared by the extra five groups of images of table 3

For analyzing the performance of inventive algorithm more objectively, table 2,4-3 and 4-4 illustrate other five groups of registered imageses pair Contrast and experiment.Data from these tables also confirm above-mentioned experimental analysis, although the method that i.e. present invention is carried is mutually Have both advantages and disadvantages with the method contrasted in information and structural similarity, but but there is preferable edge conservation degree, thus merge After image object feature more clear, subjective vision effect is preferable.But inventive algorithm is not examined in fusion rule part Consider image-region importance, thus there is also the probability promoting syncretizing effect further.

The QSSIM merging contrast experiment is compared by the extra five groups of images of table 4

The present invention proposes a kind of new infrared image enhancing method, i.e. based on uniform discrete warp wavelet and pulse-couple god Visible ray and infrared image fusion method through network.Being unique in that of this method:

Make full use of uniform discrete warp wavelet characteristic multiple dimensioned, multidirectional to decompose source images, more efficiently catch Having received the minutia information of source images, more preferable basis has been laid in the fusion for next step.

Low frequency signal is used simple average to obtain fusion results by traditional fusion method, although its method is easy to use, But effect is the most not satisfactory, easily produce the situation that main body is fuzzy.For overcoming this problem, the fusion method of present invention research The similarity of coefficient of utilization block analysis low-frequency information, if information is similar, then uses simple average；The most just Gaussian function is used Differential Characteristics (DOG) select the details relatively horn of plenty of which source images, and be selected as fusion image.At high-frequency signal Fusion rule in, the inventive method has abandoned traditional direct method taking absolute value high-frequency band pass sub-band coefficients.Pass through Introduce Pulse Coupled Neural Network, neighborhood characteristic in Fire mapping image is more significant, i.e. the subband of the bigger pixel of neighboring mean value Coefficient leads to directional subband coefficient as the band of fusion image respective pixel.By introducing pulse coupled neural net in high-frequency signal Network, this fusion method can make full use of the characteristic information in infrared image and visible images.

From the point of view of experimental result, the method that the present invention proposes can preferably catch the main body of visible images and infrared image And detailed information, the vision promoting fusion image represents power；, analyzing from the objective evaluation data of experiment, the present invention carries meanwhile The fusion method gone out is also superior to traditional method.

The above, the only detailed description of the invention of the present invention, but protection scope of the present invention is not limited thereto, and any Belong to those skilled in the art in the technical scope that the invention discloses, the change that can readily occur in or replacement, all answer Contain within protection scope of the present invention.

Claims (3)

1. visible ray based on UDCT and PCNN and infrared image blending algorithm, it is characterised in that comprise the steps,

Step 1, acquisition visible light source image and infrared radiation source image, be respectively divided visible light source image and infrared radiation source image be visible Light image block and infrared image block；

Step 2, visible light source image and infrared radiation source image carrying out uniform discrete warp wavelet, every width source images obtains one respectively Organize different scale and the low frequency sub-band coefficient of different directions and high-frequency sub-band coefficient；

Step 3, calculating visible images block and the difference of Gaussian eigenvalue of infrared image block, further according to difference of Gaussian eigenvalue, Select high frequency fusion rule and by Pulse Coupled Neural Network, high-frequency sub-band coefficient merged, it is thus achieved that high frequency merges system Number；

Step 4, calculating visible images block and the energy similarity measure eigenvalue of infrared image block, further according to energy similarity measure feature Value and difference of Gaussian eigenvalue, select low frequency fusion rule and merge low frequency sub-band coefficient, it is thus achieved that low frequency fusion coefficients；

Step 5, according to low frequency fusion coefficients and high frequency fusion coefficients, it is thus achieved that fusion coefficients, further according to fusion coefficients, utilize many chis Degree conversion inverse transformation carries out image reconstruction, it is thus achieved that fusion image.

