CN103927531A

CN103927531A - Human face recognition method based on local binary value and PSO BP neural network

Info

Publication number: CN103927531A
Application number: CN201410200902.XA
Authority: CN
Inventors: 丁欢欢; 杨永红
Original assignee: Jiangsu University of Science and Technology
Current assignee: Jiangsu University of Science and Technology
Priority date: 2014-05-13
Filing date: 2014-05-13
Publication date: 2014-07-16
Anticipated expiration: 2034-05-13
Also published as: CN103927531B

Abstract

The invention discloses a face recognition method based on local binary value and particle swarm optimization BP neural network. Firstly, the non-overlapping images of each type of face in the known face database are divided into a training sample set and a test sample set. Perform normalization and local binary preprocessing; secondly, do two-dimensional discrete wavelet transform on the preprocessed image, remove the influence of the diagonal component, weight the remaining three frequency band components, and then do the fused image Two-dimensional discrete cosine transform, using the zigzag scanning method to extract its main transformation coefficient matrix; again, using the particle swarm optimization algorithm to optimize the initial weight and threshold of the BP neural network for network training; finally, sending the test sample set data into the trained The BP neural network is tested and the recognition rate is calculated. The invention has higher computing efficiency and recognition ability, and is suitable for a face recognition system.

Description

A Face Recognition Method Based on Local Binary and Particle Swarm Optimization BP Neural Network

技术领域technical field

本发明属于一种人脸识别方法，尤其涉及一种基于局部二值和粒子群优化BP神经网络的人脸识别算法，属于智能模式识别与图像处理领域。The invention belongs to a face recognition method, in particular to a face recognition algorithm based on local binary and particle swarm optimization BP neural network, and belongs to the field of intelligent pattern recognition and image processing.

背景技术Background technique

近年来，人脸识别技术得到了迅猛的发展，大量高性能算法的出现使其从实验室走向了商用。然而，到目前为止，人脸识别技术仍面临着巨大的挑战：(1)光照、背景、姿态、表情、遮挡物以及年龄变化；(2)成像条件及设备差异；(3)数据规模的限制等。因此，人脸识别技术中高识别率问题依旧未得到彻底解决。In recent years, face recognition technology has developed rapidly, and the emergence of a large number of high-performance algorithms has made it move from the laboratory to commercial use. However, so far, face recognition technology is still facing huge challenges: (1) illumination, background, posture, expression, occlusion and age changes; (2) imaging conditions and equipment differences; (3) data scale limitations wait. Therefore, the problem of high recognition rate in face recognition technology has not been completely resolved.

环境光线的变化是影响人脸识别精度的主要因素之一。研究发现，同一人脸图像在不同光照条件下的差异往往比不同人脸图像在相同光照条件下的差异要大得多。而且现有很多人脸识别系统都是在严格的光照条件下使用的，具有很大的局限性。因此人脸分类识别前光照预处理是非常有必要的，预处理的好坏直接影响到系统性能的优劣。The change of ambient light is one of the main factors affecting the accuracy of face recognition. It is found that the difference of the same face image under different lighting conditions is often much larger than that of different face images under the same lighting conditions. Moreover, many existing face recognition systems are used under strict lighting conditions, which have great limitations. Therefore, light preprocessing before face classification and recognition is very necessary, and the quality of preprocessing directly affects the performance of the system.

人脸图像一般为二维图像，所含信息数据量巨大，容易造成维数灾难，不利于分类识别。为了降低计算的复杂度，提高系统的运算速度和识别率，必须对图像进行压缩降维处理，用尽可能少的数据表示尽可能多的信息。Face images are generally two-dimensional images, which contain a huge amount of information and data, which is easy to cause the disaster of dimensionality and is not conducive to classification and recognition. In order to reduce the computational complexity and improve the system's computing speed and recognition rate, it is necessary to compress and reduce the dimension of the image, and use as little data as possible to represent as much information as possible.

分类器的选择是影响人脸识别率的又一关键性因素。分类器的种类有很多，但是由于神经网络可以获得其他方法难以实现的识别规律和隐形表达，而且它并行的处理方式可以显著提高计算速度，因此，备受广大学者关注，被广泛应用到各个领域中。虽然神经网络在一定程度上降低了运算的复杂度，但神经网络各参数的设置并没有特定的原理算法，需要我们凭借经验来取值，另外，在神经网络的收敛速度和极易陷入局部极小值等问题方面也未彻底解决。The choice of classifier is another key factor affecting the face recognition rate. There are many types of classifiers, but because the neural network can obtain recognition rules and invisible expressions that are difficult to achieve by other methods, and its parallel processing method can significantly improve the calculation speed, it has attracted the attention of scholars and has been widely used in various fields. middle. Although the neural network reduces the complexity of the operation to a certain extent, there is no specific principle and algorithm for setting the parameters of the neural network, and we need to rely on experience to select the value. Problems such as small values have not been completely resolved.

