CN101894295B - Method for simulating attention mobility by using neural network - Google Patents

Abstract

The invention discloses a method for simulating attention mobility by using a neural network, which adopts a visual image input layer, neuronal oscillator network oscillation layer and an attention mobility realizing layer, wherein the visual image input layer inputs a gray value of a gray level image into a neurodynamic network; the neuronal oscillator network oscillation layer couples dynamic systems, which are built by all oscillators in the neurodynamic network according to a FitzHugh-Nagumo model, to form the neurodynamic network; and the attention mobility realizing layer realizes the attention mobility between the synchronized different objects by changing parameters, so that the frequency output by nerves of a currently concerned object increases. On the basis of the neurodynamic system, the method forms simple simulation on human-eye attention mobility visual treatment by analyzing and modifying the FitzHugh-Nagumo model and has a great theoretical and practical significance for further research on the visual treatment mechanism of human.

Description

By the method for neuron network simulation attention mobility

Technical field

The present invention relates to process and nerual network technique taking neurodynamics as basic image, particularly relate to the neural power network that a kind of neuron oscillator with neurodynamics model description forms, for simulating simple artificial gray level image people's the method for attention mobility.

Background technology

Vision attention is selected and vision attention transfer is the important mechanisms that ensures that biosystem is finished the work with limited processing power.From image or receptive field, extract significant feature, and put them into different regions, therefrom selecting significant region is the basic task that perception is understood.This ability is exactly the vision attention selection in visual analysis.After significant region is selected, due to the adaptability of vision system, notice can forward from current significant region next significant region to, and this is attention mobility.Neuroid based on synchronized oscillation in image processing is applied more and more widely, the neuron models of pulse-couple have obtained good achievement, but the application in vision attention selection and vision attention transfer remains very important concerning current research

Human eye can carry out perception and understanding while processing visual pattern in a short period of time easily, and realize from a significant zone-transfer to the significant region of the next one, although this research is on the one hand a lot of, but researcher still knows little about it so far to mechanism wherein.Researchers have proposed neuron behavior in the real human brain of a lot of neuron oscillator network modeling, and wherein nineteen fifty-two is first model of extensively being quoted by the Hodgkin-Huxley model (hereinafter to be referred as HH model) of Hodgkin and Huxley proposition on the basis of great many of experiments.FitzHugh-Nagumo (hereinafter to be referred as FHN) model is from HH model simplification, and is proved that by multiple researchers the current potential of imictron is provided preferably, is the important models of simulation people's vision.

FHN model description a neuron by input current I _inputrelation between film potential V and the internal state variable R driving.V and R are respectively the rate of change of time t:

$\frac{\mathrm{dV}}{\mathrm{dt}}=10(V-\frac{V_3}{3}-\mathrm{\αR}+I_{\mathrm{input}})---\left(1\right)$

$\frac{\mathrm{dR}}{\mathrm{dt}}=0.8(-R+\mathrm{\βV}+1.5)$

V is the potential difference (PD) of neuronal cell film both sides, is called film potential, and R is the internal state variable that represents current potential threshold value, I _inputbeing the input current that neuron receives, is also this neuronic light stimulus of outer bound pair.Here 10 and 0.8 is the inverse of the time constant of V and R, and α > 0, and β > 0 has described respectively the action intensity from R to V and from V to R.The time constant of V is 12.5 times of R, and this has reacted the activation fact more faster than rejuvenation in aixs cylinder.Suppose that (V, R) is equilibrium point, only have when equilibrium point (V, R) is when being unsettled, around it, produce stable limit cycle.Therefore the value of α and β must meet formula (2)

8(V ²+αβ-1)＜0(2)

Or formula (3)

8(V ²+αβ-1)＞0

-(10V ²-9.2)＞0(3)

The parameter alpha and the β that change in equation (1) can control granting frequency.The increase of α or β all can cause the increase of providing frequency, but (for example α=1, β=0.6 or α=0.3, β=2 or α=0.3 in the time that α or β are a quite little value, β=0.6), equation (1) does not produce granting.

Summary of the invention

The object of the invention is to, by a kind of method with neuron network simulation attention mobility is provided, the FHN model of application transformation forms neural power network analog human eye process from a salient region to another in receptive field, and build attention mobility system in simple receptive field, single neuron is carried out to model construction, and based on single Construction of A Model biological vision neural network.

