CN102682297B

CN102682297B - Pulse coupled neural network (PCNN) face image segmenting method simulating visual cells to feel field property

Info

Publication number: CN102682297B
Application number: CN201210137335.9A
Authority: CN
Inventors: 杨娜; 王浩全
Original assignee: North University of China
Current assignee: North University of China
Priority date: 2012-05-07
Filing date: 2012-05-07
Publication date: 2014-05-14
Anticipated expiration: 2032-05-07
Also published as: CN102682297A

Abstract

The invention relates to a pulse-coupled neural network face segmentation method for disabled persons who simulates the receptive field characteristics of visual cells. Directional and scaled pulse-coupled neural network models; model parameters are adjusted according to the characteristics of disabled face images; finally, the brightness channel information of disabled face images is input into the model to generate face segmentation results that simulate human vision. Due to the optimization of the connection matrix by the cell receptive field model, the pulse-coupled neural network has directionality and scale, which improves the accuracy of segmentation and has good robustness for face segmentation under natural light. In addition, compared with other methods, the present invention also has the advantages of good separation between different image contents, good image detail preservation, fast segmentation speed and the like.

Description

PCNN Face Image Segmentation Method Simulating Visual Cell Receptive Field Characteristics

技术领域 technical field

本发明涉及图像处理领域，尤其涉及一种模拟视觉细胞感受野特性的脉冲耦合神经网络(IG-PCNN)的残疾人人脸图像分割方法。The invention relates to the field of image processing, in particular to a method for segmenting face images of disabled people using a pulse-coupled neural network (IG-PCNN) that simulates the receptive field characteristics of visual cells.

背景技术 Background technique

作为人类最重要的外部特征，人脸图像的检测和识别技术日益成为人工智能领域中的研究热点，其在国家安全、公安民政、金融海关、保险等领域具有极为广阔的应用前景，然而不论人脸检测还是识别(检测是确定人脸的存在，以及确定图像中人脸的位置。识别是在检测的基础上识别出人脸)，都要做人脸图像的分割，这是图像处理的第一步，分割的好坏对特征提取和识别具有关键性作用。因此，人脸图像分割是人脸图像检测和识别领域中的基本问题，是目标特征提取、识别与跟踪的基础。As the most important external feature of human beings, face image detection and recognition technology has increasingly become a research hotspot in the field of artificial intelligence. It has extremely broad application prospects in national security, public security and civil affairs, financial customs, insurance and other fields. Face detection or recognition (detection is to determine the existence of the face, and determine the position of the face in the image. Recognition is to recognize the face on the basis of detection), it is necessary to segment the face image, which is the first step in image processing The quality of segmentation plays a key role in feature extraction and recognition. Therefore, face image segmentation is a basic problem in the field of face image detection and recognition, and it is the basis of target feature extraction, recognition and tracking.

目前存在的人脸分割、检测、识别算法都是针对正常人的人脸。主要的方法有基于肤色的人脸检测，基于虹膜的人脸检测与识别，Harr特征的人脸检测、肤色及Adaboost的人脸检测、支持向量机核函数的人脸检测、小波变换提取特征的人脸检测与识别等算法。The currently existing face segmentation, detection, and recognition algorithms are all aimed at the faces of normal people. The main methods are face detection based on skin color, face detection and recognition based on iris, face detection of Harr feature, face detection of skin color and Adaboost, face detection of support vector machine kernel function, and wavelet transform feature extraction. Algorithms for face detection and recognition.

由于残疾人的脸部具有独特的特征，目前的人脸图像检测与识别算法不适用这类特殊的人群，比如残疾人中盲人的眼睛区域没有正常人眼睛处灰度变化丰富，因此用眼睛区域定位比较困难，眼睛区域会跟人脸的其它区域粘连，给后期的特征提取造成困难。面部有伤的残疾人，伤疤可以作为很好的识别特征，因此，如何最大限度体现伤疤区域与完好脸部区域的区别，将伤疤区域完整地分割出来，且不与其相邻完好脸部区域粘连是现有的图像分割方法所面临的挑战。同时，精神残疾人脸部表情变化大，面部五官由于抽搐会很接近，需要将面部五官分割成独立的区域，为后续特征提取奠定基础，降低识别的难度。因此，本发明提出针对残疾人人脸的图像分割算法。Due to the unique characteristics of the faces of disabled people, the current face image detection and recognition algorithms are not suitable for such special groups of people. Positioning is more difficult, and the eye area will be glued to other areas of the face, which will make it difficult to extract features in the later stage. For disabled persons with facial injuries, scars can be used as a good identification feature. Therefore, how to maximize the difference between the scar area and the intact face area, completely segment the scar area, and not stick to the adjacent intact face area is the challenge faced by existing image segmentation methods. At the same time, the facial expressions of mentally disabled people change greatly, and the facial features will be very close due to twitching. It is necessary to divide the facial features into independent areas to lay the foundation for subsequent feature extraction and reduce the difficulty of recognition. Therefore, the present invention proposes an image segmentation algorithm for disabled faces.

