CN101894295A

CN101894295A - A Method for Simulating Attention Shifts with Neural Networks

Info

Publication number: CN101894295A
Application number: CN2010101992275A
Authority: CN
Inventors: 段立娟; 房法明; 乔元华; 王海丽; 苗军; 吴春鹏
Original assignee: Beijing University of Technology
Current assignee: Beijing Hongzhou Culture Co ltd
Priority date: 2010-06-04
Filing date: 2010-06-04
Publication date: 2010-11-24
Anticipated expiration: 2030-06-04
Also published as: CN101894295B

Abstract

The invention discloses a method for simulating attention transfer by neural network, which includes a visual image input layer, a neuron oscillator network oscillation layer, and an attention transfer realization layer. The visual image input layer inputs the gray value of the grayscale image into the neural dynamic network, and the neuron oscillator network oscillation layer couples each oscillator in the neural network to the dynamic system model established according to the FitzHugh-Nagumo model to form neural dynamics In the network, the attention transfer implementation layer realizes the attention transfer between different synchronized objects by changing the parameters, and the frequency of the nerve firing of the object that is currently concerned will increase. Based on the neurodynamic system, the present invention forms a simple simulation of human attention shifting visual processing by analyzing and transforming the FitzHugh-Nagumo model, which has important theoretical and practical significance for further research on human visual processing mechanism.

Description

A Method for Simulating Attention Shifts with Neural Networks

技术领域technical field

本发明涉及以神经动力学为基础的图像处理和神经网络技术，特别是涉及一种以神经动力学模型描述的神经元振子形成的神经动力网络，用于模拟简单人造灰度图像中人的注意转移的方法。The invention relates to image processing and neural network technology based on neural dynamics, in particular to a neural dynamic network formed by neuron oscillators described by a neural dynamics model, which is used to simulate human attention in simple artificial grayscale images method of transfer.

背景技术Background technique

视觉注意选择和视觉注意转移是保证生物系统以有限的处理能力完成任务的重要机制。从图像或者感受野中提取显著的特征，并把它们分成不同的区域，从中选择显著的区域是感知理解的根本任务。这种能力就是视觉理解中的视觉注意选择。当显著的区域被选择之后，由于视觉系统的适应性，注意力会从当前显著的区域转到下一个显著的区域，这是注意转移。图像处理中基于同步振荡的神经元网络被越来越广泛的应用，脉冲耦合的神经元模型已经获得了良好的成果，但是在视觉注意选择和视觉注意转移中的应用对当前的研究来说仍然是非常重要的Visual attention selection and visual attention shift are important mechanisms to ensure that biological systems complete tasks with limited processing power. Extracting salient features from images or receptive fields and dividing them into different regions, selecting salient regions is the fundamental task of perceptual understanding. This ability is the visual attentional selection in visual comprehension. When the salient area is selected, due to the adaptability of the visual system, the attention will shift from the current salient area to the next salient area, which is attention shift. In image processing, neuron networks based on synchronous oscillations are more and more widely used, and the pulse-coupled neuron model has achieved good results, but the application in visual attention selection and visual attention transfer is still unreliable for current research. is very important

人眼处理视觉图像时可以在很短的时间内轻松的进行感知和理解，并且实现从一个显著的区域转移到下一个显著的区域，尽管这一方面的研究甚多，然而研究者对其中的机理至今仍然知之甚少。研究者们提出了很多神经元振子网络模型模拟真实的人脑中神经元行为，其中1952年在大量实验的基础上由Hodgkin和Huxley提出的Hodgkin-Huxley模型(以下简称HH模型)是第一个被广泛引用的模型。FitzHugh-Nagumo(以下简称FHN)模型是从HH模型简化而来，而且被多个研究者证明能较好的模拟神经元的电位发放，是模拟人的视觉的重要模型。When the human eye processes visual images, it can easily perceive and understand in a short period of time, and realize the transfer from one salient area to the next salient area. The mechanism is still poorly understood. Researchers have proposed many neuron oscillator network models to simulate the behavior of neurons in the real human brain, among which the Hodgkin-Huxley model (hereinafter referred to as the HH model) proposed by Hodgkin and Huxley on the basis of a large number of experiments in 1952 is the first widely cited model. The FitzHugh-Nagumo (hereinafter referred to as FHN) model is simplified from the HH model, and it has been proved by many researchers that it can better simulate the potential firing of neurons, and is an important model for simulating human vision.

