CN103679670A - A PCNN multisource image fusion method based on an improved model - Google Patents

Abstract

The invention relates to a PCNN multisource image fusion method based on an improved model. Improved content comprises that feedback inputs of various neurons in a PCNN just receive external stimulus inputs; that values of various parameters in a link domain are same to all neurons; that the values of various parameters in a variable threshold function are same to all neurons; that a threshold value look-up table and an index plan are introduced, wherein the threshold value look-up table records threshold values corresponding to network operation times, and the threshold values can be obtained by calculation prior to the network operates such that exponent arithmetic in network operation is prevented and the operation of the network is accelerated; the index plan records the ignition time of all pixels, is an integration result of spatially-adjacent similar pixels in an input image, and embodies an overall visual characteristic of the input image. The method introduces the index plan recording the ignition time of all pixels and the threshold value look-up table recording threshold values corresponding to network operation times, employs a fusion rule based on the index plan, and achieves an effect better than a conventional wavelet transform fusion method.

Description

A kind of PCNN multisource image anastomosing method based on improved model

Technical field

The present invention relates to a kind of PCNN multisource image anastomosing method based on improved model, particularly relate to a kind of PCNN multisource image anastomosing method based on improved model that visible ray, medium wave and three wave bands of LONG WAVE INFRARED merge simultaneously that is suitable for.

Background technology

Artificial neural network is a kind of novel computing model of attempting to imitate biological nervous system information processing manner.A neural network is comprised of Multilevel method unit or node, can adopt the whole bag of tricks to carry out interconnected.Oneself carries out multi-source image fusion through using artificial neural networks some scholar.At present, the application of neural network in image co-registration mainly contains: bimodal neuroid (Bimodal Neurons), multilayer perceptron (Multi-layered Perceptron) and Pulse Coupled Neural Network (Pulse-coupled Neural Network, PCNN) etc.Wherein PCNN is a kind of new neural network proposing in recent years, is referred to as in the world third generation artificial neural network.

1981, E.A.Newman, P.H.Hartline etc. have proposed 6 kinds of dissimilar bimodal neurons and (have comprised AND, OR, Visible-Enhanced Infrared, Visible-Suppressed-Infrared, Infrared-Enhanced-Visible and Infrared-Suppressed-Visible) for the fusion of visible ray and infrared image.Nineteen ninety-five, Fechner and Godlewski have proposed the image interfusion method based on multilayer perceptron neural network.By interested pixel in training multilayer perceptron identification FLIR (Forward-Looking Infrared) image, incorporated in visible images.Since the nineties in 20th century, the research by Eckhorn etc. to the visual cortex nerve impulse string synchronized oscillation phenomenon of cat, monkey, has obtained mammalian nervous meta-model, and development has formed Pulse-coupled Neural Network Model thus.This model has the advantages that the pixel similar to image two-dimensional space, gray scale is similar is divided into groups, and can reduce image local gray scale difference value, makes up the small interruption of image local.1999, BrocssardR.P. etc. proved the relation of the neuronic spark rate of PCNN and gradation of image, have confirmed the feasibility of PCNN for image co-registration.Based on this model, relevant scholar has proposed various improved models, and uses it for the fusion of various images.

At present, main the following aspects of concentrating of the research of the image interfusion method based on PCNN:

The robotization of network parameter is chosen: because the parameter that PCNN network relates to is more, and different parameters value all can affect final result.By householder method, automatically calculate the key parameter in PCNN network, can obtain better result.

Improvement to PCNN basic model: according to practical function, process the difference of object and the mode of thinking, different researchers have successively proposed different improved models.

Therefore need badly a kind of novel PCNN multisource image anastomosing method based on improved model is provided.

Summary of the invention

The technical problem to be solved in the present invention is to provide a kind of raising multi-source image syncretizing effect, makes fused image target signature more obviously, more be conducive to the PCNN multisource image anastomosing method based on improved model of target identification.