生物视觉始终是图像处理领域研究的热点，基于脉冲耦合神经网络(PCNN)的图像分割方法具有突出的优良特性：变阈值特性、非线性调制特性、同步脉冲发放及神经元捕获特性、动态脉冲发放及自动波特性等。由于上述神经元模型是基于哺乳动物视觉皮层神经元活动提出的，基于PCNN的人脸图像分割方法完全依赖于图像的自然属性，不用预先选择处理的空间范围，是一种更自然的分割方式。通过调节神经元的连接强度，可方便地对图像进行不同层次的分割且分割速度很快。该方法在图像处理特别是图像分割方面存在较强的优势。Biological vision has always been a research hotspot in the field of image processing. The image segmentation method based on pulse-coupled neural network (PCNN) has outstanding characteristics: variable threshold characteristics, nonlinear modulation characteristics, synchronous pulse emission and neuron capture characteristics, dynamic pulse emission And automatic wave characteristics. Since the above neuron model is proposed based on the neuron activity of the mammalian visual cortex, the face image segmentation method based on PCNN is completely dependent on the natural attributes of the image, and it does not need to pre-select the spatial range of processing, which is a more natural segmentation method. By adjusting the connection strength of neurons, the image can be easily segmented at different levels and the segmentation speed is very fast. This method has strong advantages in image processing, especially in image segmentation.

如图1所示，传统的PCNN模型中单个神经元由反馈输入域、耦合连接域和脉冲发生器三部分组成，最终形成一个单层的、二维横向连接的神经网络。神经元的激发受其邻域内神经元的影响，影响的范围和程度由连接系数矩阵表示。PCNN模型中有两个连接系数矩阵M_ijkl和W_ijkl，它们都表示中心神经元受周围神经元影响的大小，反映邻近神经元对中心神经元传递信息的强弱。其中W_ijkl位于PCNN的耦合连接域，主要表示神经网络内部中心神经元与相邻神经元的连接强度。M_ijkl位于PCNN的反馈输入域，主要功能是获取神经元外部灰度信息，突出外部神经元对中心神经元的影响。As shown in Figure 1, a single neuron in the traditional PCNN model consists of three parts: the feedback input domain, the coupled connection domain, and the pulse generator, and finally forms a single-layer, two-dimensional horizontally connected neural network. The excitation of a neuron is affected by the neurons in its neighborhood, and the scope and degree of influence are represented by the connection coefficient matrix. There are two connection coefficient matrices M _ijkl and W _ijkl in the PCNN model, both of which represent the size of the influence of the central neuron on the surrounding neurons, and reflect the strength of the adjacent neurons transmitting information to the central neuron. Among them _{, Wijkl} is located in the coupled connection domain of PCNN, which mainly indicates the connection strength between the central neuron and adjacent neurons in the neural network. M _ijkl is located in the feedback input domain of PCNN, and its main function is to obtain the external grayscale information of neurons and highlight the influence of external neurons on the central neuron.

目前与PCNN模型相关的文献普遍认为M_ijkl和W_ijkl的作用是一样的，在数值上是相等的。尤其在简化PCNN模型中，反馈输入域连接矩阵M_ijkl的功能被进一步弱化，将反馈输入域简化为中心神经元接收的外部激励，即F[n]＝S_ij。本发明通过对PCNN模型工作机理的分析，得出反馈输入域连接矩阵M_ijkl是PCNN模型中获取外部图像信息的重要结构，其大小直接决定了神经元耦合域的范围，大耦合域的连接系数矩阵直接影响着自动波的传播速度和传播的距离，决定了中心神经元能否被更远的神经元所捕获。但连接矩阵不具有方向性，不能在指定方向上加强中心神经元与邻域神经元的联系。At present, the literature related to the PCNN model generally believes that the functions of M _ijkl and W _ijkl are the same, and they are equal in value. Especially in the simplified PCNN model, the function of the connection matrix M _ijkl of the feedback input domain is further weakened, and the feedback input domain is simplified as the external excitation received by the central neuron, that is, F[n]=S _ij . Through the analysis of the working mechanism of the PCNN model, the present invention obtains that the feedback input domain connection matrix M _ijkl is an important structure for obtaining external image information in the PCNN model, and its size directly determines the scope of the neuron coupling domain, and the connection coefficient of the large coupling domain The matrix directly affects the propagation speed and distance of automatic waves, and determines whether the central neuron can be captured by more distant neurons. However, the connection matrix has no directionality, and cannot strengthen the connection between the central neuron and the neighboring neurons in the specified direction.