发明内容Contents of the invention

发明目的：为了克服现有的不足，本发明提出一种基于局部二值和粒子群优化BP神经网络的人脸识别方法。Purpose of the invention: In order to overcome the existing deficiencies, the present invention proposes a face recognition method based on local binary and particle swarm optimization BP neural network.

技术方案：一种基于局部二值和粒子群优化BP神经网络的人脸识别方法，包括以下步骤：Technical solution: a face recognition method based on local binary and particle swarm optimization BP neural network, comprising the following steps:

(1)将已知人脸库中每类人脸图像随机抽取一定数目作为训练样本集I_train＝{X₁,X₂,...,X_j,...,X_A}，其中，X_j为每个训练样本图像，A为训练样本数；(1) A certain number of each type of face images in the known face database is randomly selected as a training sample set I _train ={X ₁ ,X ₂ ,...,X _j ,...,X _A }, where X _j is each training sample image, A is the number of training samples;

(2)对训练样本集中每幅M×N像素的灰度图像进行几何归一化处理，将其归一化为H×H大小的图像，记为I′_train，其中0＜H≤min(M,N)；(2) Perform geometric normalization processing on each M×N pixel grayscale image in the training sample set, and normalize it into an H×H size image, denoted as I′ _train , where 0<H≤min( M,N);

(3)利用局部二值算法对几何归一化后的训练集图像I′_train提取光照不变量，去除光照影响，得到光照处理后的训练集图像I″_train；(3) Utilize the local binary algorithm to extract the illumination invariant to the training set image I' _train after geometric normalization, remove the influence of illumination, and obtain the training set image I" _train after illumination processing;

(4)对步骤(3)中经过局部二值算法处理后的图像集I″_train进行加权二维离散小波变换，得到变换后的图像集I_train,DWT；(4) carry out weighted two-dimensional discrete wavelet transform to the image set I " _train after local binary algorithm processing in step (3), obtain transformed image set I _{train, DWT} ;

(5)对步骤(4)得到的图像集I_train,DWT中每个样本图像做二维离散余弦变换，得到变换系数矩阵Y＝{Y₁,Y₂,...,Y_h,...,Y_A}，其中，Y_h为每个样本图像经过二维离散余弦变换后得到的变换系数向量，然后，将变换系数矩阵Y中每个向量利用zigzag扫描方式展开，最后，提取每个展开向量的主分量，组成最优特征向量E；(5) Perform two-dimensional discrete cosine transform on each sample image in the image set I _{train and DWT} obtained in step (4), and obtain the transformation coefficient matrix Y={Y ₁ , Y ₂ ,...,Y _h ,... ., Y _A }, where Y _h is the transform coefficient vector obtained after each sample image undergoes two-dimensional discrete cosine transform, and then expands each vector in the transform coefficient matrix Y by zigzag scanning, and finally extracts each Expand the principal components of the vector to form the optimal eigenvector E;

(6)设置神经网络参数，确定BP神经网络的拓扑结构，输入层节点数Q、隐含层节点数W、输出层节点数Z、激活函数Sigmoid函数；(6) Neural network parameters are set to determine the topology of the BP neural network, the number of input layer nodes Q, the number of hidden layer nodes W, the number of output layer nodes Z, and the activation function Sigmoid function;

(7)通过粒子群算法优化BP神经网络权值和阀值；(7) Optimizing the weight and threshold of BP neural network through particle swarm algorithm;

(8)将步骤(7)得到的全局最优值映射为神经网络的初始权值和阈值，训练BP神经网络；(8) The global optimal value that step (7) obtains is mapped to initial weight and threshold of neural network, trains BP neural network;

(9)将人脸库中其余图像作为测试样本集I_test，将其重复步骤(2)到步骤(5)的处理，然后将测试样本数据输入到步骤(8)所得到的已训练好的BP神经网络中进行测试，计算识别率。(9) use the remaining images in the face bank as the test sample set _Itest , repeat the process from step (2) to step (5), and then input the test sample data to the trained one obtained by step (8) Test in BP neural network and calculate the recognition rate.

