CN102034288A

CN102034288A - Multiple biological characteristic identification-based intelligent door control system

Info

Publication number: CN102034288A
Application number: CN 201010581472
Authority: CN
Inventors: 周治平; 李雨凇
Original assignee: Jiangnan University
Current assignee: JIANGSU CINSEC INFORMATION TECHNOLOGY Co Ltd
Priority date: 2010-12-09
Filing date: 2010-12-09
Publication date: 2011-04-27
Anticipated expiration: 2030-12-09
Also published as: CN102034288B

Abstract

The invention relates to an intelligent access control system based on multi-biological feature recognition, which includes a sound acquisition device and an image acquisition device; an access control controller respectively receives the sound signal and the face image signal collected by the image acquisition device; the output terminals of the access control controller respectively Connected with the data processor and the electric door lock; the access control controller transmits the received sound signal and face image signal to the data processor; the data processor performs feature extraction on the received sound signal and face image signal respectively , identifying the collected sound signal and face image signal, and performing normalization and classification fusion processing on the corresponding signal after the feature extraction; the data processor compares the corresponding signal after the classification fusion processing with the preset sample library; When the signal obtained after classification and fusion processing matches the corresponding signal in the sample library, the access controller opens the electronically controlled door lock. The invention improves the reliability of identification, has high accuracy, is convenient for identification, and is safe and reliable.

Description

Intelligent access control system based on multi-biometric identification

技术领域technical field

本发明涉及一种门禁系统，尤其是一种基于多生物特征识别的智能门禁系统，属于门禁系统的技术领域。The invention relates to an access control system, in particular to an intelligent access control system based on multi-biological feature identification, which belongs to the technical field of access control systems.

背景技术Background technique

目前，传统的门禁系统的身份鉴别手段包括口令、密码、证件等等。由于与被鉴别人的可分离性，易造成伪造、盗用、破译等现象。而人体的生物属性是人的本体独有的，而且人的某些生物属性如指纹、虹膜、声音等是唯一的，将其应用在门禁系统可以杜绝伪造、盗用等现象的产生。生物特征识别技术主要分为两大类：一类是生理特征识别，分别是利用指纹，掌型，虹膜，视网膜及人脸等特征进行识别；另一类是行为特征识别，包括签名和语音识别。采用单一模态的生物特征认证技术容易受到噪声的干扰和使用环境的限制，并且容易伪造仿冒，很难保证门禁系统识别的准确性。At present, the identification methods of the traditional access control system include passwords, passwords, certificates and so on. Due to the separability from the identified person, it is easy to cause forgery, misappropriation, deciphering and other phenomena. The biological attributes of the human body are unique to the human body, and some biological attributes such as fingerprints, irises, and voices are unique. Applying them to the access control system can prevent forgery and embezzlement. Biometric recognition technology is mainly divided into two categories: one is physiological feature recognition, which uses fingerprints, palms, irises, retinas and faces for recognition; the other is behavioral feature recognition, including signature and voice recognition. . The biometric authentication technology using a single mode is susceptible to noise interference and the limitation of the use environment, and it is easy to forge and counterfeit, so it is difficult to guarantee the accuracy of the identification of the access control system.

发明内容Contents of the invention

本发明的目的是克服现有技术中存在的不足，提供一种基于多生物特征识别的智能门禁系统，其提高了识别的可靠性，准确率高，识别方便，安全可靠。The purpose of the present invention is to overcome the deficiencies in the prior art, and provide an intelligent access control system based on multi-biometric feature identification, which improves the reliability of identification, has high accuracy, convenient identification, and is safe and reliable.