基于上述研究，本发明提出一种基于具有感受野特性PCNN模型的残疾人人脸图像分割方法。Based on the above research, the present invention proposes a face image segmentation method for disabled persons based on a PCNN model with receptive field characteristics.

发明内容 Contents of the invention

针对传统的基于PCNN模型的图像分割方法所存在的问题，本发明提出了一种具有感受野特性的PCNN模型用于残疾人人脸图像的分割方法。Aiming at the problems existing in the traditional image segmentation method based on the PCNN model, the present invention proposes a PCNN model with receptive field characteristics for the segmentation method of the disabled face image.

一种模拟视觉细胞感受野特性的PCNN人脸图像分割方法，所述PCNN中的神经元包括接收域、调制域和脉冲发生器三部分，所述接收域包括反馈接收域和连接接收域，反馈接收域接受图像灰度值S_ij和感受野内相邻神经元的输出脉冲Y_kl作为输入，经过感受野矩阵IG变换输出F_ij作为神经元的反馈输入，连接接收域接受感受野内相邻神经元的脉冲输出Y_kl作为输入，经过连接矩阵W变换输出L_ij作为神经元的耦合连接输入，所述方法包括以下步骤：A PCNN face image segmentation method simulating visual cell receptive field characteristics, the neuron in the PCNN includes three parts of receiving domain, modulation domain and pulse generator, and the receiving domain includes feedback receiving domain and connection receiving domain, feedback The receiving field accepts the image gray value S _ij and the output pulse Y _kl of the adjacent neuron in the receptive field as input, and the output F _ij is transformed by the receptive field matrix IG as the feedback input of the neuron, and the receiving field is connected to accept the adjacent neuron in the receptive field The pulse output Y _kl is used as input, and the output L _ij is used as the coupled connection input of the neuron through the transformation of the connection matrix W, and the method includes the following steps:

a)将采集到的残疾人人脸图像由RGB空间转换到HSV空间，提取图像的亮度通道信息；a) Convert the collected face image of the disabled from RGB space to HSV space, and extract the brightness channel information of the image;

b)将亮度通道信息中的每一个像素作为一个神经元，并将像素的灰度值S_ij作为该神经元的外部输入值，其中S_ij为归一化后的像素值；b) Use each pixel in the brightness channel information as a neuron, and use the gray value S _ij of the pixel as the external input value of the neuron, where S _ij is the normalized pixel value;

c)确定感受野范围为以当前神经元为中心大小为K×L的神经元阵列，其中K、L的取值为奇数，确定维度为K×L的感受野矩阵IG和连接矩阵W，初始化脉冲产生区的动态门限的初始值、衰减时间常数和迭代次数，其中连接矩阵W由感受野内中心神经元与相邻神经元的欧几里德距离平方的倒数确定；IG采用如下公式确定：c) Determine the range of the receptive field as a neuron array with a size of K×L centered on the current neuron, where the values of K and L are odd numbers, determine the receptive field matrix IG and the connection matrix W with dimensions K×L, and initialize The initial value of the dynamic threshold of the pulse generation area, the decay time constant and the number of iterations, where the connection matrix W is determined by the reciprocal of the square of the Euclidean distance between the central neuron and the adjacent neuron in the receptive field; IG is determined by the following formula:

其中， $S_{K \times L} = [\begin{matrix} S_{i - \frac{K - 1}{2} j - \frac{L - 1}{2}} & . & . & . & S_{i - \frac{K - 1}{2} j} & . & . & . & S_{i - \frac{K - 1}{2} j + \frac{L - 1}{2}} \\ . & . \\ . & . \\ . & . \\ S_{ij - \frac{L - 1}{2}} & . & . & . & S_{ij} & . & . & . & S_{ij + \frac{L - 1}{2}} \\ . & . \\ . & . \\ . & . \\ S_{i + \frac{K - 1}{2} j \frac{L - 1}{2}} & . & . & . & S_{i + \frac{K - 1}{2} j} & . & . & . & S_{i + \frac{K - 1}{2} j + \frac{L - 1}{2}} \end{matrix}]$ in, $S_{K \times L} = [\begin{matrix} S_{i - \frac{K - 1}{2} j - \frac{L - 1}{2}} & . & . & . & S_{i - \frac{K - 1}{2} j} & . & . & . & S_{i - \frac{K - 1}{2} j + \frac{L - 1}{2}} \\ . & . \\ . & . \\ . & . \\ S_{ij - \frac{L - 1}{2}} & . & . & . & S_{ij} & . & . & . & S_{ij + \frac{L - 1}{2}} \\ . & . \\ . & . \\ . & . \\ S_{i + \frac{K - 1}{2} j \frac{L - 1}{2}} & . & . & . & S_{i + \frac{K - 1}{2} j} & . & . & . & S_{i + \frac{K - 1}{2} j + \frac{L - 1}{2}} \end{matrix}]$