所述步骤(3)具体为：Described step (3) is specifically:

首先，将I′_train中的每幅图像分成n×n的小块，把每个小块内n×n个像素点灰度值的均值作为新图像的一个像素值，得到分块训练集图像I_{train_block}；First, divide each image in the I′ _train into n×n small blocks, and use the mean value of the gray value of n×n pixels in each small block as a pixel value of the new image to obtain the block training set image I _{train_block} ;

然后，利用局部二值算法来描述人脸图像的光照不变量特征，对于I_{train_block}中任意一图像I，I∈I_{train_block}，其任意一点(x_c,y_c)及均匀分布在该中心点周围的P个邻域点的局部二值特征算子LBP为：Then, use the local binary algorithm to describe the illumination invariant features of the face image. For any image I in I _{train_block} , I∈I _{train_block} , any point (x _c , y _c ) and uniformly distributed around the center point The local binary feature operator LBP of the P neighborhood points of is:

$\{\begin{matrix} {LBP LBP}_{R R,, P P} = = {Σ Σ}_{q q = = 00}^{P P - - 11} s the s (({g g}_{q q} - - {g g}_{c c})) \cdot &Center Dot; 22^{q q},, \\ s the s ((x x)) = = \{\begin{matrix} 11,, x x &GreaterEqual; &Greater Equal; 00 \\ 00,, x x < < 00 \end{matrix} . . \end{matrix} - - - - - - ((11))$

其中，g_c表示区域内中心像素点(x_c,y_c)的灰度值，g_q(q＝0,1,...,P-1)表示均匀分布在以中心点(x_c,y_c)为圆心，半径为R的圆周上P个采样点的灰度值。Among them, g _c represents the gray value of the center pixel point (x _c , y _c ) in the area, and g _q (q=0,1,...,P-1) represents the gray value evenly distributed at the center point (x c , y _c ). y _c ) is the gray value of P sampling points on the circle whose center is the radius R.

所述步骤(4)具体为：Described step (4) is specifically:

在图像集I″_train中选取任意一幅图像F，对其进行二维离散小波变换，得到LL,LH,HL,HH四个方向上的图像，记为F_LL,F_LH,F_HL,F_HH，则加权后的图像F′为：Select any image F in the image set I″ _train , and perform two-dimensional discrete wavelet transform on it to obtain images in the four directions of LL, LH, HL, and HH, which are denoted as F _LL , _FLH , F _HL , F _HH , then the weighted image F′ is:

F′＝a₀F_LL+a₁F_LH+a₂F_HL (2)F'＝a ₀ F _LL +a ₁ F _LH +a ₂ F _HL (2)

其中，a₀,a₁,a₂为加权系数，约束条件为a₀+a₁+a₂＝1；低频分量F_LL为原图像的平滑图像，保持了原图像的大部分信息；F_LH分量保持了原图像的垂直边缘细节；F_HL分量保持了原图像的水平边缘细节；F_HH分量保持了原图像对角线方向上的边缘细节。Among them, a ₀ , a ₁ , and a ₂ are weighting coefficients, and the constraint condition is a ₀ +a ₁ +a ₂ =1; the low-frequency component F _LL is a smooth image of the original image, which maintains most of the information of the original image; F _LH The component maintains the vertical edge details of the original image; the F _HL component maintains the horizontal edge details of the original image; the F _HH component maintains the edge details in the diagonal direction of the original image.

所述步骤(7)具体为：Described step (7) is specifically:

a.设置粒子群参数，包括粒子种群规模B、维度D、最大迭代次数T_max、学习因子c₁和c₂、惯性权重ω、最大速度υ_max、最大位置x_max、期望误差最小值ε、在允许范围内随机产生粒子的初始速度及位置；a. Set particle swarm parameters, including particle population size B, dimension D, maximum number of iterations T _max , learning factors c ₁ and c ₂ , inertia weight ω, maximum velocity υ _max , maximum position x _max , minimum expected error ε, Randomly generate the initial velocity and position of the particles within the allowable range;

b.计算每个粒子的适应度J_a；b. Calculate the fitness J _a of each particle;

c.根据适应度J_a来确定每个粒子的个体最优极值和全局最优极值，将当前适应度值与每个粒子历史最优适应度比较，适应度小者作为个体最优极值P_best，与整个种群最优适应度比较，适应度小者作为全局最优极值G_best；c. Determine the individual optimal extremum and the global optimal extremum of each particle according to the fitness J _a , compare the current fitness value with the historical optimal fitness of each particle, and the one with the smaller fitness is the individual optimal extremum The value P _best is compared with the optimal fitness of the entire population, and the one with the smaller fitness is taken as the global optimal extreme value G _best ;

d.更新每个粒子的速度和位置，并且考虑更新后的速度和位置是否在限定范围内；d. Update the velocity and position of each particle, and consider whether the updated velocity and position are within the limits;

${&upsi; &upsi;}_{id id}^{T T + + 11} = = {ω&upsi; ω&upsi;}_{id id}^{T T} + + {c c}_{11} {r r}_{11} (({p p}_{id id}^{T T} - - {x x}_{id id}^{T T})) + + {c c}_{22} {r r}_{22} (({p p}_{gd gd}^{T T} - - {x x}_{id id}^{T T})) - - - - - - ((33))$