按照本发明提供的技术方案，所述基于多生物特征识别的智能门禁系统，包括声音采集设备及图像采集设备；所述声音采集设备及图像采集设备均与门禁控制器的输入相连，所述门禁控制器分别接收声音采集设备采集的声音信号及图像采集设备采集的人脸图像信号；门禁控制器的输出端分别与数据处理器及电控门锁相连；门禁控制器将接收的声音信号及人脸图像信号传输到数据处理器内；所述数据处理器分别对接收的声音信号及人脸图像信号进行特征提取，对采集的声音信号与人脸图像信号进行识别，并将特征提取后相应信号进行归一化与分类融合处理；数据处理器将分类融合处理后相应的信号与预设的样本库相比较；当分类融合处理后得到的信号与样本库内相应信号匹配时，门禁控制器打开电控门锁。According to the technical solution provided by the present invention, the intelligent access control system based on multi-biometric feature recognition includes sound acquisition equipment and image acquisition equipment; the sound acquisition equipment and image acquisition equipment are connected to the input of the access control controller, and the access control The controller receives the sound signal collected by the sound collection device and the face image signal collected by the image collection device; the output terminals of the access control controller are respectively connected with the data processor and the electric door lock; The face image signal is transmitted to the data processor; the data processor performs feature extraction on the received sound signal and the face image signal respectively, recognizes the collected sound signal and the face image signal, and extracts the corresponding signal Perform normalization and classification fusion processing; the data processor compares the corresponding signal after classification fusion processing with the preset sample library; when the signal obtained after classification fusion processing matches the corresponding signal in the sample library, the access controller opens Electric door lock.

所述数据处理器包括DSP。所述数据处理器利用主成分分析法与线性判别分析法对人脸图像信号进行特征提取及人脸识别。The data processor includes a DSP. The data processor uses the principal component analysis method and the linear discriminant analysis method to perform feature extraction and face recognition on the face image signal.

所述数据处理器利用梅尔倒谱系数与混合高斯模型方法对声音信号进行特征提取及人脸识别。所述数据处理器利用z-score函数对识别后的声音信号与人脸图像信号进行归一化。The data processor uses Mel cepstrum coefficient and mixed Gaussian model method to perform feature extraction and face recognition on the sound signal. The data processor uses a z-score function to normalize the recognized sound signal and face image signal.

所述数据处理器利用支持向量机对归一化后的声音信号与人脸图像信号进行分类融合处理。所述门禁控制器的输出端还与声光报警器相连；当分类融合处理后得到的信号与数据处理器内样本库不匹配时，门禁控制器通过声光报警器输出声光报警信号。The data processor uses a support vector machine to classify and fuse the normalized sound signal and face image signal. The output terminal of the access control controller is also connected with the sound and light alarm; when the signal obtained after classification and fusion processing does not match the sample library in the data processor, the access control controller outputs the sound and light alarm signal through the sound and light alarm.

本发明的优点：门禁控制器同时接收声音采集设备与图像采集设备的声音信号、人脸图像信号；数据处理器对声音信号利用梅尔倒谱系数和混合高斯模型方法进行特征提取和声音识别，数据处理器对人脸图像信号利用主成分分析法和线性判别分析法进行特征提取和声音设备，数据处理器对提取和识别后的声音信号及人脸图像信号进行Z-score归一化和支持向量机分类融合，降低了识别的等错率，提高了门禁系统的安全可靠性，安全可靠。Advantages of the present invention: the access control controller simultaneously receives the sound signal and the face image signal of the sound collection device and the image collection device; the data processor uses the Mel cepstrum coefficient and the mixed Gaussian model method to perform feature extraction and sound recognition on the sound signal, The data processor uses principal component analysis and linear discriminant analysis to perform feature extraction and sound equipment on the face image signal, and the data processor performs Z-score normalization and support on the extracted and recognized sound signal and face image signal Vector machine classification fusion reduces the error rate of recognition and improves the safety and reliability of the access control system, which is safe and reliable.

附图说明Description of drawings

图1为本发明的结构框图。Fig. 1 is a structural block diagram of the present invention.

图2为本发明的工作流程图。Fig. 2 is a working flow diagram of the present invention.

具体实施方式Detailed ways

下面结合具体附图和实施例对本发明作进一步说明。The present invention will be further described below in conjunction with specific drawings and embodiments.

如图1所示：本发明包括图像采集设备、声音采集设备，电控门锁、门禁控制器、数据处理器及声光报警器。As shown in Figure 1: the present invention includes image acquisition equipment, sound acquisition equipment, electric control door lock, access control controller, data processor and sound and light alarm.

多生物特征融合是指利用生物特征不同的特性，进行某种层面的融合，其目的是克服或者规避单一特征的局限性；本发明通过对人脸图像和声音信号多位门禁的两个生物特征。通过把不同生物特征信息结合在一起进行融合认证降低了不利因素的影响，为解决单一模态的生物特征识别的不足带来了有效的解决方案。人脸图像和声音信息具有相对的唯一性和稳定性，并且采集方便，具有非接触性和非侵犯性。适合于门禁系统作为鉴别特征来使用。Multi-biological feature fusion refers to the use of different characteristics of biological features to carry out fusion at a certain level, the purpose of which is to overcome or avoid the limitations of a single feature; . By combining different biometric information for fusion authentication, the influence of unfavorable factors is reduced, and an effective solution is brought to solve the shortage of single-modal biometric identification. Face image and sound information are relatively unique and stable, and are convenient to collect, non-contact and non-invasive. It is suitable for use as an identification feature in access control systems.