$\{\begin{matrix} x x = = ((k k - - \frac{K K + + 11}{22})) cos cos θ θ + + ((l l - - \frac{L L + + 11}{22})) sin sin θ θ \\ y the y = = - - ((k k - - \frac{K K + + 11}{22})) sin sin θ θ + + ((l l - - \frac{L L + + 11}{22})) cos cos θ θ \end{matrix}$

其中，IG(k，l)表示IG矩阵第k行第l列的元素，S_K×L表示感受野幅值矩阵，S_K×L(k，l)表示所述幅值矩阵第k行第l列的元素，用感受野内神经元的归一化灰度值表示，K和L表示感受野尺度的大小，σ_x，σ_y分别为高斯包络在x和y方向上的标准差，λ表示感受野函数的波长，

是相位偏移量，θ表示最优方向，γ是方向比，其中下标i，j表示当前神经元的平面位置坐标；Among them, IG(k, l) represents the element of the kth row and lth column of the IG matrix, S _K×L represents the receptive field amplitude matrix, and S _K×L (k, l) represents the kth row and the first column of the magnitude matrix The elements in the l column are represented by the normalized gray value of the neurons in the receptive field, K and L represent the size of the receptive field scale, σ _x , σ _y are the standard deviation of the Gaussian envelope in the x and y directions respectively, λ Indicates the wavelength of the receptive field function,

is the phase offset, θ represents the optimal direction, and γ is the direction ratio, where the subscripts i and j represent the plane position coordinates of the current neuron;

d)根据IG和W得到第n次迭代神经元的反馈输入F_ij[n]和耦合连接输入L_ij[n]；d) Obtain the feedback input F _ij [n] and coupling connection input L _ij [n] of the nth iteration neuron according to IG and W;

e)确定连接强度系数β，由公式U_ij[n]＝F_ij[n](1+βL_ij[n])得到第n次迭代的内部活动项U_ij[n]，当U_ij[n]大于动态门限时，神经元激发产生脉冲输出，并更新动态门限；e) Determine the connection strength coefficient β, and get the internal activity item U _{ij [n] of the nth iteration by the formula U ij} [n]=F _ij [n](1+βL _ij [n]), when _{U ij} _[ n ] is greater than the dynamic threshold, the neuron is excited to generate a pulse output, and the dynamic threshold is updated;

f)重复步骤d，e直到最大迭代次数，神经元的脉冲输出为图像分割的结果。f) Steps d and e are repeated until the maximum number of iterations, and the pulse output of the neuron is the result of image segmentation.

通过利用Gabor函数优化反馈域连接矩阵，使脉冲耦合神经网络中的神经元具有了方向性和尺度性。中心神经元在受其邻域神经元影响时具有了尺度性和方向性，加强了中心神经元与最优方向上的周边神经元的联系，从而保证同一区域的神经元能够同步激发；同时，在感受野模型中加入幅值矩阵，可以更好地体现周边神经元对中心神经元的激发作用。因为PCNN模型分割图像的机理就是通过连接矩阵计算中心神经元的内部活动项，当内部活动项大于动态门限时，该中心神经元激发，否则中心神经元不激发。同一区域内神经元的灰度值变化相对较缓慢，幅值矩阵可以加强不同区域之间的分离度，从而提高图像分割的效果。By using the Gabor function to optimize the connection matrix of the feedback domain, the neurons in the pulse-coupled neural network have directionality and scale. When the central neuron is affected by its neighboring neurons, it has scale and directionality, which strengthens the connection between the central neuron and the peripheral neurons in the optimal direction, thereby ensuring that the neurons in the same area can be excited synchronously; at the same time, Adding the magnitude matrix to the receptive field model can better reflect the excitation effect of peripheral neurons on central neurons. Because the mechanism of the PCNN model to segment images is to calculate the internal activity item of the central neuron through the connection matrix. When the internal activity item is greater than the dynamic threshold, the central neuron fires, otherwise the central neuron does not fire. The gray value of neurons in the same region changes relatively slowly, and the magnitude matrix can strengthen the separation between different regions, thereby improving the effect of image segmentation.

附图说明： Description of drawings:

图1：传统脉冲耦合神经网络神经元结构图Figure 1: Traditional pulse-coupled neural network neuron structure diagram

图2：本发明的脉冲耦合神经网络神经元结构图Fig. 2: The neuron structure diagram of the pulse-coupled neural network of the present invention

图3：本发明的残疾人人脸分割方法总体框图Figure 3: Overall block diagram of the face segmentation method for disabled people in the present invention

图4：本发明算法与其它分割算法的残疾人人脸图像分割比较图Fig. 4: Comparison diagram of the face image segmentation of disabled persons between the algorithm of the present invention and other segmentation algorithms

(a)亮度通道图像，(b)OSTU分割，(c)传统PCNN分割，(d)本发明分割(a) brightness channel image, (b) OSTU segmentation, (c) traditional PCNN segmentation, (d) segmentation of the present invention

具体实施方式： Detailed ways:

本发明利用感受野模型优化连接矩阵的结构，提出了具有方向性和尺度性的脉冲耦合神经网络模型(IG-PCNN)，提高了神经元的图像分割能力，IG-PCNN模型神经元结构图如图2所示。The present invention utilizes the receptive field model to optimize the structure of the connection matrix, and proposes a pulse-coupled neural network model (IG-PCNN) with directionality and scale, which improves the image segmentation ability of neurons. The neuron structure diagram of the IG-PCNN model is as follows Figure 2 shows.