${x x}_{id id}^{T T + + 11} = = {x x}_{id id}^{T T} + + {&upsi; &upsi;}_{id id}^{T T} - - - - - - ((44))$

其中，i＝1,2,...,B，d＝1,2,...,D，为第T次迭代粒子i速度矢量的第d维分量；为第T次迭代粒子i位置矢量的第d维分量；为粒子i个体当前最好位置P_best的第d维分量；为当前群体最好位置G_best的第d维分量；r₁和r₂为服从[0，1]均匀分布的随机数；Where, i=1,2,...,B, d=1,2,...,D, is the d-th dimension component of the velocity vector of particle i in the T iteration; is the d-th dimension component of the position vector of particle i in the T-th iteration; is the d-th dimension component of the current best position P _best of particle i; is the d-th dimension component of the best position G _best of the current group; r ₁ and r ₂ are random numbers that obey the uniform distribution of [0, 1];

约束条件1：若 ${&upsi;}_{id}^{T + 1} > {&upsi;}_{\max},$ 则 ${&upsi;}_{id}^{T + 1} = {&upsi;}_{\max};$ 若 ${&upsi;}_{id}^{T + 1} < - {&upsi;}_{\max},$ 则 ${&upsi;}_{id}^{T + 1} = - {&upsi;}_{\max};$ Constraint 1: If ${&upsi;}_{id}^{T + 1} > {&upsi;}_{\max},$ but ${&upsi;}_{id}^{T + 1} = {&upsi;}_{\max};$ like ${&upsi;}_{id}^{T + 1} < - {&upsi;}_{\max},$ but ${&upsi;}_{id}^{T + 1} = - {&upsi;}_{\max};$

约束条件2：若 $x_{id}^{T + 1} > x_{\max},$ 则 $x_{id}^{T + 1} = x_{\max};$ 若 $x_{id}^{T + 1} < - x_{\max},$ 则 $x_{id}^{T + 1} = - x_{\max};$ Constraint 2: If $x_{id}^{T + 1} > x_{\max},$ but $x_{id}^{T + 1} = x_{\max};$ like $x_{id}^{T + 1} < - x_{\max},$ but $x_{id}^{T + 1} = - x_{\max};$

e.比较迭代次数是否达到最大T_max或者均方误差是否达到精度ε要求，若满足，则算法收敛，记录最后一次迭代的个体最优值P_best和全局最优值G_best；否则，返回步骤c。e. Compare whether the number of iterations reaches the maximum T _max or whether the mean square error meets the precision ε requirement. If so, the algorithm converges, and record the individual optimal value P _best and the global optimal value G _best of the last iteration; otherwise, return to the step c.

所述步骤b的计算方法具体为：The calculation method of the step b is specifically:

${J J}_{a a} = = \frac{11}{22 A A} {Σ Σ}_{k k = = 11}^{A A} {Σ Σ}_{t t = = 11}^{Z Z} {(({y the y}_{k k}^{t t} - - {c c}_{t t}^{k k}))}^{22} - - - - - - ((55))$

其中，a＝1,2,...,B；是第k个样本的第t个网络输出神经元的期望输出值；是第k个样本的第t个网络输出神经元的实际输出值。Among them, a=1,2,...,B; is the expected output value of the tth network output neuron of the kth sample; is the actual output value of the tth network output neuron of the kth sample.

工作原理：本发明首先将已知人脸库中每类人脸图像无重叠的分为训练样本集和测试样本集，对图像进归一化和局部二值预处理；其次，对预处理后的图像做二维离散小波变换，去除对角线分量的影响，将其余三个频带分量加权融合，再对融合后的图像做二维离散余弦变换，利用zigzag扫描方式提取其主要变换系数矩阵；再次，利用粒子群算法优化BP神经网络的初始权值和阈值进行网络训练；最后，将测试样本集数据送入到已训练好的BP神经网络中进行测试，计算识别率。Working principle: the present invention firstly divides the non-overlapping images of each type of face in the known face database into a training sample set and a test sample set, and performs normalization and local binary preprocessing on the images; secondly, the preprocessed Perform two-dimensional discrete wavelet transform on the image to remove the influence of the diagonal component, weight and fuse the remaining three frequency band components, and then perform two-dimensional discrete cosine transform on the fused image, and use the zigzag scanning method to extract its main transformation coefficient matrix; again , use the particle swarm optimization algorithm to optimize the initial weights and thresholds of the BP neural network for network training; finally, send the test sample set data into the trained BP neural network for testing, and calculate the recognition rate.

