CN103679670B - A kind of PCNN multisource image anastomosing method based on improved model - Google Patents

Abstract

The present invention relates to a kind of PCNN multisource image anastomosing method based on improved model.Have in improvement: in PCNN, the feed back input of each neuron only receives outside stimulus input；In link field, the value of each parameter is the most identical to all neurons；In variable threshold value function, the value of each parameter is the most identical to all neurons；Introduce threshold value look-up table and index map, threshold value look-up table have recorded the threshold value corresponding with network operation number of times, and these threshold values can precalculate before the network operation and obtain, it is to avoid the exponent arithmetic in the network operation, accelerates the operation of network.Index map have recorded the duration of ignition of whole pixel, is the integrated results of the similar pixel that space is adjacent in input picture, and embodiment is the overall visual signature of input picture.Invention introduces the threshold value look-up table that the index map of the duration of ignition recording whole pixel is corresponding with network operation number of times with record, have employed fusion rule based on index map, achieve than traditional more preferable effect of wavelet transform fusion.

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 be suitable for that visible ray, medium wave and three wave bands of LONG WAVE INFRARED merge simultaneously based on improved model PCNN multisource image anastomosing method.

Background technology

Artificial neural network is at a kind of novel calculating attempting imitation biological nervous system information processing manner Reason model.One neutral net is made up of Multilevel method unit or node, and various method can be used to carry out Interconnection.Some scholar is own carries out multi-source image fusion through application artificial neural network.At present, nerve net Network application in image co-registration mainly has: 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 proposed in recent years, international Upper referred to as third generation artificial neural network.

1981, E.A.Newman, P.H.Hartline etc. proposed 6 kinds different types of pair Mode neuron (include AND, OR, Visible-Enhanced Infrared, Visible-Suppressed-Infrared, Infrared-Enhanced-Visible and Infrared-Suppressed-Visible) for visible ray and the fusion of infrared image.Nineteen ninety-five, Fechner Image interfusion method based on multilayer perceptron neutral net is proposed with Godlewski.Many by training Pixel interested in layer perceptron identification prebiotic synthesis, is incorporated in visible images.From 20 The nineties in century, existing to the visual cortex nerve impulse string synchronized oscillation of cat, monkey by Eckhorn etc. The research of elephant, has obtained mammalian nervous meta-model, and thus development has defined pulse coupled neural net Network model.This model has the advantages that to be grouped the pixel that two-dimensional image spatial similarity, gray scale are similar, And image local gray scale difference value can be reduced, make up the small interruption of image local.1999, Brocssard Etc. R.P. the spark rate of PCNN neuron is demonstrated with the relation of gradation of image it was confirmed PCNN Feasibility for image co-registration.Based on this model, relevant scholar proposes various improved model, and will It is for the fusion of various images.

At present, the image interfusion method of Based PC NN is studied and is mainly concentrated the following aspects:

The automation of network parameter is chosen: the parameter related to due to PCNN network is more, and different ginseng Numerical value all can affect final result.Automatically the key in PCNN network is calculated by householder method Parameter, can obtain more preferable result.

Improvement to PCNN basic model: according to realizing function, process object and the mode of thinking are not With, different researchers successively propose different improved models.

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

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, after making fusion Image object feature becomes apparent from, is more beneficial for the PCNN multi-source image based on improved model of target identification Fusion method.

For solving above-mentioned technical problem, a kind of PCNN multi-source image based on improved model of the present invention merges Method, comprises the following steps successively:

Step one, to input three width original image A, B and C spatially carry out pixel level registration, Ensure 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_θIt is respectively L_ij[n] and θ_ijIntrinsic electromotive force in [n], θ_ij[n] is dynamic threshold, L_ij[n] is Linearly connected inputs；

α_θFor θ_ijThe damping time constant of [n]；Δ t is time sampling interval；β is bonding strength between cynapse Constant；Y_ij[n] is PCNN pulse output；Y_kl[n-1] is PCNN last time pulse output；Inside connects Meet the w in matrix W_ijklCorresponding L_ijY in [n]_klThe weight coefficient of [n-1]； N is the number of run of network, n=1,2 ..., N-1, N, N are 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, obtaining network maximum number of run N and threshold value look-up table LT (s), s is the letter of LT (s) Number variable；