FHN模型描述了一个神经元由输入电流I_input驱动的膜电位V和内部状态变量R之间的关系。V和R对时间t的变化率分别为：The FHN model describes the relationship between the membrane potential V and the internal state variable R of a neuron driven by the input current I _input . The rate of change of V and R with respect to time t is:

$\frac{dV dV}{dt dt} = = 1010 ((V V - - \frac{{V V}_{33}}{33} - - αR αR + + {I I}_{input input})) - - - - - - ((11))$

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

V 是神经元细胞膜两侧的电位差，称为膜电位，R是代表电位阈值的内部状态变量，I_input是神经元接受到的输入电流，也是外界对此神经元的光刺激。这里10和0.8是V和R的时间常数的倒数，并且α＞0，β＞0分别描述了从R到V和从V到R的动作强度。V的时间常数是R的12.5倍，这反应了轴突中激活过程要比恢复过程快得多的事实。假设(V，R)是平衡点，只有当平衡点(V，R)是不稳定的时候，在其周围产生稳定的极限环。因此α和β的取值必须满足公式(2)V is the potential difference on both sides of the neuron cell membrane, called the membrane potential, R is the internal state variable representing the potential threshold, I _input is the input current received by the neuron, and it is also the external light stimulus to the neuron. Here 10 and 0.8 are the reciprocals of the time constants of V and R, and α > 0, β > 0 describe the motion intensity from R to V and from V to R, respectively. The time constant of V is 12.5 times that of R, reflecting the fact that the activation process in axons is much faster than the recovery process. Assuming (V, R) is an equilibrium point, only when the equilibrium point (V, R) is unstable, a stable limit cycle is generated around it. Therefore, the values of α and β must satisfy the formula (2)

8(V²+αβ-1)＜0(2)8(V ² +αβ-1)＜0(2)

或者公式(3)or formula (3)

8(V²+αβ-1)＞08(V ² +αβ-1)>0

-(10V²-9.2)＞0(3)-(10V ² -9.2)＞0(3)

改变方程(1)中的参数α和β可以控制发放频率。α或者β的增加都会引起发放频率的增加，但当α或者β是一个相当小的值时(例如α＝1，β＝0.6或者α＝0.3，β＝2或者α＝0.3，β＝0.6)，方程(1)不产生发放。The firing frequency can be controlled by changing the parameters α and β in equation (1). The increase of α or β will cause the increase of firing frequency, but when α or β is a relatively small value (such as α=1, β=0.6 or α=0.3, β=2 or α=0.3, β=0.6) , Equation (1) does not produce a release.

发明内容Contents of the invention

本发明的目的在于，通过提供一种用神经网络模拟注意转移的方法，应用改造的FHN模型形成神经动力网络模拟人眼在感受野中从一个显著性区域到另一个的过程，并构建简单感受野中注意转移系统，对单个神经元进行模型构建，并基于单个模型构造生物视觉神经网络。The purpose of the present invention is, by providing a method for simulating attention transfer with a neural network, applying the transformed FHN model to form a neural dynamic network to simulate the process of the human eye from one salient region to another in the receptive field, and construct a simple receptive field Attention transfer system, model construction of single neuron, and construct biological visual neural network based on single model.

本发明用神经网络模拟注意转移的方法，是采用以下技术手段实现的。The method for simulating attention transfer by neural network in the present invention is realized by adopting the following technical means.

步骤1、视觉图像输入层将灰度图像的灰度值输入到神经动力网络中，图像中的像素点与神经动力网络上的神经元振子存在一一对应关系，每个像素对应的神经动力学系统由FHN模型描述。Step 1. The visual image input layer inputs the grayscale value of the grayscale image into the neural dynamic network. There is a one-to-one correspondence between the pixels in the image and the neuron oscillators on the neural dynamic network, and the neural dynamics corresponding to each pixel The system is described by the FHN model.