The method of neuron network simulation attention mobility for the present invention is to adopt following technological means to realize.

Step 1, visual pattern input layer are input to the gray-scale value of gray level image in neural power network, there is one-to-one relationship in the neuron oscillator on the pixel in image and neural power network, neurodynamics system corresponding to each pixel is by FHN model description.

Step 2, neuron oscillator network oscillating layer, the dynamic system model coupling that each oscillator in neural network is set up according to FHN model forms neural power network, and i film potential V and voltage threshold R capable, a j row neuron oscillator are wherein respectively the rate of change of time;

$\left\{\begin{array}{c}\frac{{\mathrm{dV}}_{i,j}}{\mathrm{dt}}=10((V_{i,j}+\mathrm{\Δ}{V}_{i,j})-\frac{{(V_{i,j}+\mathrm{\Δ}{V}_{i,j})}^3}{3}-\mathrm{\α}{R}_{i,j}+I_{i,j})\\ \frac{d{R}_{i,j}}{\mathrm{dt}}=0.8(-(R_{i,j}+\mathrm{\Δ}{R}_{i,j})+\mathrm{\β}{V}_{i,j}+1.5)\end{array}\right.---\left(4\right)$

Wherein, V is the potential difference (PD) of neuronal cell film both sides, is called film potential, and R is the internal state variable that represents current potential threshold value, and in following formula, V and R represent identical implication, I _{i, j}represent that i is capable, a j row neuron receives from extraneous light stimulus, equals the gray-scale value in gray level image on numerical value.

Step 3, attention mobility realize layer and in vibration, adjust α and β according to being defined in of following formula (5) and (6), thereby the corresponding neuron pool of the object of realizing current attention is provided frequency and is raise, and neuron pool corresponding to other objects produced to inhibition signal, make it provide frequency and slow down, distinguish current concern object and other objects with this.

${\mathrm{\α}}_{p,q}\left(\mathrm{\τ}\right)={\mathrm{\α}}_{p,q}(\mathrm{\τ}-1)+\frac{h_1\left({\mathrm{\α}}_{p,q}(\mathrm{\τ}-1)\right)}{M\left(\mathrm{\τ}\right)}\underset{i,j\∈\mathrm{\Δ}\left(\mathrm{\τ}\right)}{\mathrm{\Σ}}{I}_{i,j}{f}_1\left(\right||V_{i,j}-V_{p,q}|\left|\right)---\left(5\right)$

${\mathrm{\β}}_{p,q}\left(\mathrm{\τ}\right)={\mathrm{\β}}_{p,q}(\mathrm{\τ}-1)+\frac{h_2\left({\mathrm{\β}}_{p,q}(\mathrm{\τ}-1)\right)}{M\left(\mathrm{\τ}\right)}\underset{i,j\∈\mathrm{\Δ}\left(\mathrm{\τ}\right)}{\mathrm{\Σ}}{I}_{i,j}{f}_2\left(\right||V_{i,j}-V_{p,q}|\left|\right)---\left(6\right)$

Wherein, (p, q) pixel of the capable q row of p in presentation video, τ is the moment that has at least a neuron providing, τ-1st, the previous moment of τ, M (τ) is the neuronic quantity of moment τ at the state of granting, and Δ (τ) is the neuronic set of moment τ at the state of granting, I _{i, j}represent that i is capable, a j row neuron receives from extraneous light stimulus, equals the gray-scale value in gray level image on numerical value.

Described FHN model is the abbreviation of FitzHugh-Nagumo model.

Aforesaid visual pattern input layer, before gray-scale value being input to neural power network, has carried out normalization, and the gray-scale value after normalization is in [0,1] scope.