如图1所示：所述图像采集设备与声音采集设备的输出端与门禁控制器的输入端相连，所述图像彩采集设备采用摄像头，作为特征采集工具，采集来访者的人脸图像；声音采集设备采用麦克风，作为特征采集工具，采集来访者的声音信号。门禁控制器的输出端与数据处理器相连，所述数据处理器包括DSP(数字信号处理器)，所述数据处理器能够进行人脸识别和声音识别，所述数据处理器内预先存储有相关的人脸信息和声音信息。门禁控制器接收图像采集设备采集的人脸图像信号及声音采集设备采集的声音信号，并将上述信号传输到数据处理器内。所述数据处理器对接收的声音信号利用梅尔谱倒频系数(MFCC)及混合高斯模型(GMM)进行特征提取和声音识别；数据处理器对接收的人脸图像信号利用主成分分析法(PCA)及线性判别分析法(LDA)进行特征提取和人脸识别。为了降低分类比较的等错率，数据处理器通过Z-score函数对对识别后的声音信号和人脸图像信号进行归一化处理，消除利用不同生物特征识别的结果差异，减小由于类别不同而引起的误差；数据处理器在进行归一化处理后，再利用支持向量机(SVM)方法对上述归一化结果进行分类，达到决策层融合处理。数据处理器将上述融合处理后的声音信号及人脸图像信号和数据处理器内预存的样本库间进行比较，当采集的声音信号及人脸图像信号与样本库内相应的信号相匹配时，数据处理器向门禁控制器发出开门指令，门禁控制器打开电控门锁；当采集的声音信号和人脸图像信号与样本库不匹配时，门禁控制器通过声光报警器发出声光报警信号，确保门禁系统的安全。As shown in Figure 1: the output end of described image acquisition equipment and sound acquisition equipment is connected with the input end of access controller, and described image color acquisition equipment adopts camera, as feature acquisition tool, gathers the face image of visitor; The collection device uses a microphone as a feature collection tool to collect the visitor's voice signal. The output end of access control controller is connected with data processor, and described data processor comprises DSP (digital signal processor), and described data processor can carry out face recognition and voice recognition, and relevant facial and voice information. The access control controller receives the face image signal collected by the image collection device and the sound signal collected by the sound collection device, and transmits the above signals to the data processor. Described data processor utilizes mel spectrum cepstral coefficient (MFCC) and mixed Gaussian model (GMM) to carry out feature extraction and sound recognition to the sound signal that receives; Data processor utilizes principal component analysis method ( PCA) and linear discriminant analysis (LDA) for feature extraction and face recognition. In order to reduce the equal error rate of classification comparison, the data processor normalizes the recognized sound signal and face image signal through the Z-score function, eliminating the difference in the results of recognition using different biometric features, and reducing the difference caused by different categories. The error caused; after the data processor performs the normalization process, the support vector machine (SVM) method is used to classify the above-mentioned normalization results to achieve the decision-making level fusion process. The data processor compares the above-mentioned fusion-processed sound signal and face image signal with the sample library prestored in the data processor, and when the collected sound signal and face image signal match the corresponding signal in the sample library, The data processor sends an instruction to open the door to the access controller, and the access controller opens the electronically controlled door lock; when the collected sound signal and face image signal do not match the sample library, the access controller sends an audible and visual alarm signal through the audible and visual alarm , to ensure the security of the access control system.