步骤2、神经元振子网络振荡层，将神经网络中的每个振子根据FHN模型建立起来的动力学系统模型耦合形成神经动力网络，其中的第i行、第j列个神经元振子的膜电位V和电压阈值R对时间的变化率分别为；Step 2, neuron oscillator network oscillation layer, coupling each oscillator in the neural network with the dynamic system model established according to the FHN model to form a neural dynamic network, in which the membrane potential of the i-th row and j-th column neuron oscillator The rate of change of V and the voltage threshold R to time are respectively;

$\{\begin{matrix} \frac{{dV dV}_{i i,, j j}}{dt dt} = = 1010 (((({V V}_{i i,, j j} + + Δ Δ {V V}_{i i,, j j})) - - \frac{{(({V V}_{i i,, j j} + + Δ Δ {V V}_{i i,, j j}))}^{33}}{33} - - α α {R R}_{i i,, j j} + + {I I}_{i i,, j j})) \\ \frac{d d {R R}_{i i,, j j}}{dt dt} = = 0.8 0.8 ((- - (({R R}_{i i,, j j} + + Δ Δ {R R}_{i i,, j j})) + + β β {V V}_{i i,, j j} + + 1.5 1.5)) \end{matrix} - - - - - - ((44))$

其中，V是神经元细胞膜两侧的电位差，称为膜电位，R是代表电位阈值的内部状态变量，以下的公式中V和R代表相同的含义，I_i，j表示第i行、第j列个神经元接受到的来自外界的光刺激，数值上等于灰度图像中的灰度值。Among them, V is the potential difference on both sides of the neuron cell membrane, called the membrane potential, and R is the internal state variable representing the potential threshold. In the following formulas, V and R represent the same meaning, and I _{i, j} represent the i-th row, the The light stimulus received by the j neurons from the outside world is numerically equal to the gray value in the gray image.

步骤3、注意转移实现层根据以下公式(5)和(6)的定义在振荡中调整α和β，从而实现当前注意的物体对应神经元群发放频率升高，并对其他物体对应的神经元群产生抑制性信号，使其发放频率放缓，以此区别当前关注物体与其他物体。Step 3, the attention transfer implementation layer adjusts α and β in the oscillation according to the definition of the following formulas (5) and (6), so as to increase the firing frequency of the neuron group corresponding to the object currently paying attention, and increase the firing frequency of the neuron group corresponding to other objects. The swarm produces inhibitory signals that slow down its firing frequency, thereby distinguishing the currently focused object from other objects.

${α α}_{p p,, q q} ((τ τ)) = = {α α}_{p p,, q q} ((τ τ - - 11)) + + \frac{{h h}_{11} (({α α}_{p p,, q q} ((τ τ - - 11))))}{M m ((τ τ))} \underset{i i,, j j &Element; &Element; Δ Δ ((τ τ))}{Σ Σ} {I I}_{i i,, j j} {f f}_{11} ((| | | | {V V}_{i i,, j j} - - {V V}_{p p,, q q} | | | |)) - - - - - - ((55))$

${β β}_{p p,, q q} ((τ τ)) = = {β β}_{p p,, q q} ((τ τ - - 11)) + + \frac{{h h}_{22} (({β β}_{p p,, q q} ((τ τ - - 11))))}{M m ((τ τ))} \underset{i i,, j j &Element; &Element; Δ Δ ((τ τ))}{Σ Σ} {I I}_{i i,, j j} {f f}_{22} ((| | | | {V V}_{i i,, j j} - - {V V}_{p p,, q q} | | | |)) - - - - - - ((66))$

其中，(p，q)表示图像中p行q列的像素，τ是至少有一个神经元在发放的时刻，τ-1是τ的前一个时刻，M(τ)是时刻τ在发放状态的神经元的数量，Δ(τ)是时刻τ在发放状态的神经元的集合，I_i，j表示第i行、第j列个神经元接受到的来自外界的光刺激，数值上等于灰度图像中的灰度值。Among them, (p, q) represents the pixels in the p row and q column in the image, τ is the moment when at least one neuron is firing, τ-1 is the previous moment of τ, and M(τ) is the time when τ is firing. The number of neurons, Δ(τ) is the set of neurons in firing state at time τ, I _{i, j} represents the light stimulation from the outside world received by neurons in row i and column j, numerically equal to grayscale Grayscale values in the image.