For solving the problems of the technologies described above, a kind of PCNN multisource image anastomosing method based on improved model of the present invention, comprises the following steps successively:

Step 1, three width original image A, B and the C of input are spatially carried out to Pixel-level registration, guarantee that three width image sizes are X * Y;

Step 2, setting network parameter W, V _l, β, V _θ, α _θvalue with Δ t;

V _land V _θbe respectively L _ij[n] and θ _ijintrinsic electromotive force in [n], θ _ij[n] is dynamic threshold, L _ij[n] is the linear input that connects;

α _θfor θ _ijthe damping time constant of [n]; Δ t is time sampling interval; β is strength of joint constant between cynapse; Y _ij[n] is PCNN pulse output; Y _kl[n-1] is the last pulse output of PCNN; W in inner connection matrix W _ijklcorresponding L _ijy in [n] _klthe weighting coefficient of [n-1]; N is the number of run of network, n=1, and 2 ..., N-1, N, N is maximum number of run;

Step 3, in every width input picture, search S _{ij_max}, S _{ij_min}; S _{ij_max}< V _θ, S _{ij_min}> 0;

Step 4, obtain the maximum number of run N of network and threshold value look-up table LT (s), s is the function variable of LT (s);

$N=\frac{t_2-t_1}{\mathrm{\&Delta;t}}+1$

$t_1=\frac{1}{{\mathrm{\&alpha;}}_{\mathrm{\&theta;}}}\ln \left[\frac{V_{\mathrm{\&theta;}}}{S_{\mathrm{ij}\_\max }}\right]$

$t_2=\frac{1}{{\mathrm{\&alpha;}}_{\mathrm{\&theta;}}}\ln \left[\frac{V_{\mathrm{\&theta;}}}{S_{\mathrm{ij}\_\min }}\right]$

$\mathrm{LT}\left(s\right)=V_{\mathrm{\&theta;}}{e}^{(-(-\mathrm{s\&Delta;t}+t_2){\mathrm{\&alpha;}}_{\mathrm{\&theta;}})}$

In formula: t ₁and t ₂be respectively the autogenous ignition time of gray-scale value maximum pixel and minimum pixel in image;

Step 5, utilize following formula moving model;

F _ij[n]＝S _ij

L _ij[n]＝V _L∑w _ijklY _kl[n-1]

U _ij[n]＝F _ij[n](1+βL _ij[n])

I _ij[n]＝N-n

In formula: U _ij[n] is internal activity item, Y _ij[n] is PCNN pulse output, I _ij[n] is index value; When n=1, L _ij[1]=0, U _ij[1]=F _ij[1]=S _ij, θ _ij[1]=LT (N-1)=S _{ij_max}, corresponding feed back input intermediate value is S _{ij_max}neuron by autogenous ignition; After neuron firing, output Y _ij[1]=1, θ _ij[2] become V _θ, the neuronic index value of lighting a fire is labeled as I _ij=N-1;

By that analogy, when n=N is arrived in the network operation, threshold value θ _ij[N]=LT (0)=S _{ij_min}, for feed back input, be S _{ij_min}neuron autogenous ignition, the neuronic index value of lighting a fire is labeled as I _ij=0;

Step 6, obtain respectively the key map I of three width original image A, B and C _a, I _band I _c;

Work as I _a, I _band I _cthe absolute value of the mutual difference of index value of respective pixel is all less than or equal to representative value e, and the pixel value of fused images is got the weighted mean value of three width image respective pixel;

Work as I _a, I _band I _cwhen the absolute value of the mutual difference of index value of respective pixel has the situation that is greater than representative value e, if wherein the mutual difference of index value of two width image respective pixel is less than or equal to representative value e, the pixel value of fused images is got the weighted mean value of above-mentioned two width image respective pixel;

In other situations, the pixel value of fused images is got the pixel value that index value is larger.

In step 2, guarantee the feed back input F of dendron _ij[n] receives only outside input stimulus signal S _ij; Guarantee W, V _l, β, V _θ, α _θall identical to all neurons with the value of Δ t.

e=2。

Three width original image A, B and C are respectively visible ray, medium-wave infrared and LONG WAVE INFRARED image.