原理如下：The principle is as follows:

参见图2，在以位于坐标(i，j)处的神经元为中心大小为K×L的神经元阵列所构成的感受野内，具有感受野特性的脉冲耦合神经网络(improved Gabor-pulse coupled neural networks，IG-PCNN)模型中单个神经元工作机理可描述为：Referring to Figure 2, in the receptive field formed by the neuron array with the size of K×L centered on the neuron at coordinates (i, j), the improved Gabor-pulse coupled neural network (PCPNN) with receptive field characteristics networks, IG-PCNN) model of a single neuron working mechanism can be described as:

${S S}_{K K \times \times L L} = = [\begin{matrix} {S S}_{i i - - \frac{K K - - 11}{22} j j - - \frac{L L - - 11}{22}} & . . & . . & . . & {S S}_{i i - - \frac{K K - - 11}{22} j j} & . . & . . & . . & {S S}_{i i - - \frac{K K - - 11}{22} j j + + \frac{L L - - 11}{22}} \\ . . & . . \\ . . & . . \\ . . & . . \\ {S S}_{ij ij - - \frac{L L - - 11}{22}} & . . & . . & . . & {S S}_{ij ij} & . . & . . & . . & {S S}_{ij ij + + \frac{L L - - 11}{22}} \\ . . & . . \\ . . & . . \\ . . & . . \\ {S S}_{i i + + \frac{K K - - 11}{22} j j \frac{L L - - 11}{22}} & . . & . . & . . & {S S}_{i i + + \frac{K K - - 11}{22} j j} & . . & . . & . . & {S S}_{i i + + \frac{K K - - 11}{22} j j + + \frac{L L - - 11}{22}} \end{matrix}] - - - - - - ((33))$

$\{\begin{matrix} x x = = ((k k - - \frac{K K + + 11}{22})) cos cos θ θ + + ((l l - - \frac{L L + + 11}{22})) sin sin θ θ \\ y the y = = - - ((k k - - \frac{K K + + 11}{22})) sin sin θ θ + + ((l l - - \frac{L L + + 11}{22})) cos cos θ θ \end{matrix} - - - - - - ((44))$

${L L}_{ij ij} [[n no]] = = \underset{kl kl}{Σ Σ} W W ((k k,, l l)) {Y Y}_{kl kl} [[n no - - 11]] - - - - - - ((55))$

U_ij[n]＝F_in[n](1+βL_ij[n]) (6)U _ij [n]＝F _in [n](1+βL _ij [n]) (6)

${Y Y}_{ij ij} [[n no]] = = \{\begin{matrix} 11 & {U u}_{ij ij} [[n no]] > > {E E.}_{ij ij} [[n no]] \\ 00 & else else \end{matrix} - - - - - - ((77))$

${E E.}_{ij ij} [[n no]] = = {e e}^{- - {α α}_{E E.}} {E E.}_{ij ij} [[n no - - 11]] + + {V V}_{E E.} {Y Y}_{ij ij} [[n no]] - - - - - - ((88))$

反馈输入域中的连接矩阵M用感受野矩阵IG代替，式(1)～(8)中F_ij是第(i，j)个神经元的反馈输入，F_ij[n]表示第n次迭代时的反馈输入，IG(k，l)表示感受野矩阵IG第k行第l列的元素，S_ij为神经元接收外界激励，用坐标为(i，j)的神经元处的像素的灰度值表示，在感受野模型中S_K×L表示感受野幅值矩阵，用区域内神经元的归一化灰度值表示，K和L表示感受野尺度的大小，L_ij[n]表示第n次迭代时的连接输入；σ_x，σ_y分别为高斯包络在x和y方向上的标准差；λ表示感受野函数的波长，