有益效果：本发明通过局部二值算法去除光照影响，加权离散小波变换与离散余弦变换相结合来进行特征提取，再利用粒子群优化神经网络的权值和阈值来进行分类识别，具有较强的鲁棒性和寻优能力。与现有技术相比，本发明具有较高的运算效率和识别能力，适用于人脸识别系统。Beneficial effects: the present invention removes the influence of illumination through local binary algorithm, combines weighted discrete wavelet transform and discrete cosine transform to extract features, and then uses particle swarm optimization neural network weights and thresholds to classify and identify, which has strong Robustness and optimization ability. Compared with the prior art, the invention has higher computing efficiency and recognition ability, and is suitable for a face recognition system.

附图说明Description of drawings

图1是本发明框图；Fig. 1 is a block diagram of the present invention;

图2是本发明中粒子群优化BP神经网络的流程图。Fig. 2 is the flowchart of particle swarm optimization BP neural network in the present invention.

具体实施方式Detailed ways

如图1、2所示，一种基于局部二值和粒子群优化BP神经网络的人脸识别方法包括以下步骤：As shown in Figures 1 and 2, a face recognition method based on local binary and particle swarm optimization BP neural network includes the following steps:

步骤1：将已知人脸库中每类人脸图像随机抽取一定数目作为训练样本集I_train＝{X₁,X₂,...,X_j,...,X_A}，其中，X_j为每个训练样本图像，A为训练样本数，其余作为测试样本集I_test；Step 1: Randomly select a certain number of face images of each type in the known face database as a training sample set I _train ={X ₁ ,X ₂ ,...,X _j ,...,X _A }, where X _j is each training sample image, A is the number of training samples, and the rest are used as the test sample set _Itest ;

步骤2：对训练集中每幅M×N像素的灰度图像进行几何归一化处理，将其归一化为H×H大小的图像，记为I′_train，其中0＜H≤min(M,N)；Step 2: Perform geometric normalization on each grayscale image of M×N pixels in the training set, and normalize it into an image of H×H size, denoted as I′ _train , where 0<H≤min(M ,N);

步骤3：利用局部二值算法对归一化后的训练集图像I′_train提取光照不变量，去除光照影响，得到光照处理后的训练集图像I″_train，其过程为：Step 3: Use the local binary algorithm to extract the illumination invariant from the normalized training set image I′ _train , remove the influence of illumination, and obtain the training set image I″ _train after illumination processing. The process is as follows:

(1)首先，将I′_train中的每幅图像分成n×n的小块，把每个小块内n×n个像素点灰度值的均值作为新图像的一个像素值，得到分块训练集图像I_{train_block}；(1) First, divide each image in the I′ _train into small blocks of n×n, and use the mean value of the gray value of n×n pixels in each small block as a pixel value of the new image to obtain the block training set image I _{train_block} ;

(2)利用局部二值算法来描述人脸图像的光照不变量特征，对于I_{train_block}中任意一图像I，I∈I_{train_block}，其任意一点(x_c,y_c)及均匀分布在该中心点周围的P个邻域点的局部二值特征算子LBP表示为：(2) Use the local binary algorithm to describe the illumination invariant features of the face image. For any image I in I _{train_block} , I∈I _{train_block} , any point (x _c , y _c ) and uniformly distributed at the center point The local binary feature operator LBP of the surrounding P neighborhood points is expressed as:

其中，g_c表示区域内中心像素点(x_c,y_c)的灰度值，g_q(q＝0,1,...,P-1)表示均匀分布在以中心点(x_c,y_c)为圆心，半径为R的圆周上P个采样点的灰度值；Among them, g _c represents the gray value of the center pixel point (x _c , y _c ) in the area, and g _q (q=0,1,...,P-1) represents the gray value evenly distributed at the center point (x c , y _c ). y _c ) is the center of the circle, the gray value of P sampling points on the circle whose radius is R;

步骤4：对上述经过局部二值处理后的图像集I″_train进行加权二维离散小波变换，得到变换后的图像集I_train,DWT：Step 4: Perform weighted two-dimensional discrete wavelet transform on the above-mentioned image set I″ _train after local binary processing to obtain the transformed image set I _train,DWT :

其中，a₀,a₁,a₂为加权系数，约束条件为a₀+a₁+a₂＝1；低频分量F_LL为原图像的平滑图像，保持了原图像的大部分信息；F_LH分量保持了原图像的垂直边缘细节；F_HL分量保持了原图像的水平边缘细节；F_HH分量保持了原图像对角线方向上的边缘细节；由于人脸图像是非刚性的，对角线方向信息的稳定性较差，含噪声较多，非常不利于特征提取，所以将其舍去；Among them, a ₀ , a ₁ , and a ₂ are weighting coefficients, and the constraint condition is a ₀ +a ₁ +a ₂ =1; the low-frequency component F _LL is a smooth image of the original image, which maintains most of the information of the original image; F _LH The component maintains the vertical edge details of the original image; the F _HL component maintains the horizontal edge details of the original image; the F _HH component maintains the edge details in the diagonal direction of the original image; since the face image is non-rigid, the diagonal direction The stability of the information is poor, and it contains more noise, which is very unfavorable for feature extraction, so it is discarded;