$N=\frac{t_2-t_1}{\mathrm{\Δt}}+1$

$t_1=\frac{1}{{\mathrm{\α}}_{\mathrm{\θ}}}\ln \left[\frac{V_{\mathrm{\θ}}}{S_{\mathrm{ij}\_\max }}\right]$

$t_2=\frac{1}{{\mathrm{\α}}_{\mathrm{\θ}}}\ln \left[\frac{V_{\mathrm{\θ}}}{S_{\mathrm{ij}\_\min }}\right]$

$\mathrm{LT}\left(s\right)=V_{\mathrm{\θ}}{e}^{(-(-\mathrm{s\Δt}+t_2){\mathrm{\α}}_{\mathrm{\θ}})}$

In formula: t₁And t₂It is respectively gray value maximum pixel and the autogenous ignition time of minimum pixel in image；

Step 5, utilize following equation 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； As n=1, L_ij[1]=0, then 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, export Y_ij[1]=1, θ_ij[2] V is become_θ, the index value of igniting neuron is labeled as I_ij=N-1；

By that analogy, when the network operation to n=N, threshold θ_ij[N]=LT(0)=S_{ij_min}, defeated for feedback Enter for S_{ij_min}Neuron autogenous ignition, igniting neuron index value be labeled as I_ij=0；

Step 6, respectively obtain the index 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 both less than equal to representative value E, then the pixel value of fused images takes the weighted average of three width image respective pixel；

Work as I_A、I_BAnd I_CThe absolute value of the mutual difference of index value of respective pixel has more than representative value e's During situation, if wherein the mutual difference of index value of two width image respective pixel is less than or equal to representative value e, then The pixel value of fused images takes the weighted average of above-mentioned two width image respective pixel；

In the case of other, then the pixel value of fused images takes the pixel value that index value is bigger.

In step 2, it is ensured that the feed back input F of dendron_ij[n] only receives outside input stimulus signal S_ij；Ensure W、V_L、β、V_θ、α_θThe most 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.

Basic model, on the basis of analyzing PCNN image interfusion method, is simplified by the present invention accordingly And improvement, obtain the PCNN image interfusion method of a kind of new improvement.Have in wherein simplifying and improving: 1. in PCNN, the feed back input of each neuron only receives outside stimulus input；2. each parameter in link field Value the most identical to all neurons；3. in variable threshold value function, the value of each parameter is to all neurons all Identical；4. introducing threshold value look-up table and index map, threshold value look-up table have recorded corresponding with network operation number of times Threshold value, these threshold values can precalculate before the network operation and obtain, it is to avoid the finger in the network operation Number computing, accelerates the operation of network.Index map have recorded the duration of ignition of whole pixel, is input figure The integrated results of the similar pixel that space is adjacent in Xiang, embodiment is the overall visual signature of input picture.

The present invention introduce in improved model the duration of ignition recording whole pixel index map and record with The threshold value look-up table that network operation number of times is corresponding, under the conditions of identical fusion rule, becomes than traditional small echo Change fusion method and have more preferable effect.

The indices of the present invention is better than the indices of the former fusion rule of WT+ to a certain extent, especially It is the index such as average and standard variance, improves clearly, illustrate the validity of improved method.

Accompanying drawing explanation

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

Fig. 2 is certain seashore original image and fusion results of Based PC NN.

Detailed description of the invention

The present invention is further detailed explanation with embodiment below in conjunction with the accompanying drawings.

The basic thought of the present invention is: two width or several original images to input are respectively adopted PCNN mould Type is calculated the index map of correspondence, then uses index map and original image and merges decision-making accordingly, Finally give fused images.The PCNN fusion method schematic diagram of three width images is as shown in Figure 1.