The present invention, analyzing on PCNN image interfusion method basis, simplifies accordingly and improves basic model, obtains a kind of new improved PCNN image interfusion method.Wherein simplify and improve in have: 1. in PCNN, each neuronic feed back input receives only outside stimulus input; 2. in link field, the value of each parameter is all identical to all neurons; 3. in variable threshold value function, the value of each parameter is all identical to all neurons; 4. introduce threshold value look-up table and key map, threshold value look-up table has recorded the threshold value corresponding with network operation number of times, and these threshold values can calculate in advance before the network operation, has avoided the exponent arithmetic in the network operation, has accelerated the operation of network.Key map has recorded the duration of ignition of whole pixels, is the integrated results of the adjacent similar pixel in space in input picture, embodiment be the whole visual signature of input picture.

The present invention has introduced key map and the record threshold value look-up table corresponding with network operation number of times of the duration of ignition of recording whole pixels in improved model, under identical fusion rule condition, than traditional wavelet transform fusion, has better effect.

Indices of the present invention is better than the indices of the former fusion rule of WT+ to a certain extent, and the index such as average and standard variance particularly improves very obviously, and the validity of improving one's methods has been described.

Accompanying drawing explanation

Fig. 1 is the PCNN fusion method schematic diagram of three width images.

Fig. 2 is certain seashore original image and the fusion results based on PCNN.

Embodiment

Below in conjunction with drawings and Examples, the present invention is further detailed explanation.

Basic thought of the present invention is: two width or several original images to input adopt respectively PCNN model to calculate corresponding key map, then key map and original images by using are merged to decision-making accordingly, finally obtain fused images.The PCNN fusion method schematic diagram of three width images as shown in Figure 1.

Specifically, the present invention comprises the following steps successively:

Step 1, three width original image A, B and the C of input are spatially carried out to Pixel-level registration, guarantee that three width image sizes are X * Y; Three width original image A, B and C are respectively visible ray, medium-wave infrared and LONG WAVE INFRARED image;

Step 2, setting network parameter W, V _l, β, V _θ, α _θvalue with Δ t; Guarantee the feed back input F of dendron _ij[n] receives only outside input stimulus signal S _ij; Guarantee each parameter W, V in link field _l, β value all identical to all neurons; Guarantee each V parameter in variable threshold value function _θ, α _θall identical to all neurons with the value of Δ t;

V _land V _θbe respectively L _ij[n] and θ _ijintrinsic electromotive force (amplification coefficient) in [n], θ _ij[n] is dynamic threshold, L _ij[n] is the linear input that connects;

α _θfor θ _ijthe damping time constant of [n]; Δ t is time sampling interval; β is strength of joint constant between cynapse; Y _ij[n] is PCNN pulse output; Y _kl[n-1] is the last pulse output of PCNN; W in inner connection matrix W _ijklcorresponding L _ijy in [n] _klthe weighting coefficient of [n-1];

N is the number of run of network, n=1, and 2 ..., N-1, N, N is maximum number of run;

Step 3, in every width input picture, search S _{ij_max}, S _{ij_min}; V _θ> S _{ij_max}; If S _{ij_min}be less than or equal to 0, by linearity, adjust S _{ij_max}, S _{ij_min}and input picture, make S _{ij_min}be greater than 0;

Step 4, obtain the maximum number of run N of network and threshold value look-up table LT (s), s is the function variable of LT (s);

$N=\frac{t_2-t_1}{\mathrm{\&Delta;t}}+1$

$t_1=\frac{1}{{\mathrm{\&alpha;}}_{\mathrm{\&theta;}}}\ln \left[\frac{V_{\mathrm{\&theta;}}}{S_{\mathrm{ij}\_\max }}\right]$

$t_2=\frac{1}{{\mathrm{\&alpha;}}_{\mathrm{\&theta;}}}\ln \left[\frac{V_{\mathrm{\&theta;}}}{S_{\mathrm{ij}\_\min }}\right]$

$\mathrm{LT}\left(s\right)=V_{\mathrm{\&theta;}}{e}^{(-(-\mathrm{s\&Delta;t}+t_2){\mathrm{\&alpha;}}_{\mathrm{\&theta;}})}$

In formula: t ₁and t ₂be respectively the autogenous ignition time of gray-scale value maximum pixel and minimum pixel in image;

Step 5, utilize following formula moving model;

F _ij[n]＝S _ij

L _ij[n]＝V _L∑w _ijklY _kl[n-1]