数据处理器对人脸图像信号利用主成分分析法(PCA)及线性判别分析法(LDA)进行特征提取和人脸识别，其中，主成分分析(PCA)是一种统计方法，它借助于一个正交变换，将分量相关的原随机向量转化成分量不相关的新随机向量，在代数上表现为将原随机向量的协方差阵变换成对角形阵，在几何上表现为将原坐标系变换成新的正交坐标系，样本点散布最开。线性判别分析(LDA)在模式识别领域有着重大的影响，线性判别分析方法可应用于图像降维，它是一种基于样本的类别进行降维的方法。它的投影矩阵通过最大化类间分布，同时最小化类内散布获得。The data processor uses Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) to perform feature extraction and face recognition on the face image signal. Among them, Principal Component Analysis (PCA) is a statistical method. It uses a Orthogonal transformation transforms the original random vector with related components into a new random vector with uncorrelated components. Algebraically, it is expressed as transforming the covariance matrix of the original random vector into a diagonal matrix. Geometrically, it is expressed as transforming the original coordinate system Form a new orthogonal coordinate system, and the sample points are scattered the most. Linear discriminant analysis (LDA) has a great influence in the field of pattern recognition. The linear discriminant analysis method can be applied to image dimensionality reduction. It is a method for dimensionality reduction based on the category of samples. Its projection matrix is obtained by maximizing the between-class distribution while minimizing the within-class scatter.

线性判别分析人脸识别方法描述如下：The linear discriminant analysis face recognition method is described as follows:

假设原始图像库中共有N个图像，假设每个图像共有d个像素点，则原始人脸图像向量可以表示为X₁，X₂，Λ X_N(向量维数设为d)，其中N₁个人脸图像属于1类，N₂个人脸图像属于2类，N_C个人脸图像属于C类(其中N₁+N₂+…+N_C＝N)。则各类人脸图像的均值为：Assuming that there are N images in the original image library, assuming that each image has d pixels in total, the original face image vector can be expressed as X ₁ , X ₂ , Λ X _N (vector dimension is set to d), where N ₁ Face images belong to category 1, N ₂ facial images belong to category 2, and N _C facial images belong to category C (where N ₁ +N ₂ +...+ _NC =N). Then the average value of all kinds of face images is:

${μ μ}_{i i} = = \frac{11}{N N} {Σ Σ}_{i i = = 11}^{N N} {X x}_{i i},, ((i i = = 1,2 1,2,, L L,, C C)) - - - - - - ((11))$

总的人脸图像的均值为：The mean of the total face image is:

$μ μ = = \frac{11}{N N} {Σ Σ}_{i i = = 11}^{N N} {X x}_{i i} - - - - - - ((22))$

样本类间离散度矩阵S_B和类内离散度矩阵S_w定义为：The sample between-class dispersion matrix S _B and the intra-class dispersion matrix S _w are defined as:

${S S}_{B B} = = {Σ Σ}_{i i = = 11}^{C C} {N N}_{i i} (({μ μ}_{i i} - - μ μ)) {(({μ μ}_{i i} - - μ μ))}^{T T} - - - - - - ((33))$

${S S}_{W W} = = {Σ Σ}_{i i = = 11}^{C C} \underset{{X x}_{k k} &Element; &Element; {X x}_{i i}}{Σ Σ} (({X x}_{k k} - - {μ μ}_{i i})) {(({X x}_{k k} - - {μ μ}_{i i}))}^{T T} - - - - - - ((44))$

如果S_w(类内离散度矩阵)是非奇异的，则要获得类间离散度与类内离散度的比值最大的投影方向的W_opt满足下式：If S _w (intra-class scatter matrix) is non-singular, then the W _opt of the projection direction to obtain the largest ratio of inter-class scatter to intra-class scatter satisfies the following formula:

${W W}_{opt opt} = = arg arg max max \frac{| | {W W}^{T T} {S S}_{B B} W W | |}{| | {W W}^{T T} {S S}_{W W} W W | |} = = [[{w w}_{11},, {w w}_{22},, K K,, {w w}_{m m}]] - - - - - - ((55))$

其中{w_i|i＝1，2，L，m}是满足下式的S_B和S_w对应的m个最大特征值{λ_i|i＝1，2，L，m}所对应的特征向量：S_BW_i＝λ_iS_WW_i(i＝1，2，L，m)。注意到该矩阵最多只有C-1个非零特征值，C是类别数目。Where {w _i |i=1, 2, L, m} is the feature corresponding to the m largest eigenvalues {λ _i |i=1, 2, L, m} corresponding to S _B and S _w satisfying the following formula Vector: S _B W _i =λ _i S _W W _i (i=1, 2, L, m). Note that the matrix has at most C-1 non-zero eigenvalues, where C is the number of categories.