所述的FHN模型为FitzHugh-Nagumo模型的简称。The FHN model is an abbreviation of the FitzHugh-Nagumo model.

前述的视觉图像输入层，在将灰度值输入到神经动力网络之前，进行了归一化，归一化之后的灰度值在[0，1]范围内。The aforementioned visual image input layer is normalized before the gray value is input to the neural dynamic network, and the gray value after normalization is within the range of [0, 1].

只有当平衡点(V，R)是不稳定的时候才会有稳定的极限环产生，α和β的取值必须满足公式(2)Only when the equilibrium point (V, R) is unstable will there be a stable limit cycle, and the values of α and β must satisfy the formula (2)

8(V²+αβ-1)＜0(2)8(V ² +αβ-1)＜0(2)

或者公式(3)or formula (3)

8(V²+αβ-1)＞08(V ² +αβ-1)>0

-(10V²-9.2)＞0°(3)-(10V ² -9.2)＞0°(3)

前述的神经元振子网络振荡层，在所述的公式(4)中(i，j)表示图像中的第i行、第j列，1≤i ≤M，1≤j≤N；The aforementioned neuron oscillator network oscillation layer, in the formula (4), (i, j) represents the i-th row and j-th column in the image, 1≤i≤M, 1≤j≤N;

其中，M和N是图像的宽和高；ΔV_i，j和ΔR_i，j代表周围神经元的影响，它们由以下式子定义：Among them, M and N are the width and height of the image; ΔV _i,j and ΔR _i,j represent the influence of surrounding neurons, which are defined by the 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)+Δxi _{, j} = γ _{i-1, j-1; i, j} (xi _{-1, j-1} -xi _{, j} )+γ _{i-1, j; i, j} (xi _{-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-1, j+1; i, j} (xi _{-1, j+1} -xi _{, j} )+γ _{i, j-1; i, j} (xi _{, j-1} -xi _{, 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)+γ _{i, j+1; i, j} (xi _{, j+1} -xi _{, j} )+γ _{i+1, j-1; i, j} (xi _{+1, j-1} -xi _{, 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)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} )

其中

in

其中x表示V或者R。Where x represents V or R.

前述的注意转移实现层，在以下公式(9)和(10)中θ₁＞θ₂＞0In the aforementioned attention transfer implementation layer, θ ₁ >θ ₂ >0 in the following formulas (9) and (10)

${h h}_{11} ((α α)) = = \{\begin{matrix} {θ θ}_{11} & α α < < {θ θ}_{α α} \\ {θ θ}_{22} & α α &GreaterEqual; &Greater Equal; {θ θ}_{α α} \end{matrix} - - - - - - ((99))$

${h h}_{22} ((β β)) = = \{\begin{matrix} {θ θ}_{11} & β β &GreaterEqual; &Greater Equal; {θ θ}_{β β} \\ {θ θ}_{22} & β β < < {θ θ}_{β β} \end{matrix} - - - - - - ((1010))$

f₁(x)＝a₁x+b₁(a₁＞0，b₁＜0)(11)f ₁ (x)=a ₁ x+b ₁ (a ₁ >0, b ₁ <0)(11)

f₂(x)＝a₂x+b₂(a₂＜0，b₂＞0)(12)f ₂ (x)=a ₂ x+b ₂ (a ₂ <0, b ₂ >0)(12)

实施过程参数为a₁＝2，b₁＝-4，a₂＝-4，b₂＝2，θ₁＝0.1，θ₂＝0.01，θ_α＝0.8，θ_β＝4。The implementation process parameters are 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.

本发明一种用神经网络模拟注意转移的方法，与以往文献中使用FHN模型的方法及其应用领域相比，具有以下明显的优势和有益的效果：A kind of method of neural network simulation attention transfer of the present invention, compared with the method and its application field using FHN model in the past literature, has the following obvious advantages and beneficial effects:

1、神经元之间的耦合方式不同于以往文献。1. The coupling mode between neurons is different from the previous literature.