U _ij[n]＝F _ij[n](1+βL _ij[n])

I _ij[n]＝N-n

In formula: U _ij[n] is internal activity item, Y _ij[n] is PCNN pulse output, I _ij[n] is index value;

When n=1, L _ij[1]=0, U _ij[1]=F _ij[1]=S _ij, θ _ij[1]=LT (N-1)=S _{ij_max}, corresponding feed back input intermediate value is S _{ij_max}neuron by autogenous ignition; After neuron firing, output Y _ij[1]=1, θ _ij[2] become V _θ, the neuron that makes to light a fire will can not lighted a fire again.Meanwhile, the neuronic index value of lighting a fire is labeled as I _ij=N-1, no longer changes.Along with the increase of the number of run n of network, threshold value θ _ij[n] reduces gradually, and the neuron of igniting will encourage adjacent neuron by link field.When n=N is arrived in the network operation, threshold value θ _ij[N]=LT (0)=S _{ij_min}, for feed back input, be S _{ij_min}neuron, even if there is no adjacent neurons excitation in the situation that, also will autogenous ignition, the neuronic index value of lighting a fire is labeled as I _ij=0; By the network operation of maximum N time, all neurons will be lighted a fire and only once.

Can find out, if neuron S _ijwhen the n time operation, light a fire, its key map I _ij[n] is fixed as N-n, and no longer the operation along with network changes.Key map I _ij[n] recorded all neuronic durations of ignition, is the result that input picture space-time is integrated.In addition, because the variable threshold value in the network operation is to obtain by look-up table mode, do not need complicated exponent arithmetic, shortened the time of the network operation.The automatic acquisition of network operation maximum times, has not only guaranteed the determinacy of the network operation, and makes each neuron firing only once;

Step 6, adopt corresponding convergence strategy, obtain fused images.The convergence strategy that this method adopts is as follows: because the key map of every width original image has represented the whole visual signature of original image, so carry out synthetic setting convergence strategy by the key map of three width original images.

Obtain the key map I of three width original image A, B and C _a, I _band I _c;

Work as I _a, I _band I _cthe absolute value of the mutual difference of index value of respective pixel is all less than or equal to representative value e, and the pixel value of fused images is got the weighted mean value of three width image respective pixel;

Work as I _a, I _band I _cwhen the absolute value of the mutual difference of index value of respective pixel has the situation that is greater than representative value e, if wherein the mutual difference of index value of two width image respective pixel is less than or equal to representative value e, the pixel value of fused images is got the weighted mean value of above-mentioned two width image respective pixel;

In other situations, the pixel value of fused images is got the pixel value that index value is larger.

Preferred e=2.

In order to verify the performance based on improving PCNN model image fusion method of proposition, selected three width visible rays, LONG WAVE INFRARED and the medium-wave infrared image through registration of certain seashore (image size is 320 * 256) as treating fused images.The (a) and (b) of Fig. 2 and (c) be respectively primary visible light to be merged, LONG WAVE INFRARED and the medium-wave infrared image of certain seashore, (d) of Fig. 2 adopts 4 layers of wavelet transformation (WT), low frequency coefficient is averaged, the fusion results that high frequency coefficient adopts the fusion rule of region energy operator to obtain; (e) of Fig. 2 adopts the fusion results of improving PCNN model.In order better fusion results to be analyzed relatively, adopted the product of average, standard deviation, entropy and structural information and transinformation content

objective evaluation criteria calculate, the result obtaining is as shown in table 1.Can find out, the fusion results based on improving PCNN model is better than the fusion results based on the former fusion rule of WT+ on local detail, profile (personage on seashore limit) and overall brightness.From three tables of objective evaluation, can find out, the indices based on improving PCNN model is better than the indices, the particularly index such as average and standard variance of the former fusion rule of WT+ to a certain extent, improves very obviously, and the validity of improving one's methods has been described.