利用线性投影算子，每一幅人脸图像可以映射得到一个低维的特征向量。该向量的元素通过图像向量与投影算子的每一个列向量分别做内积运算得到。Using the linear projection operator, each face image can be mapped to obtain a low-dimensional feature vector. The elements of this vector are obtained by performing the inner product operation between the image vector and each column vector of the projection operator.

无论PCA方法还是LDA方法都需要大量的实验样本，而实际生活中并不可能完全满足样本数量的要求，为了解决小样本可能存在的问题，数据处理器采用PCA+LDA的方法来进行人脸识别。Both the PCA method and the LDA method require a large number of experimental samples, but in real life it is impossible to fully meet the requirements of the number of samples. In order to solve the possible problems of small samples, the data processor uses the PCA+LDA method for face recognition. .

设对pⁱ _j(j＝1，2，...，S；i＝1，2，...，K)为PCA得到的第i类人脸第j个人脸向量的特征投影，S是每类的样本数，K是训练样本总数。Let p ⁱ _j (j=1, 2, ..., S; i = 1, 2, ..., K) be the feature projection of the i-th type of face vector obtained by PCA, S is The number of samples for each class, K is the total number of training samples.

首先计算各类样本的均值μ_i和总样本均值μ，从各样本的图像减去对应的类均值，即各类训练样本中心化，然后，从各类均值中减去总样本均值得到μ_i，把所有中心化的训练样本图像组成数据矩阵，并通过PCA方法为这个数据矩阵寻找正交基。设求出的正交基为U，将所有中心化的图像投影到正交基上。First calculate the mean value μ _i of various samples and the mean value μ of the total sample, subtract the corresponding class mean value from the image of each sample, that is, the centering of various training samples, and then subtract the mean value of the total sample from the mean value of each type to obtain μ _i , compose all the centralized training sample images into a data matrix, and use the PCA method to find an orthogonal basis for this data matrix. Let the obtained orthogonal base be U, and project all the centered images onto the orthogonal base.

其中，

标识样本特征值，把所有的中心化的均值投影到正交基上，完成了PCA的过程：in,

Identify the sample eigenvalues, project all the centered mean values onto the orthogonal basis, and complete the PCA process:

最后将求得的参数代入LDA的以下公式中：Finally, the obtained parameters are substituted into the following formula of LDA:

求解类间集散度矩阵S_B和类内离散度矩阵S_W。其中：Solve the inter-class scatter matrix S _B and the intra-class scatter matrix S _W . in:

${S S}_{W W} = = {Σ Σ}_{I I = = 11}^{C C} {S S}_{I I} - - - - - - ((99))$

计算广义特征值Λ和对应的特征向量V：Compute the generalized eigenvalue Λ and the corresponding eigenvector V:

S_BV＝λS_WV (11)S _B V = λS _W V (11)

根据对应特征值由大到小的顺序排列特征向量，仅保留前K-1个特征向量，这就是Fisher基向量。把旋转过的原始图像投影到Fisher基向量上，就是说先把原始图像投影到正交基U上，再把得到的投影图像继续投影到Fisher基向量W上。然后进行分类识别工作，计算出人脸识别的识别率。Arrange the eigenvectors according to the order of the corresponding eigenvalues from large to small, and only keep the first K-1 eigenvectors, which are the Fisher basis vectors. Project the rotated original image onto the Fisher basis vector, that is, first project the original image onto the orthogonal basis U, and then continue to project the obtained projected image onto the Fisher basis vector W. Then carry out classification and recognition work, and calculate the recognition rate of face recognition.

梅尔倒谱系数(Mel-Frequency Cepstrum Coefficient，MFCC)是对人耳滤波功能的一个模拟，在低频部分呈线性的增长，在高频部分呈指数性的增长，是最有效的一种说话人识别的特征。目前，对于文本无关的声纹识别主要采用MFCC特征。梅尔倒谱系数特征参数提取过程为预处理分帧、FFT(快速傅里叶变换)、梅尔滤波器、取对数及DCT(离散余弦变换)变换过程。Mel-Frequency Cepstrum Coefficient (MFCC) is a simulation of the filtering function of the human ear. It increases linearly in the low frequency part and increases exponentially in the high frequency part. It is the most effective speaker Recognized features. At present, MFCC features are mainly used for text-independent voiceprint recognition. The process of extracting the feature parameters of Mel cepstral coefficients is the process of preprocessing subdivision, FFT (Fast Fourier Transform), Mel filter, taking logarithm and DCT (Discrete Cosine Transform).