Only have and just have stable limit cycle generation when being unsettled when equilibrium point (V, R), the value of α and β must meet formula (2)

8(V ²+αβ-1)＜0(2)

Or formula (3)

8(V ²+αβ-1)＞0

-(10V ²-9.2)＞0°(3)

Aforesaid neuron oscillator network oscillating layer, i in described formula (4) in (i, j) presentation video is capable, j row, 1≤i≤M, 1≤j≤N;

Wherein, M and N are the wide and high of image; Δ V _{i, j}with Δ R _{i, j}represent the impact of peripheral nerve unit, they are defined by following formula:

Δx _i，j＝γ _{i-1，j-1；i，j}(x _i-1，j-1-x _i，j)+γ _{i-1，j；i，j}(x _i-1，j-x _i，j)+

γ _{i-1，j+1；i，j}(x _i-1，j+1-x _i，j)+γ _{i，j-1；i，j}(x _i，j-1-x _i，j)+

γ _{i，j+1；i，j}(x _i，j+1-x _i，j)+γ _{i+1，j-1；i，j}(x _i+1，j-1-x _i，j)+

(7)

γ _{i+1，j；i，j}(x _i+1，j-x _i，j)+γ _{i+1，j+1；i，j}(x _i+1，j+1-x _i，j)

Wherein

Wherein x represents V or R.

Aforesaid attention mobility realizes layer, θ in following formula (9) and (10) ₁> θ ₂> 0

$h_1\left(\mathrm{\&alpha;}\right)=\left\{\begin{array}{cc}{\mathrm{\&theta;}}_1& \mathrm{\&alpha;}<{\mathrm{\&theta;}}_{\mathrm{\&alpha;}}\\ {\mathrm{\&theta;}}_2& \mathrm{\&alpha;}\&GreaterEqual;{\mathrm{\&theta;}}_{\mathrm{\&alpha;}}\end{array}\right.---\left(9\right)$

$h_2\left(\mathrm{\&beta;}\right)=\left\{\begin{array}{cc}{\mathrm{\&theta;}}_1& \mathrm{\&beta;}\&GreaterEqual;{\mathrm{\&theta;}}_{\mathrm{\&beta;}}\\ {\mathrm{\&theta;}}_2& \mathrm{\&beta;}<{\mathrm{\&theta;}}_{\mathrm{\&beta;}}\end{array}\right.---\left(10\right)$

f ₁(x)＝a ₁x+b ₁(a ₁＞0，b ₁＜0)(11)

f ₂(x)＝a ₂x+b ₂(a ₂＜0，b ₂＞0)(12)

Implementation process parameter is a ₁=2, b ₁=-4, a ₂=-4, b ₂=2, θ ₁=0.1, θ ₂=0.01, θ _α=0.8, θ _β=4.

A kind of method with neuron network simulation attention mobility of the present invention, compared with using the method and application thereof of FHN model in previous literature, has following obvious advantage and useful effect:

1, the document different from the past of the coupling scheme between neuron.

2, FHN model is used for simulating people's attention mobility mechanism.

Brief description of the drawings

Fig. 1 attention mobility system architecture schematic diagram;

Fig. 2 is image and the corresponding schematic diagram of neural power network;

Fig. 3 is input picture schematic diagram;

Fig. 4 is the successively concern object figure of method to Fig. 3 in the present invention.

Embodiment

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is illustrated.

This system is divided into three parts, visual pattern input layer, and neuron oscillator network oscillating layer, attention mobility realizes layer.Relation between three layers is as shown in Fig. 1 attention mobility system architecture schematic diagram.

First this system is set up each pixel in a width visual pattern and all neuron oscillators in neural power network associated one to one, then allows this group neuron oscillator produce vibration.Because the object that brightness is large is stronger to human eye retina's light stimulus, so I _inputthe increase increase that causes providing frequency, and can other regions be produced and be suppressed in distribution process.

The single neuron by FHN model description arranged with the form of matrix and carried out model reconstruction, forming neural power network.In newly-established model each neuron oscillator except edge with its periphery 8 fields in neuron have coupled relation, the neuron oscillator at edge according to the difference of its adjacent vibration generators respectively with 3 or 5 neuron couplings.In neural power network, film potential V and the voltage threshold R of the neuron oscillator that i is capable, j is listed as are respectively the rate of change of time:

$\left\{\begin{array}{c}\frac{{\mathrm{dV}}_{i,j}}{\mathrm{dt}}=10((V_{i,j}+\mathrm{\&Delta;}{V}_{i,j})-\frac{{(V_{i,j}+\mathrm{\&Delta;}{V}_{i,j})}^3}{3}-\mathrm{\&alpha;}{R}_{i,j}+I_{i,j})\\ \frac{d{R}_{i,j}}{\mathrm{dt}}=0.8(-(R_{i,j}+\mathrm{\&Delta;}{R}_{i,j})+\mathrm{\&beta;}{V}_{i,j}+1.5)\end{array}\right.---\left(4\right)$

The pixel that in equation, in (i, j) presentation video, i is capable, j is listed as, also represents the neuron oscillator that in neural power network, i is capable, j is listed as, 1≤i≤M, 1≤j≤N (M and N are the wide and high of image).Δ V _{i, j}with Δ R _{i, j}representative is the neuronic impact of 8 neighborhoods around, and they are defined by following formula:

Δx _i，j＝γ _{i-1，j-1，i，j}(x _i-1，j-1-x _i，j)+γ _{i-1，j；i，j}(x _i-1，j-x _i，j)+

γ _{i-1，j+1；i，j}(x _i-1，j+1-x _i，j)+γ _{i，j-1；i，j}(x _i，j-1-x _i，j)+

γ _{i，j+1；i，j}(x _i，j+1-x _i，j)+γ _{i+1，j-1；i，j}(x _i+1，j-1-x _i，j)+

(7)

γ _{i+1，j；i，j}(x _i+1，j-x _i，j)+γ _{i+1，j+1；i，j}(x _i+1，j+1-x _i，j)

Wherein wherein x represents V or R.

For the neuron of realizing corresponding obvious object is provided more frequently, the neuron of other objects is with lower frequency granting or do not provide such fact, first we allow neural power Netowrk tape preset parameter α and β move, until the neuron of corresponding same area is synchronous, this has also shown to cut apart task.Afterwards, whenever any neuron granting, its can produce two kinds of signals to own and other neurons: to own and together with the neuron of granting produce excitatoty signal, what to provide with it is the signal of inhibition.

In system when operation,, we can control the change of parameter alpha and β, make gathering way of α reduce large many of speed than β.High frequency period oscillation phase will be jumped in the most significant like this region, and other regions can be relatively quiet, notes also can transferring to the most significant region.After receiving attention, this region can be suppressed to allow other regions to become significantly.

From to the analysis of equation (1), we can draw above, α and β can control neuronic activity.

After a neuron granting, can cause the change of periphery neuron α and β parameter, thereby realize the transfer of noting.Wherein the variation of α and β is defined by following equation:

${\mathrm{\&alpha;}}_{p,q}\left(\mathrm{\&tau;}\right)={\mathrm{\&alpha;}}_{p,q}(\mathrm{\&tau;}-1)+\frac{h_1\left({\mathrm{\&alpha;}}_{p,q}(\mathrm{\&tau;}-1)\right)}{M\left(\mathrm{\&tau;}\right)}\underset{i,j\&Element;\left(\mathrm{\&tau;}\right)}{\mathrm{\&Sigma;}}{I}_{i,j}{f}_1\left(\right||V_{i,j}-V_{p,q}|\left|\right)---\left(5\right)$

${\mathrm{\&beta;}}_{p,q}\left(\mathrm{\&tau;}\right)={\mathrm{\&beta;}}_{p,q}(\mathrm{\&tau;}-1)+\frac{h_2\left({\mathrm{\&beta;}}_{p,q}(\mathrm{\&tau;}-1)\right)}{M\left(\mathrm{\&tau;}\right)}\underset{i,j\&Element;\mathrm{\&Delta;}\left(\mathrm{\&tau;}\right)}{\mathrm{\&Sigma;}}{I}_{i,j}{f}_2\left(\right||V_{i,j}-V_{p,q}|\left|\right)---\left(6\right)$

Wherein

$h_1\left(\mathrm{\&alpha;}\right)=\left\{\begin{array}{cc}{\mathrm{\&theta;}}_1& \mathrm{\&alpha;}<{\mathrm{\&theta;}}_{\mathrm{\&alpha;}}\\ {\mathrm{\&theta;}}_2& \mathrm{\&alpha;}\&GreaterEqual;{\mathrm{\&theta;}}_{\mathrm{\&alpha;}}\end{array}\right.,{\mathrm{\&theta;}}_1>{\mathrm{\&theta;}}_2>0--\left(9\right)$