Table 1 certain seashore fusion results objective evaluation based on PCNN

Claims (4)

1. the PCNN multisource image anastomosing method based on improved model, comprises the following steps successively:

Step 1, three width original image A, B and the C of input are spatially carried out to Pixel-level registration, guarantee that three width image sizes are X * Y;

Step 2, setting network parameter W, V _l, β, V _θ, α _θvalue with Δ t;

V _land V _θbe respectively L _ij[n] and θ _ijintrinsic electromotive force in [n], θ _ij[n] is dynamic threshold, L _ij[n] is the linear input that connects;

α _θfor θ _ijthe damping time constant of [n]; Δ t is time sampling interval; β is strength of joint constant between cynapse; Y _ij[n] is PCNN pulse output; Y _kl[n-1] is the last pulse output of PCNN; W in inner connection matrix W _ijklcorresponding L _ijy in [n] _klthe weighting coefficient of [n-1];

N is the number of run of network, n=1, and 2 ..., N-1, N, N is maximum number of run;

Step 3, in every width input picture, search S _{ij_max}, S _{ij_min}; S _{ij_max}< V _θ, S _{ij_min}> 0;

Step 4, obtain the maximum number of run N of network and threshold value look-up table LT (s), s is the function variable of LT (s);

$N=\frac{t_2-t_1}{\mathrm{\&Delta;t}}+1$

$t_1=\frac{1}{{\mathrm{\&alpha;}}_{\mathrm{\&theta;}}}\ln \left[\frac{V_{\mathrm{\&theta;}}}{S_{\mathrm{ij}\_\max }}\right]$

$t_2=\frac{1}{{\mathrm{\&alpha;}}_{\mathrm{\&theta;}}}\ln \left[\frac{V_{\mathrm{\&theta;}}}{S_{\mathrm{ij}\_\min }}\right]$

$\mathrm{LT}\left(s\right)=V_{\mathrm{\&theta;}}{e}^{(-(-\mathrm{s\&Delta;t}+t_2){\mathrm{\&alpha;}}_{\mathrm{\&theta;}})}$

In formula: t ₁and t ₂be respectively the autogenous ignition time of gray-scale value maximum pixel and minimum pixel in image;

Step 5, utilize following formula moving model;

F _ij[n]＝S _ij

L _ij[n]＝V _L∑w _ijklY _kl[n-1]

U _ij[n]＝F _ij[n](1+βL _ij[n])

I _ij[n]＝N-n

In formula: U _ij[n] is internal activity item, Y _ij[n] is PCNN pulse output, I _ij[n] is index value;

When n=1, L _ij[1]=0, U _ij[1]=F _ij[1]=S _ij, θ _ij[1]=LT (N-1)=S _{ij_max}, corresponding feed back input intermediate value is S _{ij_max}neuron by autogenous ignition; After neuron firing, output Y _ij[1]=1, θ _ij[2] become V _θ, the neuronic index value of lighting a fire is labeled as I _ij=N-1;

By that analogy, when n=N is arrived in the network operation, threshold value θ _ij[N]=LT (0)=S _{ij_min}, for feed back input, be S _{ij_min}neuron autogenous ignition, the neuronic index value of lighting a fire is labeled as I _ij=0;

Step 6, obtain respectively the key map I of three width original image A, B and C _a, I _band I _c;

Work as I _a, I _band I _cthe absolute value of the mutual difference of index value of respective pixel is all less than or equal to representative value e, and the pixel value of fused images is got the weighted mean value of three width image respective pixel;

Work as I _a, I _band I _cwhen the absolute value of the mutual difference of index value of respective pixel has the situation that is greater than representative value e, if wherein the mutual difference of index value of two width image respective pixel is less than or equal to representative value e, the pixel value of fused images is got the weighted mean value of above-mentioned two width image respective pixel;

In other situations, the pixel value of fused images is got the pixel value that index value is larger.

2. a kind of PCNN multisource image anastomosing method based on improved model according to claim 1, is characterized in that: in described step 2, guarantee the feed back input F of dendron _ij[n] receives only outside input stimulus signal S _ij; Guarantee W, V _l, β, V _θ, α _θall identical to all neurons with the value of Δ t.

3. a kind of PCNN multisource image anastomosing method based on improved model according to claim 1, is characterized in that: e=2.

4. a kind of PCNN multisource image anastomosing method based on improved model according to claim 1, is characterized in that: three width original image A, B and C are respectively visible ray, medium-wave infrared and LONG WAVE INFRARED image.

Images