统计模型中的概率密度函数是对说话人特征在特征空间这种分布的完整描述，一个说话人的特征分布概率密度函数就可以成为这一说话人的模板。The probability density function in the statistical model is a complete description of the distribution of the speaker's features in the feature space, and the probability density function of a speaker's feature distribution can be a template for this speaker.

GMM(混合高斯模型)是概率统计模型，通过对目标说话人特征分布的统计描述来区分说话人，其统计参数能有效地表示话者的特征信息，因此可以利用GMM来做特征空间的映射。它具有与文本无关、处理速度快、识别效果好等优点，已成功地运用于与文本无关的说话人识别和确认中。GMM (Mixed Gaussian Model) is a probabilistic statistical model that distinguishes speakers through the statistical description of the target speaker's feature distribution. Its statistical parameters can effectively represent the speaker's feature information, so GMM can be used to map the feature space. It has the advantages of being text-independent, fast processing speed, and good recognition effect, and has been successfully used in text-independent speaker recognition and confirmation.

GMM是M个成员高斯概率密度的加权和，表示为：GMM is the weighted sum of M member Gaussian probability densities, expressed as:

$p p ((x x | | λ λ)) = = {Σ Σ}_{i i = = 11}^{M m} {a a}_{i i} {b b}_{i i} ((x x)) - - - - - - ((1212))$

其中，x是D维随机向量；b_i(x)(i＝1，2，...，M)是每个成员的高斯概率密度函数；a_i(i＝1，2，...，M)是混合权值。Among them, x is a D-dimensional random vector; b _i (x) (i=1, 2, ..., M) is the Gaussian probability density function of each member; a _i (i = 1, 2, ..., M) is the mixed weight.

完整的GMM可表示为：The complete GMM can be expressed as:

λ_i＝{a_i，μ_i，∑_i}，(i＝1，2，L，M) (13)λ _i = {a _i , μ _i , Σ _i }, (i=1, 2, L, M) (13)

每个成员密度函数是一个D维变量的高斯分布函数，表示为：Each member density function is a Gaussian distribution function of a D-dimensional variable, expressed as:

${b b}_{i i} ((x x)) = = \frac{11}{{((22 π π))}^{D D. / / 22} {| | Σ Σ | |}^{11 / / 22}} exp exp [[- - \frac{11}{22} {((x x - - {μ μ}_{i i}))}^{T T} {Σ Σ}^{- - 11} ((x x - - {μ μ}_{i i}))]] - - - - - - ((1414))$

对于一个长度为T的测试语音时间序列X＝(x₁，x₂，...，x_T)，它的GMM似然概率可表示为：For a test speech time sequence X=(x ₁ , x ₂ ,..., x _T ) whose length is T, its GMM likelihood probability can be expressed as:

$P P ((X x | | {λ λ}_{i i})) = = {Π Π}_{t t = = 11}^{T T} p p (({x x}_{t t} | | {λ λ}_{i i})) - - - - - - ((1515))$

$L (X | λ_{i}) = \lg [P (X | λ_{i})]$ (16) $L (x | λ_{i}) = \lg [P (x | λ_{i})]$ (16)

$= = {Σ Σ}_{t t = = 11}^{T T} lg lg [[p p (({x x}_{t t} | | {λ λ}_{i i}))]]$

识别时运用贝叶斯定理，在N个未知话者的模型中，得到似然概率最大的模型对应的话者即为识别结果：Bayesian theorem is used in the recognition. Among the models of N unknown speakers, the speaker corresponding to the model with the largest likelihood probability is the recognition result:

${t t}^{* *} = = \underset{11 \leq \leq i i \leq \leq N N}{arg arg} {{max max [[L L ((X x | | {λ λ}_{i i}))]]}} - - - - - - ((1717))$

在实际应用中，每个GMM模型的规模通常取为30-50个高斯分布。In practical applications, the scale of each GMM model is usually taken as 30-50 Gaussian distributions.