2、把FHN模型用于模拟人的注意转移机制。2. The FHN model is used to simulate the human attention shifting mechanism.

附图说明Description of drawings

图1注意转移系统结构示意图；Fig. 1 is a schematic structural diagram of attention transfer system;

图2为图像与神经动力网络对应示意图；Fig. 2 is a schematic diagram corresponding to an image and a neural dynamic network;

图3为输入图像示意图；Fig. 3 is a schematic diagram of an input image;

图4为本发明中方法对图3的依次关注物体图。Fig. 4 is a sequential attention object map of Fig. 3 by the method in the present invention.

具体实施方式Detailed ways

以下结合附图对本发明的具体实施方式加以说明。Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings.

该系统分为三个部分，视觉图像输入层，神经元振子网络振荡层，注意转移实现层。三层之间的关系如图1注意转移系统结构示意图所示。The system is divided into three parts, the visual image input layer, the neuron oscillator network oscillation layer, and the attention transfer implementation layer. The relationship between the three layers is shown in Fig. 1, the structural diagram of the attention transfer system.

该系统首先将一幅视觉图像中的各个像素与神经动力网络中的所有神经元振子建立一对一关联，然后让这一组神经元振子产生振荡。由于亮度大的物体对人眼视网膜的光刺激更强烈，所以I_input的增大导致发放频率的增大，而发放过程中会对其他区域产生抑制。The system first establishes a one-to-one association between each pixel in a visual image and all the neuron oscillators in the neurodynamic network, and then makes this group of neuron oscillators oscillate. Since objects with high brightness stimulate the retina of the human eye more strongly, the increase in I _input leads to an increase in the firing frequency, and other areas will be inhibited during the firing process.

把单个的由FHN模型描述的神经元以矩阵的形式排列并进行模型重构，形成神经动力网络。新建立的模型中每一个神经元振子除边缘外都与其周边8领域内的神经元有耦合关系，边缘的神经元振子根据其相邻振子的不同分别与3个或者5个神经元振子耦合。神经动力网络中第i行、第j列的神经元振子的膜电位V和电压阈值R对时间的变化率分别为：Arrange the individual neurons described by the FHN model in the form of a matrix and reconstruct the model to form a neural dynamic network. In the newly established model, each neuron oscillator has a coupling relationship with the neurons in the surrounding 8 fields except the edge, and the neuron oscillators at the edge are coupled with 3 or 5 neuron oscillators according to the difference between the adjacent oscillators. The rate of change of the membrane potential V and the voltage threshold R of the neuron oscillator in the i-th row and j-th column of the neurodynamic network with respect to time are:

方程中(i，j)表示图像中第i行、第j列的像素，也代表神经动力网络中第i行、第j列的神经元振子，1≤i≤M，1≤j≤N(M和N是图像的宽和高)。ΔV_i，j和ΔR_i，j代表周围8邻域的神经元的影响，它们由以下式子定义：In the equation (i, j) represents the pixel in the i-th row and j-th column in the image, and also represents the neuron oscillator in the i-th row and j-th column in the neural dynamic network, 1≤i≤M, 1≤j≤N( M and N are the width and height of the image). ΔV _i,j and ΔR _i,j represent the influence of neurons in the surrounding 8 neighborhoods, which are defined by the 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)+Δx _i,j =γ _i-1,j-1,i,j (xi _-1,j-1 -xi _,j )+γ _{i-1,j; i,j} (xi _-1,j -x _{i, j} )+

(7)...

其中

其中x表示V或者R。in

Where x represents V or R.

为了实现对应显著物体的神经元更频繁的发放，其他物体的神经元以较低频率发放或者不发放这样的事实，我们首先让神经动力网络带固定参数α和β运行，直到对应相同区域的神经元同步，这也表明完成了分割任务。之后，不论何时任何神经元发放，它会对自己和其他神经元产生两种信号：对自己和一起发放的神经元产生兴奋性的信号，没有随它一起发放的是抑制性的信号。To realize the fact that neurons corresponding to salient objects fire more frequently and neurons on other objects fire less frequently or not, we first run the neural dynamics network with fixed parameters α and β until neurons corresponding to the same region Meta-sync, which also indicates the completion of the split task. Then, whenever any neuron fires, it generates two kinds of signals to itself and to other neurons: an excitatory signal to itself and to the neuron that fires with it, and an inhibitory signal to the neuron that doesn't fire with it.