步骤5：对图像集I_train,DWT中每个样本图像做二维离散余弦变换，得到变换系数矩阵Y＝{Y₁,Y₂,...,Y_h,...,Y_A}，其中，Y_h为每个样本图像经过二维离散余弦变换后得到的变换系数向量；然后，将变换系数矩阵Y中每个向量利用zigzag扫描方式展开；最后，提取每个展开向量的主分量，组成最优特征向量E，即为神经网络的输入；Step 5: Perform a two-dimensional discrete cosine transform on each sample image in the image set I _train,DWT to obtain the transform coefficient matrix Y={Y ₁ ,Y ₂ ,...,Y _h ,...,Y _A }, Among them, Y _h is the transformation coefficient vector obtained after two-dimensional discrete cosine transformation of each sample image; then, each vector in the transformation coefficient matrix Y is expanded by zigzag scanning; finally, the principal component of each expansion vector is extracted, Form the optimal feature vector E, which is the input of the neural network;

步骤6：如图2所示：设置神经网络参数，确定BP神经网络的拓扑结构：输入层节点数Q、隐含层节点数W、输出层节点数Z；激活函数Sigmoid函数；Step 6: As shown in Figure 2: set the neural network parameters and determine the topology of the BP neural network: the number of input layer nodes Q, the number of hidden layer nodes W, the number of output layer nodes Z; the activation function Sigmoid function;

步骤7：设置粒子群参数：粒子种群规模B，其中20≤B≤100；维度D，其中D＝QW+WZ+W+Z；最大迭代次数T_max；学习因子c₁和c₂，其中1≤c₁≤2，1≤c₂≤2，一般取c₁＝c₂；惯性权重ω，其中0＜ω＜1；最大速度υ_max；最大位置x_max；在允许范围内随机产生粒子的初始速度、位置；期望误差最小值ε；Step 7: Set particle swarm parameters: particle population size B, where 20≤B≤100; dimension D, where D=QW+WZ+W+Z; maximum number of iterations T _max ; learning factors c ₁ and c ₂ , where 1 ≤c ₁ ≤2, 1≤c ₂ ≤2, generally take c ₁ =c ₂ ; inertia weight ω, where 0<ω<1; maximum velocity υ _max ; maximum position x _max ; within the allowable range randomly generated particles Initial velocity, position; minimum value of expected error ε;

步骤8：计算每个粒子的适应度：先输入一个粒子，计算所有样本均方差，即该粒子的适应度，即：Step 8: Calculate the fitness of each particle: first input a particle, calculate the mean square error of all samples, that is, the fitness of the particle, namely:

${J J}_{a a} = = \frac{11}{22 A A} {Σ Σ}_{k k = = 11}^{A A} {Σ Σ}_{t t = = 11}^{Z Z} {(({y the y}_{k k}^{t t} - - {c c}_{t t}^{k k}))}^{22} - - - - - - ((33))$

其中，a＝1,2,...,B；是第k个样本的第t个网络输出神经元的期望输出值；是第k个样本的第t个网络输出神经元的实际输出值；同理，继续输入其它粒子，直至计算出所有粒子的适应度；Among them, a=1,2,...,B; is the expected output value of the tth network output neuron of the kth sample; is the actual output value of the tth network output neuron of the kth sample; similarly, continue to input other particles until the fitness of all particles is calculated;

步骤9：根据适应度J_a来确定每个粒子的个体最优极值和全局最优极值，将当前适应度值与每个粒子历史最优适应度比较，适应度小者作为个体最优极值P_best，与整个种群最优适应度比较，适应度小者作为全局最优极值G_best；Step 9: Determine the individual optimal extremum and the global optimal extremum of each particle according to the fitness J _a , compare the current fitness value with the historical optimal fitness of each particle, and the one with the smaller fitness is the individual optimal The extreme value P _best is compared with the optimal fitness of the entire population, and the one with the smallest fitness is the global optimal extreme value G _best ;

步骤10：更新每个粒子的速度和位置，并且考虑更新后的速度和位置是否在限定范围内；Step 10: Update the velocity and position of each particle, and consider whether the updated velocity and position are within the limit;