在人脸识别和声纹识别过程中，每个子识别系统会给出一个匹配结果s(score)来表征测试样本与匹配模板的相似度。由于处理的样本数据有差异，人脸识别系统与声纹识别系统的决策形式上会存在差异。为了消除数值形式差异对最后分类结果的影响，首先利用归一化函数把来自不同分类器的不同值域区间与度量方法的s转化为同一值域区间可以互相比较的s′，然后利用分类能力较强的SVM分类器对归一化的s′进行分类。在本文中，采用归一化函数采用Z-score函数。In the process of face recognition and voiceprint recognition, each sub-recognition system will give a matching result s(score) to represent the similarity between the test sample and the matching template. Due to the difference in the sample data processed, there will be differences in the decision-making form of the face recognition system and the voiceprint recognition system. In order to eliminate the impact of the differences in numerical forms on the final classification results, first use the normalization function to transform s from different range intervals and measurement methods from different classifiers into s′ that can be compared with each other in the same range interval, and then use the classification ability The stronger SVM classifier classifies the normalized s′. In this paper, the Z-score function is adopted as a normalization function.

具体应用于实验的Z-score函数的形式：The form of the Z-score function specifically applied to the experiment:

${s the s}^{' '} = = \frac{s the s - - mean mean (({{s the s}}))}{std std (({{s the s}}))} - - - - - - ((1818))$

在这里s′是归一化后的匹配结果，mean( )和std( )分别代表求均值与求均方差运算，{s}是一个分类器对所有测试样本的匹配打分集合。这种归一化方法比较简单，只需要估计均值和方差就可以对s进行归一化。Here s′ is the normalized matching result, mean( ) and std( ) respectively represent mean value and mean square error calculation, and {s} is a set of matching scores for all test samples by a classifier. This normalization method is relatively simple, and s can be normalized only by estimating the mean and variance.

对归一化后的结果采用了SVM进行分类，充分利用了归一化后s′的信息；与直接采用SVM分类器相比，进行归一化可以将结果调整到可以比较的同一数值区间，使分类结果更加准确。采用归一化SVM融合方式的系统性能会比直接相加方式有较大提高，而且当分类数越多提高越明显。这是由于当分类数越大时，每个子系统的性能会变差，各子系统之间的关系也变复杂。如果只对各子系统的决策分数直接相加的时候，系统性能会随着分类数的增多而越来越差。反而SVM融合模型可以较为充分地描述各子系统之间的非线性关系，所以性能会有所改进。因而，当分类数越大时，SVM融合的优势就越明显。SVM is used to classify the normalized results, making full use of the information of s′ after normalization; compared with directly using the SVM classifier, normalization can adjust the results to the same value range that can be compared. Make the classification result more accurate. The performance of the system using the normalized SVM fusion method will be greatly improved compared with the direct addition method, and the improvement will be more obvious when the number of classifications increases. This is because when the classification number is larger, the performance of each subsystem will become worse, and the relationship between subsystems will also become complicated. If only the decision scores of each subsystem are directly added, the system performance will become worse and worse as the number of classifications increases. On the contrary, the SVM fusion model can fully describe the nonlinear relationship between subsystems, so the performance will be improved. Therefore, when the number of categories is larger, the advantages of SVM fusion are more obvious.

如图2所示：工作时，输入声像检测设备持续检测声音信号或图像信号；当有来访者时，声音采集设备与图像采集设备启动，分别采集来访者的声音信号和人脸图像信号。门禁控制器将声音信号和人脸图像信号输出到数据处理器内，数据处理器对声音信号利用梅尔倒谱系数和混合高斯模型进行特征提取和声音识别；同时利用PCA和LDA对人脸图像信号进行特征提取和人脸识别。数据处理器对识别后的声音信号和人脸图像识别后，通过Z-score函数进行归一化处理，并通过SVM对归一化的数据进行融合处理，从而形成决策层融合。当决策层融合后的数据与数据处理器内的样本库进行比较，当融合数据与样本库的数据相匹配时，数据处理器向门禁控制器发出匹配信号，门禁控制器打开电控门锁，完成开门动作；当融合数据与样本库内的数据不匹配时，数据处理器启动计数器，对比较次数进行比较；当比较达到设定次数时，门禁控制器通过声光报警器输出声光报警信号；当比较次数小于设定次数时，门禁控制器重新接收声音采集设备和图像采集设备的采集信号。As shown in Figure 2: when working, the input sound and image detection equipment continuously detects the sound signal or image signal; when there is a visitor, the sound collection device and image collection device start to collect the visitor's sound signal and face image signal respectively. The access control controller outputs the sound signal and face image signal to the data processor, and the data processor uses the Mel cepstrum coefficient and the mixed Gaussian model to perform feature extraction and sound recognition on the sound signal; meanwhile, PCA and LDA are used to analyze the face image The signal is used for feature extraction and face recognition. After the data processor recognizes the recognized sound signal and face image, it performs normalization processing through the Z-score function, and performs fusion processing on the normalized data through SVM, thereby forming a decision-making layer fusion. When the fused data at the decision-making level is compared with the sample library in the data processor, when the fused data matches the data in the sample library, the data processor sends a matching signal to the access controller, and the access controller opens the electronically controlled door lock. Complete the door opening action; when the fusion data does not match the data in the sample library, the data processor starts the counter to compare the number of comparisons; when the comparison reaches the set number of times, the access control controller outputs an audible and visual alarm signal through the audible and visual alarm ; When the number of comparisons is less than the set number of times, the access control controller receives the acquisition signals from the sound acquisition device and the image acquisition device again.