系统运行时，我们可以控制参数α和β的改变，使得α的增加速度比β的减小速度大的多。这样最显著的区域将会跳到高频周期振荡相位，而其他区域会相对安静，注意也会转移到最显著的区域。当接受到注意之后，此区域会受到抑制以允许其他区域成为显著的。When the system is running, we can control the changes of parameters α and β, so that the increasing speed of α is much larger than the decreasing speed of β. In this way, the most prominent area will jump to the high-frequency periodic oscillation phase, while other areas will be relatively quiet, and attention will also be shifted to the most prominent area. After receiving attention, this area is suppressed to allow other areas to become salient.

从上面对方程(1)的分析我们可以得出，α和β可以控制神经元的活动性。From the above analysis of equation (1), we can conclude that α and β can control the activity of neurons.

一个神经元发放后会引起周边神经元α和β参数的改变，从而实现了注意的转移。其中α和β的变化由以下方程定义：After a neuron fires, it will cause changes in the α and β parameters of peripheral neurons, thereby realizing the transfer of attention. where the variation of α and β is defined by the following equations:

${α α}_{p p,, q q} ((τ τ)) = = {α α}_{p p,, q q} ((τ τ - - 11)) + + \frac{{h h}_{11} (({α α}_{p p,, q q} ((τ τ - - 11))))}{M m ((τ τ))} \underset{i i,, j j &Element; &Element; ((τ τ))}{Σ Σ} {I I}_{i i,, j j} {f f}_{11} ((| | | | {V V}_{i i,, j j} - - {V V}_{p p,, q q} | | | |)) - - - - - - ((55))$

其中in

${h h}_{11} ((α α)) = = \{\begin{matrix} {θ θ}_{11} & α α < < {θ θ}_{α α} \\ {θ θ}_{22} & α α &GreaterEqual; &Greater Equal; {θ θ}_{α α} \end{matrix},, {θ θ}_{11} > > {θ θ}_{22} > > 00 - - - - ((99))$

${h h}_{22} ((β β)) = = \{\begin{matrix} {θ θ}_{11} & β β &GreaterEqual; &Greater Equal; {θ θ}_{β β} \\ {θ θ}_{22} & β β < < {θ θ}_{β β} \end{matrix} {θ θ}_{11} > > {θ θ}_{22} > > 00 - - - - - - ((1010))$

(p，q)表示图像中p行q列的像素，τ是至少有一个神经元在发放的时刻，M(τ)是时刻τ在发放状态的神经元的数量，Δ(τ)是时刻τ在发放状态的神经元的集合，I_i，j表示神经元接受到的来自外界的光刺激，数值上等于灰度图像中的灰度值。通过设置a₁＞0，b₁＜0和a₂＜0，b₂＞0，以及函数h₁，h₂，每个发放中的神经元(i，j)向(p，q)发放兴奋性或者抑制性的信号。(p, q) represents the pixel in row p and column q in the image, τ is the moment when at least one neuron is firing, M(τ) is the number of neurons firing at time τ, Δ(τ) is the moment τ The set of neurons in the firing state, I _{i, j} represents the light stimulus received by the neuron from the outside world, which is numerically equal to the gray value in the gray image. By setting a ₁ >0, b ₁ <0 and a ₂ <0, b ₂ >0, and functions h ₁ , h ₂ , neuron (i, j) in each firing fires to (p, q) Sexual or inhibitory signals.