Specifically, the present invention comprises the following steps successively:

Step one, to input three width original image A, B and C spatially carry out pixel level registration, Ensure that three width image sizes are X × Y；Three width original image A, B and C be respectively visible ray, in Ripple is infrared and LONG WAVE INFRARED image；

Step 2, setting network parameter W, V_L, β, V_θ, α_θValue with Δ t；Ensure the feedback of dendron Input F_ij[n] only receives outside input stimulus signal S_ij；Ensure each parameter W, V in link field_L, the taking of β It is worth the most identical to all neurons；Ensure each parameter V in variable threshold value function_θ、α_θWith the value of Δ t to all Neuron is the most identical；

V_LAnd V_θIt is respectively L_ij[n] and θ_ijIntrinsic electromotive force (amplification coefficient) in [n], θ_ij[n] is dynamic threshold Value, L_ij[n] is linearly connected input；

α_θFor θ_ijThe damping time constant of [n]；Δ t is time sampling interval；β is bonding strength between cynapse Constant；Y_ij[n] is PCNN pulse output；Y_kl[n-1] is PCNN last time pulse output；Inside connects Meet the w in matrix W_ijklCorresponding L_ijY in [n]_klThe weight coefficient of [n-1]；

N is the number of run of network, n=1,2 ..., N-1, N, N are 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}Little In equal to 0, by Serial regulation S_{ij_max}、S_{ij_min}And input picture, make S_{ij_min}More than 0；

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

$N=\frac{t_2-t_1}{\mathrm{\Δt}}+1$

$t_1=\frac{1}{{\mathrm{\α}}_{\mathrm{\θ}}}\ln \left[\frac{V_{\mathrm{\θ}}}{S_{\mathrm{ij}\_\max }}\right]$

$t_2=\frac{1}{{\mathrm{\α}}_{\mathrm{\θ}}}\ln \left[\frac{V_{\mathrm{\θ}}}{S_{\mathrm{ij}\_\min }}\right]$

$\mathrm{LT}\left(s\right)=V_{\mathrm{\θ}}{e}^{(-(-\mathrm{s\Δt}+t_2){\mathrm{\α}}_{\mathrm{\θ}})}$

In formula: t₁And t₂It is respectively gray value maximum pixel and the autogenous ignition time of minimum pixel in image；

Step 5, utilize following equation 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；

As n=1, L_ij[1]=0, then 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, export Y_ij[1]=1, θ_ij[2] V is become_θSo that igniting neuron will not be the most ignited.Meanwhile, the rope of igniting neuron Draw value and be labeled as I_ij=N-1, no longer changes.Along with the increase of the number of run n of network, threshold θ_ij[n] by The least, the neuron of igniting will encourage adjacent neuron by link field.When the network operation to n=N Time, threshold θ_ij[N]=LT(0)=S_{ij_min}, it is S for feed back input_{ij_min}Neuron, even if there is no phase In the case of adjacent neuron excitation, also will autogenous ignition, the index value of igniting neuron is labeled as I_ij=0； I.e. by the network operation of most n times, all of neuron will ignited and also only once.

If it can be seen that neuron S_ijLight a fire when n-th is run, then its index map I_ij[n] is solid It is set to N-n, no longer changes along with the operation of network.Index map I_ij[n] have recorded all neurons The duration of ignition, is the result of input picture space-time integration.Additionally, due to the variable threshold value in the network operation is Obtained by look-up table mode, it is not necessary to complicated exponent arithmetic, shorten the time of the network operation.Net Network runs the automatic acquisition of maximum times, not only ensure that the certainty of the network operation, and makes each Neuron firing is only once；

Step 6, use corresponding convergence strategy, obtain fused images.The convergence strategy that this method uses As follows: owing to the index map of every width original image represents the overall visual signature of original image, therefore to lead to The index map crossing three width original images carrys out synthetic setting convergence strategy.

Obtain the index 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 both less than equal to representative value E, then the pixel value of fused images takes the weighted average of three width image respective pixel；

Work as I_A、I_BAnd I_CThe absolute value of the mutual difference of index value of respective pixel has more than representative value e's During situation, if wherein the mutual difference of index value of two width image respective pixel is less than or equal to representative value e, then The pixel value of fused images takes the weighted average of above-mentioned two width image respective pixel；

In the case of other, then the pixel value of fused images takes the pixel value that index value is bigger.

Preferably e=2.