$h_2\left(\mathrm{\&beta;}\right)=\left\{\begin{array}{cc}{\mathrm{\&theta;}}_1& \mathrm{\&beta;}\&GreaterEqual;{\mathrm{\&theta;}}_{\mathrm{\&beta;}}\\ {\mathrm{\&theta;}}_2& \mathrm{\&beta;}<{\mathrm{\&theta;}}_{\mathrm{\&beta;}}\end{array}\right.{\mathrm{\&theta;}}_1>{\mathrm{\&theta;}}_2>0---\left(10\right)$

f ₁(x)＝a ₁x+b ₁(a ₁＞0，b ₁＜0)(11)

f ₂(x)＝a ₂x+b ₂(a ₂＜0，b ₂＞0)(12)

(p, q) pixel of the capable q row of p in presentation video, τ has a neuron at least in the moment of providing, and M (τ) is the neuronic quantity of moment τ at the state of granting, Δ (τ) is the neuronic set of moment τ at the state of granting, I _{i, j}represent neuron receive from extraneous light stimulus, on numerical value, equal the gray-scale value in gray level image.By a is set ₁> 0, b ₁< 0 and a ₂< 0, b ₂> 0, and function h ₁, h ₂, the neuron (i, j) in each granting is provided the signal of excitability or inhibition to (p, q).

First be to set up visual pattern input layer.According to the system of formula (1) and formula (4) establishment, set up neural power network, and the pixel in the each neuron on network and image is set up to relation one to one, and the gray-scale value of image also becomes the input value of neuromotor system, the I in corresponding formula (1) _input.Accompanying drawing 2 is neural power network and gray level image corresponding diagram one by one, each small circle is a pixel in representative image both, also a neuron in neural power network corresponding to representative image, stain represents the neuron of denumerable number, horizontal line represents the connection between neuron.

Next is neuron oscillator network oscillating layer.Solve by 4 rank Runge-Kutta methods by each neuron oscillator in the neural power network to whole, the oscillator of realizing in same object reaches synchronized oscillation, and the oscillator between different objects desynchronizes.

Finally that attention mobility realizes layer.Change α and β according to the definition in formula (5), formula (6), result is to select successively according to the object in the sized images of conspicuousness, and when the object of selecting is noted, the granting frequency of corresponding neural oscillator can become large.

Next by the concrete image graph 3 of a width, implementing procedure is described.Fig. 3 is a width gray-scale map, formed by 4 targets, and be 1. wherein the sun, corresponding grey scale value 255, is 2. tree, and corresponding grey scale value 195, is 3. the chain of mountains, and corresponding grey scale value 130, is 4. sky, corresponding grey scale value 65.First Fig. 3 is normalized to operation, is namely gone up pixel value span and become [0,1] from [0,255].Visual pattern after normalization is input in the dynamical system (4) establishing, and the dynamical system establishing refers to according to formula (4), (5), the system that (6) are set up.Next utilize the classical Runge-Kutta method in 4 rank to solve to (4) formula, parameter alpha=1 in solution procedure, β=1.25 remain unchanged first voluntarily, until the neuron oscillator in the regional area in neural power network reaches synchro-resonance, in solution procedure, solving result V and R obtain periodic solution.Realize in attention mobility the parameter using in layer and be respectively a ₁=2, b ₁=-4, a ₂=-4, b ₂=2, θ ₁=0.1, θ ₂=0.01, θ _α=0.8, θ _β=4.Pass through iterative, the granting frequency that in Fig. 3,4 objects are corresponding according to 1. 2. 3. order 4. increase respectively, when increasing, suppresses on certain object the granting frequency of other objects, the granting frequency that is other 3 objects slows down, and realizes thus attention mobility and 1. 2. 3. 4. between 4 objects, is constantly switching.The object of selecting successively in Fig. 3 is paid close attention to as shown in object figure successively as Fig. 4.

Finally it should be noted that: above embodiment is only in order to illustrate the present invention and unrestricted technical scheme described in the invention; Therefore, although this instructions has been described in detail the present invention with reference to each above-mentioned embodiment,, those of ordinary skill in the art should be appreciated that still and can modify or be equal to replacement the present invention; And all do not depart from technical scheme and the improvement thereof of the spirit and scope of invention, it all should be encompassed in the middle of claim scope of the present invention.