首先是建立视觉图像输入层。根据公式(1)和公式(4)确立的系统，建立神经动力网络，并将网络上的每个神经元与图像中的像素建立一一对应的关系，而图像的灰度值也成为神经动力系统的输入值，对应公式(1)中的I_input。附图2为神经动力网络与灰度图像一一对应图，每一个小圆圈既代表图像中的一个像素，也代表图像对应的神经动力网络中的一个神经元，黑点代表可数个数的神经元，横线代表神经元之间的连接。The first is to build the visual image input layer. According to the system established by formula (1) and formula (4), establish a neural power network, and establish a one-to-one correspondence between each neuron on the network and the pixel in the image, and the gray value of the image also becomes the neural power The input value of the system corresponds to I _input in formula (1). Attached Figure 2 is a one-to-one correspondence between the neural dynamic network and the grayscale image. Each small circle represents not only a pixel in the image, but also a neuron in the corresponding neural dynamic network of the image. Black dots represent countable Neurons, horizontal lines represent connections between neurons.

其次是神经元振子网络振荡层。通过对整个的神经动力网络中每个神经元振子用4阶Runge-Kutta法求解，实现同一物体内的振子达到同步振荡，而不同物体间的振子则去同步。The second is the neuron oscillator network oscillation layer. By using the fourth-order Runge-Kutta method to solve each neuron oscillator in the entire neural dynamic network, the oscillators in the same object can achieve synchronous oscillation, while the oscillators between different objects are desynchronized.

最后是注意转移实现层。根据公式(5)、公式(6)中的定义改变α和β，结果是按照显著性的大小图像中的物体依次选出，选出的物体被注意到时对应神经振子的发放频率会变大。The last is the attention transfer implementation layer. Change α and β according to the definition in formula (5) and formula (6), the result is that the objects in the image are selected in order according to the size of the salience, and the firing frequency of the corresponding neural oscillator will become larger when the selected object is noticed .

接下来用一幅具体图像图3说明实施流程。图3是一幅灰度图，由4个目标组成，其中①是太阳，对应灰度值255，②是树，对应灰度值195，③是山峦，对应灰度值130，④是天空，对应灰度值65。首先将图3进行归一化操作，也就是将其上像素值取值范围从[0，255]变为[0，1]。将归一化后的视觉图像输入到建立好的动力学系统(4)中，建立好的动力学系统是指按照公式(4)，(5)，(6)建立起来的系统。接下来对(4)式利用4阶经典Runge-Kutta法进行求解，求解过程中参数α＝1，β＝1.25首先自行保持不变，直至神经动力网络中的局部区域中的神经元振子达到同步共振，在求解过程中求解结果V和R得到周期解。在注意转移实现层中使用的参数分别为a₁＝2，b₁＝-4，a₂＝-4，b₂＝2，θ₁＝0.1，θ₂＝0.01，θ_α＝0.8，θ_β＝4。通过迭代求解，图3中4个物体对应的发放频率依照①②③④的顺序分别增大，在某个物体增大的同时抑制其他物体的发放频率，是其他3个物体的发放频率变缓，由此实现注意转移在①②③④4个物体间不断切换。在图3中依次选出的物体如图4依次关注物体图所示。Next, a specific image, Figure 3, is used to illustrate the implementation process. Figure 3 is a grayscale image consisting of 4 objects, among which ① is the sun, corresponding to a grayscale value of 255, ② is a tree, corresponding to a grayscale value of 195, ③ is a mountain, corresponding to a grayscale value of 130, ④ is the sky, Corresponds to a grayscale value of 65. Firstly, the normalization operation is performed on Figure 3, that is, the range of pixel values on it is changed from [0, 255] to [0, 1]. The normalized visual image is input into the established dynamic system (4), and the established dynamic system refers to the system established according to formulas (4), (5), and (6). Next, formula (4) is solved by the 4th-order classic Runge-Kutta method. During the solution process, the parameters α=1 and β=1.25 are kept unchanged until the neuron oscillators in the local area of the neural dynamical network are synchronized. Resonance, during the solution process, solve the results V and R to obtain a periodic solution. The parameters used in the attention transfer realization layer are 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. Through iterative solution, the release frequencies corresponding to the four objects in Figure 3 increase in the order of ①②③④. When an object increases, the release frequency of other objects is suppressed, and the release frequency of the other 3 objects slows down. Realize attention transfer and switch continuously among ①②③④4 objects. The objects sequentially selected in Fig. 3 are shown in Fig. 4, which is followed by objects sequentially.