Visible ray based on UDCT and PCNN the most according to claim 1 and infrared image blending algorithm, it is characterised in that Described step 3, comprises the steps,

Step 3.1, calculating visible images block and the difference of Gaussian eigenvalue F of infrared image block_DoG,

$F_{DoG}=g_2-g_1=G_{{\mathrm{\&sigma;}}_2}*I-G_{{\mathrm{\&sigma;}}_1}*I=(G_{{\mathrm{\&sigma;}}_2}-G_{{\mathrm{\&sigma;}}_1})*I$

Wherein, Represent that parameter is σ respectively₁And σ₂Two Gaussian functions, I is Visible images block or infrared image block；

Step 3.2, according to difference of Gaussian eigenvalue, using equation below as high frequency fusion rule,

$C_{j,l}^F(i,j)=\left\{\begin{array}{cc}{C}_{j,l}^A(i,j),& D_A(i,j)>D_B(i,j)\\ C_{j,l}^B(i,j),& D_A(i,j)<D_B(i,j)\end{array}\right.$

Wherein, with two-dimensional image coordinate (i, j) centered by, D_A(i,j)、D_B(i j) is respectively visible images block and infrared image The difference of Gaussian eigenvalue of block, X takes F, A or B,Corresponding X value is respectively high frequency fusion coefficients, visible light source image A High-frequency sub-band coefficient or the high-frequency sub-band coefficient of infrared radiation source image B；

Step 3.3, by Pulse Coupled Neural Network, high-frequency sub-band coefficient is merged, it is thus achieved that high frequency fusion coefficients.

Visible ray based on UDCT and PCNN the most according to claim 2 and infrared image blending algorithm, it is characterised in that Described step 4, comprises the steps,

Step 4.1, by equation below, calculate visible images block and infrared image block energy similarity measure eigenvalue S (i, j)

$S(i,j)=\frac{min\{E\left(F_A\right(i,j\left)\right),E\left(F_B\right(i,j\left)\right)\}}{max\{E\left(F_A\right(i,j\left)\right),E\left(F_B\right(i,j\left)\right)\}}$

Wherein, F_A(i,j),F_B(i, j) is respectively visible images block and infrared image block, and with two-dimensional image coordinate, (i, in j) being The heart, min{E (F_A(i,j)),E(F_B(i, j)) } it is visible images block and the minimum value function of infrared image block energy, max{E (F_A(i,j)),E(F_B(i, j)) } it is visible images block and the max function of infrared image block energy；

Step 4.2, default similarity threshold, it is judged that (whether i, j) more than similarity threshold for energy similarity measure eigenvalue S；

If (i, j) more than similarity threshold, chooses average rule and merges low frequency sub-band step 4.2.1 energy similarity measure eigenvalue S Coefficient；

If (i, j) less than or equal to similarity threshold, chooses energy the biggest to step 4.2.2 energy similarity measure eigenvalue S Visible images block or infrared image block, and it is composed the biggest weighting；

Step 4.3, combine difference of Gaussian eigenvalue, by updating equation below

$C_{j_0}^F(i,j)={\mathrm{\&alpha;}}_1{C}_{j_0}^A(i,j)+{\mathrm{\&alpha;}}_2{C}_{j_0}^B(i,j)$

For fusion formula

$C_{j0}^F(i,j)\left\{\begin{array}{cc}\frac{C_{j0}^A(i,j)+C_{j0}^B(i,j)}{2},& S\left(i,j\right)>T\\ C_{j0}^A(i,j),& D_{j0}^A(i,j)>D_{j0}^B(i,j),S(i,j)<T\\ C_{j0}^B(i,j),& D_{j0}^A(i,j)<D_{j0}^B(i,j),S(i,j)<T\end{array}\right.$

And threshold value T is set, thus obtain low frequency fusion coefficientsWherein, α₁,α₂For weight coefficient, X takes F, A or B, to scheme As two-dimensional coordinate (i, j) centered by,Corresponding X value is respectively low frequency of low frequency fusion coefficients, visible light source image A With coefficient or the low frequency sub-band coefficient of infrared radiation source image B.