In order to verify the performance based on improvement PCNN model image fusion method of proposition, select certain sea The three width visible rays, LONG WAVE INFRARED and the medium-wave infrared figure that are registered of bank (image size is 320 × 256) As image to be fused.The (a) and (b) of Fig. 2 and (c) are respectively the to be fused original of certain seashore can Seeing light, LONG WAVE INFRARED and medium-wave infrared image, (d) of Fig. 2 is to use 4 layers of wavelet transformation (WT), Low frequency coefficient is averaged, and high frequency coefficient uses the fusion results that the fusion rule of region energy operator obtains； (e) of Fig. 2 is the fusion results using and improving PCNN model.In order to preferably fusion results be entered Row com-parison and analysis, have employed the product of average, standard deviation, entropy and structural information and transinformation contentObjective evaluation criteria calculated, the result obtained is as shown in table 1.It can be seen that Based on improve PCNN model fusion results local detail, profile (personage on seashore limit) and Fusion results based on the former fusion rule of WT+ it is better than on overall brightness.From the three of objective evaluation tables It can be seen that be better than the former fusion of WT+ to a certain extent based on the indices improving PCNN model The indices of rule, the particularly index such as average and standard variance, improves clearly, illustrates to change Enter the validity of method.

Certain seashore fusion results objective evaluation of table 1 Based PC NN

Claims (4)

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

Step one, to input three width original image A, B and C spatially carry out pixel level registration, Ensure 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_θIt is respectively L_ij[n] and θ_ijIntrinsic electromotive force in [n], θ_ij[n] is dynamic threshold, L_ij[n] is Linearly connected inputs；

α_θFor θ_ijThe damping time constant of [n]；Δ t is time sampling interval；β is bonding strength between cynapse Constant；Y_ij[n] is PCNN pulse output；Y_kl[n-1] is PCNN last time pulse output；Inside connects Meet the w in matrix W_ijklCorresponding L_ijY in [n]_klThe weight coefficient of [n-1]；

N is the number of run of network, n=1,2 ..., N-1, N, N are 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, obtaining network maximum number of run N and threshold value look-up table LT (s), s is the letter of LT (s) Number variable；

$N=\frac{t_2-t_1}{\mathrm{\Δt}}+1$

$t_1=\frac{1}{{\mathrm{\α}}_{\mathrm{\θ}}}\ln \left[\frac{V_{\mathrm{\θ}}}{S_{\mathrm{ij}\_\max }}\right]$

$t_2=\frac{1}{{\mathrm{\α}}_{\mathrm{\θ}}}\ln \left[\frac{V_{\mathrm{\θ}}}{S_{\mathrm{ij}\_\min }}\right]$

$\mathrm{LT}\left(s\right)=V_{\mathrm{\θ}}{e}^{(-(-\mathrm{s\Δt}+t_2){\mathrm{\α}}_{\mathrm{\θ}})}$

In formula: t₁And t₂It is respectively gray value maximum pixel and the autogenous ignition time of minimum pixel in image；

Step 5, utilize following equation 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；

As n=1, L_ij[1]=0, then 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, export Y_ij[1]=1, θ_ij[2] V is become_θ, the index value of igniting neuron is labeled as I_ij=N-1；

By that analogy, when the network operation to n=N, threshold θ_ij[N]=LT(0)=S_{ij_min}, defeated for feedback Enter for S_{ij_min}Neuron autogenous ignition, igniting neuron index value be labeled as I_ij=0；

Step 6, respectively obtain the index 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 both less than equal to representative value E, then the pixel value of fused images takes the weighted average of three width image respective pixel；

Work as I_A、I_BAnd I_CThe absolute value of the mutual difference of index value of respective pixel has more than representative value e's During situation, if wherein the mutual difference of index value of two width image respective pixel is less than or equal to representative value e, then The pixel value of fused images takes the weighted average of above-mentioned two width image respective pixel；

In the case of other, then the pixel value of fused images takes the pixel value that index value is bigger.

A kind of PCNN multisource image anastomosing method based on improved model the most according to claim 1, It is characterized in that: in described step 2, it is ensured that the feed back input F of dendron_ij[n] only receives outside input stimulus Signal S_ij；Ensure W, V_L、β、V_θ、α_θThe most identical to all neurons with the value of Δ t.

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

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