Claims (2)

1. by a method for neuron network simulation attention mobility, it is characterized in that comprising the following steps:

Step 1, visual pattern input layer are input to the gray-scale value of gray level image in neural power network, there is one-to-one relationship in the neuron oscillator on the pixel in image and neural power network, neurodynamics system corresponding to each pixel is by FHN model description; Described visual pattern input layer, before gray-scale value being input to neural power network, has carried out normalization, and the gray-scale value after normalization is in [0,1] scope;

Step 2, neuron oscillator network oscillating layer, the dynamic system model coupling that each oscillator in neural network is set up according to FHN model forms neural power network, and i film potential V and voltage threshold R capable, a j row neuron oscillator are wherein respectively the rate of change of time:

Wherein, V is the potential difference (PD) of neuronal cell film both sides, is called film potential, and R is the internal state variable that represents current potential threshold value, and in following formula, V and R represent identical implication, I _{i, j}represent neuron receive from extraneous light stimulus, on numerical value, equal the gray-scale value in gray level image; α > 0, β > 0 has described respectively the action intensity from R to V and from V to R; Described neuron oscillator network oscillating layer, i in described formula (1) in (i, j) presentation video is capable, j row, 1≤i≤M, 1≤j≤N; Wherein, M and N are the wide and high of image; Δ V _{i, j}with Δ R _{i, j}represent the impact of peripheral nerve unit, they are defined by following formula:

Δx _i，j＝γ _{i-1，j-1；i，j}(x _i-1，j-1-x _i，j)+γ _{i-1，j；i，j}(x _i-1，j-x _i，j)+

γ _{i-1，j+1；i，j}(x _i-1，j+1-x _i，j)+γ _{i，j-1；i，j}(x _i，j-1-x _i，j)+

γ _{i，j+1；i，j}(x _i，j+1-x _i，j)+γ _{i+1，j-1；i，j}(x _i+1，j-1-x _i，j)+ (6)

γ _{i+1，j；i，j}(x _i+1，j-x _i，j)+γ _{i+1，j+1；i，j}(x _i+1，j+1-x _i，j)

Wherein

Wherein x represents V or R;

Step 3, attention mobility realize layer and in vibration, adjust α and β according to being defined in of following formula (2) and (3), thereby the corresponding neuron pool of the object of realizing current attention is provided frequency and is raise, and neuron pool corresponding to other objects produced to inhibition signal, make it provide frequency and slow down, distinguish current concern object and other objects with this;

Wherein, (p, q) pixel of the capable q row of p in presentation video, τ is the moment that has at least a neuron providing, τ-1st, the previous moment of τ, M (τ) is the neuronic quantity of moment τ at the state of granting, and Δ (τ) is the neuronic set of moment τ at the state of granting, I _{i, j}represent that i is capable, a j row neuron receives from extraneous light stimulus, equals the gray-scale value in gray level image on numerical value; Described FHN model is the abbreviation of FitzHugh-Nagumo model; Realize layer, h in attention mobility ₁(α _{p, q}(τ-1)), h ₂(β _{p, q}(τ-1)), f ₁(|| V _{i, j}-V _{p, q}||), f ₂(|| V _{i, j}-V _{p, q}||) implication of function is as follows:

f ₁(x)＝a ₁x+b ₁(a ₁＞0，b ₁＜0) (10)

f ₂(x)＝a ₂x+b ₂(a ₂＜0，b ₂＞0) (11)

In above formula (8), (9), (10), (11), θ ₁> θ ₂> 0, a ₁=2, b ₁=-4, a ₂=-4, b ₂=2, θ ₁=0.1, θ ₂=0.01, θ _α=0.8, θ _β=4.

2. the method with neuron network simulation attention mobility according to claim 1, is characterized in that:

Only have and just have stable limit cycle generation when being unsettled when equilibrium point (V, R), the value of α and β must meet formula (4)

8(V ²+αβ-1)＜0 (4)

Or formula (5)

8(V ²+αβ-1)＞0

(5)

-(10V ²-9.2)＞0。