是相位偏移量，θ表示最优方向，γ是方向比；L_ij是连接输入，Y_kl[n-1]是第n-1次迭代时相邻神经元的输出，W为连接矩阵；β为突触之间的连接强度系数；α_E为衰减时间常数；U_ij[n]是内部活动项，其大小决定神经元是否输出脉冲，当神经元内部活动项U_ij大于动态门限E_ij时，神经元激发产生脉冲输出Y_kl，否则神经元不产生脉冲输出。The connection matrix M in the feedback input domain is replaced by the receptive field matrix IG. In formulas (1)-(8), F _ij is the feedback input of the (i, j)th neuron, and F _ij [n] represents the nth iteration Feedback input at time, IG(k, l) represents the element of the kth row and lth column of the receptive field matrix IG, S _ij is the neuron receiving external excitation, and the gray value of the pixel at the neuron with the coordinates (i, j) In the receptive field model, S _K×L represents the receptive field amplitude matrix, expressed by the normalized gray value of neurons in the region, K and L represent the size of the receptive field scale, and _Lij [n] represents The connection input at the nth iteration; σ _x , σ _y are the standard deviations of the Gaussian envelope in the x and y directions respectively; λ represents the wavelength of the receptive field function,

is the phase offset, θ represents the optimal direction, γ is the direction ratio; L _ij is the connection input, Y _kl [n-1] is the output of the adjacent neuron at the n-1th iteration, W is the connection matrix; β is the connection strength coefficient between synapses; α _E is the decay time constant; U _ij [n] is the internal activity item, and its size determines whether the neuron outputs pulses. When the internal activity item U _ij of the neuron is greater than the dynamic threshold E _ij When , the neuron excites to generate a pulse output Y _kl , otherwise the neuron does not generate a pulse output.

残疾人人脸图像分割算法实现步骤：Implementation steps of face image segmentation algorithm for disabled persons:

Step1：首先将采集到的残疾人人脸图像由RGB空间转换到HSV空间，提取图像的亮度通道信息，并将其作为模拟视觉细胞感受野特性的脉冲耦合神经网络模型的外部输入。Step1: First, convert the collected face images of the disabled from RGB space to HSV space, extract the brightness channel information of the image, and use it as the external input of the pulse-coupled neural network model that simulates the receptive field characteristics of visual cells.

Step2：将亮度通道信息中的每一个像素作为一个神经元，并将像素的灰度值作为该神经元的外部输入值，其中S_ij为归一化后的像素值。Step2: Take each pixel in the brightness channel information as a neuron, and use the gray value of the pixel as the external input value of the neuron, where S _ij is the normalized pixel value.

Step3：根据图像的特点，确定感受野模型中的参数，最优方向θ＝0°和最佳尺度K＝L＝3，λ＝20，

γ＝0.5。Step3: According to the characteristics of the image, determine the parameters in the receptive field model, the optimal direction θ=0° and the optimal scale K=L=3, λ=20,

γ = 0.5.

Step4：感受野模型优化脉冲耦合神经网络反馈域连接矩阵，神经元感受野模型IG由公式(2)计算得到，利用公式(3)获取幅值矩阵：Step4: The receptive field model optimizes the pulse-coupled neural network feedback domain connection matrix. The neuron receptive field model IG is calculated by the formula (2), and the amplitude matrix is obtained by using the formula (3):

${S S}_{33 \times \times 33} = = [\begin{matrix} {S S}_{i i - - 11 j j - - 11} & {S S}_{i i - - 11 j j} & {S S}_{i i - - 11 j j + + 11} \\ {S S}_{ij ij - - 11} & {S S}_{ij ij} & {S S}_{ij ij + + 11} \\ {S S}_{i i + + 11 j j - - 11} & {S S}_{i i + + 11 j j 11} & {S S}_{i i + + 11 j j + + 11} \end{matrix}]$

反馈域连接矩阵为：The feedback domain connectivity matrix is:

$IG IG ((k k,, l l)) = = {S S}_{K K \times \times L L} ((k k,, l l)) cos cos ((\frac{22 π π}{2020} x x)) \cdot &Center Dot; exp exp [[- - \frac{11}{22} ((\frac{{x x}^{22}}{{σ σ}_{x x}^{22}} + + \frac{{0.5 0.5}^{22} {y the y}^{22}}{{σ σ}_{y the y}^{22}}))]]$

根据公式(1)计算此时的F_ij[n]。F _ij [n] at this time is calculated according to formula (1).

Step5：初始化PCNN模型中的其他参数。初始阈值Eij[0]为图像的最佳阈值；衰减时间常数α_E＝0.185；V_E＝2；n＝4。连接域权值矩阵W由相邻神经元的欧几里得距离平方的倒数确定，Step5: Initialize other parameters in the PCNN model. The initial threshold Eij[0] is the optimal threshold of the image; decay time constant α _E =0.185; V _E =2; n=4. The connection domain weight matrix W is determined by the inverse of the square of the Euclidean distance between adjacent neurons,

$W W = = [\begin{matrix} 11 / / \sqrt{22} & 11 & 11 / / \sqrt{22} \\ 11 & 00 & 11 \\ 11 / / \sqrt{22} & 11 & 11 / / \sqrt{22} \end{matrix}]$

并计算此时的L_ij[n]。And calculate L _ij [n] at this time.