最后应说明的是：以上实施例仅用以说明本发明而并非限制本发明所描述的技术方案；因此，尽管本说明书参照上述的各个实施例对本发明已进行了详细的说明，但是，本领域的普通技术人员应当理解，仍然可以对本发明进行修改或等同替换；而一切不脱离发明的精神和范围的技术方案及其改进，其均应涵盖在本发明的权利要求范围当中。Finally, it should be noted that: the above embodiments are only used to illustrate the present invention rather than limit the technical solutions described in the present invention; Those of ordinary skill in the art should understand that the present invention can still be modified or equivalently replaced; and all technical solutions and improvements that do not depart from the spirit and scope of the invention should be covered by the claims of the present invention.

Claims

1. A method for simulating attention transfer with neural network, is characterized in that comprising the following steps:

Step 1. The visual image input layer inputs the grayscale value of the grayscale image into the neural dynamic network. There is a one-to-one correspondence between the pixels in the image and the neuron oscillators on the neural dynamic network, and the neural dynamics corresponding to each pixel The system is described by the FHN model;

Step 2, neuron oscillator network oscillation layer, coupling each oscillator in the neural network with the dynamic system model established according to the FHN model to form a neural dynamic network, in which the membrane potential of the i-th row and j-th column neuron oscillator The rate of change of V and the voltage threshold R with respect to time are:

Among them, V is the potential difference on both sides of the neuron cell membrane, called the membrane potential, and R is the internal state variable representing the potential threshold. In the following formulas, V and R represent the same meaning, and I _{i, j} represent the neuron received The light stimulus from the outside is numerically equal to the gray value in the gray image;

Step 3. The attention transfer implementation layer adjusts α and β in the oscillation according to the definition of the following formulas (2) and (3), so as to increase the firing frequency of the neuron group corresponding to the object currently paying attention, and to increase the firing frequency of the neuron group corresponding to other objects. The group produces inhibitory signals to slow down its emission frequency, so as to distinguish the current attention object from other objects;

Among them, (p, q) represents the pixels in the p row and q column in the image, τ is the moment when at least one neuron is firing, τ-1 is the previous moment of τ, and M(τ) is the time when τ is firing. The number of neurons, Δ(τ) is the set of neurons in firing state at time τ, I _{i, j} represents the light stimulation from the outside world received by neurons in row i and column j, numerically equal to grayscale the grayscale value in the image;

The FHN model is an abbreviation of the FitzHugh-Nagumo model.

2. the method for simulating attention shifting with neural network according to claim 1, is characterized in that: described visual image input layer, before grayscale value is input to neural power network, has carried out normalization, normalization The gray value after conversion is in the range of [0, 1].

3. according to claim 1, use neural network simulation to pay attention to the method of shifting, it is characterized in that: only when equilibrium point (V, R) is unstable time just can have stable limit cycle to produce, the getting of α and β The value must satisfy formula (4)

8(V ² +αβ-1)＜0(4)

or formula (5)

8(V ² +αβ-1)>0

-(10V ² -9.2) > 0° (5).

4. the method for imitating attention with neural network according to claim 1 is characterized in that: described neuron oscillator network oscillation layer, in described formula (1), (i, j) represent in the image row i, column j, 1≤i≤M, 1≤j≤N;

Among them, M and N are the width and height of the image; ΔV _i,j and ΔR _i,j represent the influence of surrounding neurons, which are defined by the following formula:

Δxi _{, j} = γ _{i-1, j-1; i, j} (xi _{-1, j-1} -xi _{, j} )+γ _{i-1, j; i, j} (xi _{-1, j} -x _{i, j} )+

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

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

in

Where x represents V or R.

5. the method for simulating attention transfer with neural network according to claim 1, is characterized in that: described attention transfer realizes layer, in following formula (8) and (9) θ ₁ > θ ₂ > 0

f ₁ (x)=a ₁ x+b ₁ (a ₁ >0, b ₁ <0)(10)

f ₂ (x)=a ₂ x+b ₂ (a ₂ <0, b ₂ >0)(11)

The implementation process parameters are 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.