${&upsi; &upsi;}_{id id}^{T T + + 11} = = {ω&upsi; ω&upsi;}_{id id}^{T T} + + {c c}_{11} {r r}_{11} (({p p}_{id id}^{T T} - - {x x}_{id id}^{T T})) + + {c c}_{22} {r r}_{22} (({p p}_{gd gd}^{T T} - - {x x}_{id id}^{T T})) - - - - - - ((44))$

${x x}_{id id}^{T T + + 11} = = {x x}_{id id}^{T T} + + {&upsi; &upsi;}_{id id}^{T T} - - - - - - ((55))$

约束条件2：若 $x_{id}^{T + 1} > x_{\max},$ 则 $x_{id}^{T + 1} = x_{\max};$ 若 $x_{id}^{T + 1} < - x_{\max},$ 则 $x_{id}^{T + 1} > x_{\max},$ Constraint 2: If $x_{id}^{T + 1} > x_{\max},$ but $x_{id}^{T + 1} = x_{\max};$ like $x_{id}^{T + 1} < - x_{\max},$ but $x_{id}^{T + 1} > x_{\max},$

步骤11：比较迭代次数是否达到最大T_max或者均方误差是否达到精度ε要求，若满足，则算法收敛，记录最后一次迭代的个体最优值P_best和全局最优值G_best；否则，返回步骤9，继续迭代；Step 11: Compare whether the number of iterations reaches the maximum T _max or whether the mean square error meets the precision ε requirement. If so, the algorithm converges, and record the individual optimal value P _best and the global optimal value G _best of the last iteration; otherwise, return Step 9, continue to iterate;

步骤12：将全局最优值映射为神经网络的初始权值和阈值，训练网络；Step 12: Map the global optimal value to the initial weight and threshold of the neural network, and train the network;

步骤13：将测试集I_test重复步骤2到步骤5的处理，将测试样本数据输入到已训练好的BP神经网络中进行测试，根据BP神经网络的输出结果，计算识别率。Step 13: Repeat steps 2 to 5 for the test set I _test , input the test sample data into the trained BP neural network for testing, and calculate the recognition rate according to the output of the BP neural network.

Claims

1. a face recognition method based on local binary and particle swarm optimization BP neural network, is characterized in that, comprises the following steps:

(1) Randomly select a certain number of each type of face images in the known face library as a training sample set I _train ={X ₁ ,X ₂ ,...,X _j ,...,X _A }, where X _j is each training sample image, A is the number of training samples;

(2) Perform geometric normalization processing on each M×N pixel grayscale image in the training sample set, and normalize it into an H×H size image, denoted as I′ _train , where 0<H≤min( M,N);

(3) Utilize the local binary algorithm to extract the illumination invariant to the training set image I' _train after geometric normalization, remove the influence of illumination, and obtain the training set image I" _train after illumination processing;

(4) carry out weighted two-dimensional discrete wavelet transform to the image set I " _train after local binary algorithm processing in step (3), obtain transformed image set I _{train, DWT} ;

(5) Perform two-dimensional discrete cosine transform on each sample image in the image set I _{train and DWT} obtained in step (4), and obtain the transformation coefficient matrix Y={Y ₁ , Y ₂ ,...,Y _h ,... ., Y _A }, where Y _h is the transform coefficient vector obtained after each sample image undergoes two-dimensional discrete cosine transform, and then expands each vector in the transform coefficient matrix Y by zigzag scanning, and finally extracts each Expand the principal components of the vector to form the optimal eigenvector E;

(6) Neural network parameters are set to determine the topology of the BP neural network, the number of input layer nodes Q, the number of hidden layer nodes W, the number of output layer nodes Z, and the activation function Sigmoid function;

(7) Optimizing the weight and threshold of BP neural network through particle swarm algorithm;

(8) The global optimal value that step (7) obtains is mapped to initial weight and threshold of neural network, trains BP neural network;

(9) use the remaining images in the face bank as the test sample set _Itest , repeat the process from step (2) to step (5), and then input the test sample data to the trained one obtained by step (8) Test in BP neural network and calculate the recognition rate.

2. a kind of face recognition method based on local binary value and particle swarm optimization BP neural network according to claim 1, is characterized in that, described step (3) is specially:

First, divide each image in the I′ _train into n×n small blocks, and use the mean value of the gray value of n×n pixels in each small block as a pixel value of the new image to obtain the block training set image I _{train_block} ;

Then, use the local binary algorithm to describe the illumination invariant features of the face image. For any image I in I _{train_block} , I∈I _{train_block} , any point (x _c , y _c ) and uniformly distributed around the center point The local binary feature operator LBP of the P neighborhood points of is:

\{\begin{matrix} {LBP LBP}_{R R,, P P} = = {Σ Σ}_{q q = = 00}^{P P - - 11} s the s (({g g}_{q q} - - {g g}_{c c})) \cdot &Center Dot; 22^{q q},, \\ s the s ((x x)) = = \{\begin{matrix} 11,, x x &GreaterEqual; &Greater Equal; 00 \\ 00,, x x < < 00 \end{matrix} . . \end{matrix} - - - - - - ((11))

Among them, g _c represents the gray value of the center pixel point (x _c , y _c ) in the area, and g _q (q=0,1,...,P-1) represents the gray value evenly distributed at the center point (x c , y _c ). y _c ) is the gray value of P sampling points on the circle whose center is the radius R.