Step6：链接强度系数β的计算。计算每一个神经元在其3×3邻域内的链接强度系数，

x_kl表示以(i，j)为中心的邻域神经元的灰度值，

表示以(i，j)为中心的K×L区域神经元灰度值的均值，m表示K×L区域内神经元个数，并计算此时的U_ij[1]，U_ij[1]＝F_ij[1](1+βL_ij[1])。Step6: Calculation of link strength coefficient β. Calculate the link strength coefficient of each neuron in its 3×3 neighborhood,

x _kl represents the gray value of the neighborhood neuron centered on (i, j),

Indicates the mean value of the gray value of neurons in the K×L area centered on (i, j), m indicates the number of neurons in the K×L area, and calculates U _ij [1], U _ij [1] at this time =F _ij [1](1+βL _ij [1]).

Step7：比较U_ij[1]与E_ij[0]的大小，若U_ij[1]＞E_ij[0]，则Y_ij[1]＝1即神经元I_ij激发，标记该神经元为激发，且该神经元始终保持激发状态。Step7: Compare the size of U _ij [1] and E _ij [0], if U _ij [1] > E _ij [0], then Y _ij [1] = 1, that is, the neuron I _ij is excited, and the neuron is marked as fires, and the neuron remains fired all the time.

Step8：迭代次数n＝n+1，计算新的

并重复Step1～Step7，直到指定的迭代次数n停止。本文取n＝4，此时得到的Y_ij[n]为最终的残疾人人脸分割图像。Step8: The number of iterations n=n+1, calculate the new

And repeat Step1~Step7 until the specified number of iterations n stops. In this paper, n=4, and the Y _ij [n] obtained at this time is the final segmented image of the handicapped face.

多种分割方法比较：Comparison of multiple segmentation methods:

如图4所示为本发明的残疾人人脸图像分割比较，即采用OSTU、传统脉冲耦合神经网络、本发明分割方法三种方法对残疾人人脸图像分割结果的比较。选取两种具有代表性的经典分割方法与本专利分割算法的比较结果图，从图中可以看出，OTSU阈值法(即最大类间方差法)的核心是按图像的灰度特性，将图像分成背景和目标两部分，但当部分目标错分为背景或部分背景错分为目标时都会导致两部分差别变小，所以容易造成人脸区域得不到很好地分割或是将图像的部分背景分割出来的情况。传统PCNN分割方法由于同一区域内神经元之间的相互作用，分割结果很好地去除了背景、较好地保留了人脸区域的信息，但相邻区域之间的分离度不理想造成了分割目标之间相互粘连，为后续的图像特征提取造成困难。本发明方法利用初级视皮层感受野模型优化连接矩阵的结构，实现了对脉冲耦合神经网络模型方向和尺度的选择，加强了同一区域神经元之间的相互作用，使分割结果更接近人类视觉的分割结果，解决了分割中出现的过分割与欠分割问题，具有广泛的应用价值。As shown in Figure 4, it is the comparison of the face image segmentation of the disabled in the present invention, that is, the comparison of the segmentation results of the face images of the disabled by using OSTU, the traditional pulse-coupled neural network, and the segmentation method of the present invention. Select two representative classic segmentation methods and the comparison results of the patented segmentation algorithm. From the figure, it can be seen that the core of the OTSU threshold method (ie, the maximum inter-class variance method) is to divide the image according to the grayscale characteristics of the image Divided into two parts, the background and the target, but when part of the target is wrongly divided into the background or part of the background is wrongly divided into the target, the difference between the two parts will become smaller, so it is easy to cause the face area to be not well segmented or part of the image The situation where the background is segmented out. Due to the interaction between neurons in the same area, the traditional PCNN segmentation method removes the background well and retains the information of the face area well, but the separation between adjacent areas is not ideal, resulting in segmentation problems. The objects stick to each other, which makes it difficult for the subsequent image feature extraction. The method of the invention utilizes the primary visual cortex receptive field model to optimize the structure of the connection matrix, realizes the selection of the direction and scale of the pulse-coupled neural network model, strengthens the interaction between neurons in the same area, and makes the segmentation result closer to that of human vision The segmentation result solves the problem of over-segmentation and under-segmentation in segmentation, and has wide application value.

有益效果Beneficial effect

1)该方法可以很好地解决一般图像分割方法对光照敏感、图像细节信息不突出、以及分割中常出现的过分割与欠分割的问题。1) This method can well solve the problems that general image segmentation methods are sensitive to illumination, image details are not prominent, and over-segmentation and under-segmentation often occur in segmentation.

2)具有感受野特性的脉冲耦合神经网络模型，使图像分割模拟了人类视觉细胞图像分割的功能。2) The pulse-coupled neural network model with receptive field characteristics enables image segmentation to simulate the function of human visual cell image segmentation.