本发明通过对人脸图像信号和声音信号进行采集，并采用现有的处理方法对人脸图像和声音信号进行特征提取和识别，可以应用于机场、银行、公安机关、考勤系统或其他方面。The present invention can be applied to airports, banks, public security organs, attendance systems or other aspects by collecting face image signals and sound signals, and adopting existing processing methods to perform feature extraction and recognition on face images and sound signals.

本发明门禁控制器同时接收声音采集设备与图像采集设备的声音信号、人脸图像信号；数据处理器对声音信号利用梅尔倒谱系数和混合高斯模型方法进行特征提取和声音识别，数据处理器对人脸图像信号利用主成分分析法和线性判别分析法进行特征提取和声音设备，数据处理器对提取和识别后的声音信号及人脸图像信号进行Z-score归一化和支持向量机分类融合，降低了识别的等错率，提高了门禁系统的安全可靠性，安全可靠。The access control controller of the present invention simultaneously receives the sound signal and the face image signal of the sound collection device and the image collection device; the data processor uses the Mel cepstrum coefficient and the mixed Gaussian model method to perform feature extraction and sound recognition on the sound signal, and the data processor Use principal component analysis and linear discriminant analysis to perform feature extraction and sound equipment on the face image signal, and the data processor performs Z-score normalization and support vector machine classification on the extracted and recognized sound signal and face image signal Fusion reduces the error rate of identification and improves the safety and reliability of the access control system, which is safe and reliable.

Claims

1. An intelligent access control system based on multi-biometric identification, comprising sound acquisition equipment and image acquisition equipment, characterized in that: said sound acquisition equipment and image acquisition equipment are all connected to the input of access controller, and said access control The device respectively receives the sound signal collected by the sound collection device and the face image signal collected by the image collection device; the output terminals of the access control controller are respectively connected with the data processor and the electric control door lock; the access control controller will receive the sound signal and the face image signal The image signal is transmitted to the data processor; the data processor performs feature extraction on the received sound signal and face image signal respectively, recognizes the collected sound signal and face image signal, and performs corresponding signal extraction after feature extraction. Normalization and classification fusion processing; the data processor compares the corresponding signal after classification fusion processing with the preset sample library; when the signal obtained after classification fusion processing matches the corresponding signal in the sample library, the access controller turns on the power switch Control the door lock.

2. The intelligent access control system based on multi-biometric feature identification according to claim 1, characterized in that: the data processor comprises a DSP.

3. the intelligent access control system based on multi-biometric feature recognition according to claim 1, is characterized in that: said data processor utilizes principal component analysis method and linear discriminant analysis method to carry out feature extraction and face recognition to face image signal .

4. The intelligent access control system based on multi-biological feature recognition according to claim 1, characterized in that: said data processor uses Mel cepstral coefficient and mixed Gaussian model method to perform feature extraction and face recognition on sound signals.

5. The intelligent access control system based on multi-biometric feature recognition according to claim 1, characterized in that: said data processor uses a z-score function to normalize the recognized sound signal and face image signal.

6. The intelligent access control system based on multi-biometric feature recognition according to claim 1, characterized in that: the data processor uses a support vector machine to classify and fuse the normalized sound signal and face image signal.

7. The intelligent access control system based on multi-biometric identification according to claim 1 is characterized in that: the output end of the access control controller is also connected with the sound and light alarm; the signal and data processing obtained after human fusion processing When the sample library in the device does not match, the access controller will output an audible and visual alarm signal through the audible and visual alarm.