3. a kind of face recognition method based on local binary value and particle swarm optimization BP neural network according to claim 1, is characterized in that, described step (4) is specially:

Select any image F in the image set I″ _train , and perform two-dimensional discrete wavelet transform on it to obtain images in the four directions of LL, LH, HL, and HH, which are denoted as F _LL , _FLH , F _HL , F _HH , then the weighted image F′ is:

F'＝a ₀ F _LL +a ₁ F _LH +a ₂ F _HL (2)

Among them, a ₀ , a ₁ , and a ₂ are weighting coefficients, and the constraint condition is a ₀ +a ₁ +a ₂ =1; the low-frequency component F _LL is a smooth image of the original image, which maintains most of the information of the original image; F _LH The component maintains the vertical edge details of the original image; the F _HL component maintains the horizontal edge details of the original image; the F _HH component maintains the edge details in the diagonal direction of the original image.

4. a kind of face recognition method based on local binary value and particle swarm optimization BP neural network according to claim 1, is characterized in that, described step (7) is specially:

a. Set particle swarm parameters, including particle population size B, dimension D, maximum number of iterations T _max , learning factors c ₁ and c ₂ , inertia weight ω, maximum velocity υ _max , maximum position x _max , minimum expected error ε, Randomly generate the initial velocity and position of the particles within the allowable range;

b. Calculate the fitness J _a of each particle;

c. Determine the individual optimal extremum and the global optimal extremum of each particle according to the fitness J _a , compare the current fitness value with the historical optimal fitness of each particle, and the one with the smaller fitness is the individual optimal extremum The value P _best is compared with the optimal fitness of the entire population, and the one with the smaller fitness is taken as the global optimal extreme value G _best ;

d. Update the velocity and position of each particle, and consider whether the updated velocity and position are within the limits;

{&upsi; &upsi;}_{id id}^{T T + + 11} = = {ω&upsi; ω&upsi;}_{id id}^{T T} + + {c c}_{11} {r r}_{11} (({p p}_{id id}^{T T} - - {x x}_{id id}^{T T})) + + {c c}_{22} {r r}_{22} (({p p}_{gd gd}^{T T} - - {x x}_{id id}^{T T})) - - - - - - ((33))

{x x}_{id id}^{T T + + 11} = = {x x}_{id id}^{T T} + + {&upsi; &upsi;}_{id id}^{T T} - - - - - - ((44))

Where, i=1,2,...,B, d=1,2,...,D, is the d-th dimension component of the velocity vector of particle i in the T iteration; is the d-th dimension component of the position vector of particle i in the T-th iteration; is the d-th dimension component of the current best position P _best of particle i; is the d-th dimension component of the best position G _best of the current group; r ₁ and r ₂ are random numbers that obey the uniform distribution of [0, 1];

Constraint 1: If

{&upsi;}_{id}^{T + 1} > {&upsi;}_{\max},

but

{&upsi;}_{id}^{T + 1} = {&upsi;}_{\max};

like

{&upsi;}_{id}^{T + 1} < - {&upsi;}_{\max},

but

{&upsi;}_{id}^{T + 1} = - {&upsi;}_{\max};

Constraint 2: If

x_{id}^{T + 1} > x_{\max},

but

x_{id}^{T + 1} = x_{\max};

like

x_{id}^{T + 1} < - x_{\max},

but

x_{id}^{T + 1} = - x_{\max};

e. Compare whether the number of iterations reaches the maximum T _max or whether the mean square error meets the precision ε requirement. If so, the algorithm converges, and record the individual optimal value P _best and the global optimal value G _best of the last iteration; otherwise, return to the step c.

5. a kind of face recognition method based on local binary value and particle swarm optimization BP neural network according to claim 4, is characterized in that, the computing method of described step b is specifically:

{J J}_{a a} = = \frac{11}{22 A A} {Σ Σ}_{k k = = 11}^{A A} {Σ Σ}_{t t = = 11}^{Z Z} {(({y the y}_{k k}^{t t} - - {c c}_{t t}^{k k}))}^{22} - - - - - - ((55))

Among them, a=1,2,...,B; is the expected output value of the tth network output neuron of the kth sample; is the actual output value of the tth network output neuron of the kth sample.