3)通过对大量残疾人人脸图像的实验仿真，验证了该分割方法对残疾人人脸图像分割的有效性。3) Through the experimental simulation of a large number of disabled face images, the effectiveness of the segmentation method for disabled face images is verified.

4)利用分割评价准则对分割结果进行评价，进一步验证了脸部各区域之间分离度好、图像细节保持好及边缘分割准确率高的优点。4) The segmentation results are evaluated by the segmentation evaluation criteria, which further verifies the advantages of good separation between face regions, good image detail preservation and high edge segmentation accuracy.

5)解决了分割中常出现的欠分割与过分割问题，为残疾人身份认证系统的应用打下良好的基础。5) It solves the under-segmentation and over-segmentation problems that often occur in segmentation, and lays a good foundation for the application of the identification system for the disabled.

6)本项发明是残疾人脸部信息获取及处理的关键技术，解决了残疾人人脸分割的问题，同时也适合正常人人脸的分割。6) This invention is a key technology for obtaining and processing face information of disabled people, which solves the problem of face segmentation for disabled people, and is also suitable for face segmentation of normal people.

Claims

1. the Pulse Coupled Neural Network facial image dividing method of an analog vision cell receptive field characteristic, neuron in described Pulse Coupled Neural Network comprises acceptance domain, modulation domain and pulse producer three parts, described acceptance domain comprises feedback acceptance domain and is connected acceptance domain, feedback acceptance domain reception gradation of image value S _ijoutput pulse Y with adjacent neurons in receptive field _klas input, through receptive field matrix IG conversion output F _ijas neuronic feed back input, connect the pulse output Y of adjacent neurons in acceptance domain reception receptive field _klas input, through connection matrix W conversion output L _ijas the neuronic input that is of coupled connections, said method comprising the steps of:

A), the disabled person's facial image collecting is transformed into HSV space by rgb space, extracts the luminance channel information of image;

B), using each pixel in luminance channel information as a neuron, and by the gray-scale value S of pixel _ijas this neuronic outside input value, wherein S _ijfor the pixel value after normalization;

C), determine that receptive field scope is that size is the neuron array of K × L centered by current neuron, wherein K, the value of L is odd number, determine that dimension is receptive field matrix IG and the connection matrix W of K × L, initialization pulse produces initial value, damping time constant and the iterations of dynamic threshold in district, and wherein connection matrix W is definite by the inverse of the Euclidean distance of center neuron and adjacent neurons in receptive field square; Wherein IG adopts following formula to determine:

Wherein,

S_{K \times L} = [\begin{matrix} S_{i - \frac{K - 1}{2} j - \frac{L - 1}{2}} & . . . & S_{i - \frac{k - 1}{2} j} & . . . & S_{i - \frac{K - 1}{2} j + \frac{L - 1}{2}} \\ . & . & \cdot & \cdot \\ . & . & \cdot & \cdot \\ . & . & \cdot & \cdot \\ S_{ij - \frac{L - 1}{2}} & . . . & S_{ij} & . . . & S_{ij + \frac{L - 1}{2}} \\ . & . & \cdot & . \\ . & . & \cdot & . \\ . & . & \cdot & . \\ S_{i + \frac{K - 1}{2} j - \frac{L - 1}{2}} & . . . & S_{i + \frac{K - 1}{2} j} & . . . & S_{i + \frac{K - 1}{2} j + \frac{L - 1}{2}} \end{matrix}]

\{\begin{matrix} x = (k - \frac{K + 1}{2}) \cos θ + (l - \frac{L + 1}{2}) \sin θ \\ y = - (k - \frac{K + 1}{2}) \sin θ + (l - \frac{L + 1}{2}) \cos θ \end{matrix}

Wherein, IG (k, l) represents the element of the capable l row of IG matrix k, S _{k × L}represent receptive field amplitude matrix, S _{k × L}(k, l) represents the element of the capable l row of described amplitude matrix k, and with neuronic Normalized Grey Level value representation in receptive field, K and L represent the size of receptive field yardstick, σ _x, σ _ybe respectively Gaussian envelope standard deviation in the x and y direction, λ represents the wavelength of receptive field function,

be phase pushing figure, θ represents optimal direction, and γ is direction ratio, wherein subscript i, and j represents current neuronic planimetric position coordinate;

D), obtain the neuronic feed back input F of iteration the n time according to IG and W _ij[n] and the input L that is of coupled connections _ij[n];

E), calculate strength of joint factor beta, by formula U _ij[n]=F _ij[n] (1+ β L _ij[n]) obtain the internal activity item U of the n time iteration _ij[n], works as U _ijwhen [n] is greater than dynamic threshold, neuron excites and produces pulse output, and Regeneration dynamics thresholding;

F), repeating step d, e is until maximum iteration time, and neuronic pulse is output as the result that image is cut apart.

2. method according to claim 1, is characterized in that: described optimal direction θ=0 °, and best scale K=L=3.