KR101658528B1 - NIGHT VISION FACE RECOGNITION METHOD USING 2-Directional 2-Dimensional Principal Component Analysis ALGORITHM AND Polynomial-based Radial Basis Function Neural Networks - Google Patents

Abstract

The present invention relates to a method of recognizing a face in an environment without illumination, comprising the steps of: (1) acquiring an image using a night vision camera; (2) a data preprocessing step of detecting a face region in the image data obtained in the step (1) and removing a disturbance value; (3) The data preprocessed in the above step (2) is subjected to a two-dimensional 2D principal component analysis (hereinafter referred to as '2D (2D) PCA' Reducing the data dimension while using both directions intact; (4) Polynomial-based Radial Basis Function Neural Networks (hereinafter referred to as " pRBFNNs ") to recognize faces in the image data with reduced data dimensions And is characterized by its constitution.
According to the two-way two-dimensional principal component analysis algorithm proposed by the present invention and the nighttime facial recognition method using the optimal polynomial radial basis function based neural network, it is possible to recognize the face even in the environment without illumination by acquiring the image through the night vision camera Do.
Further, according to the present invention, the face area is detected in the obtained image data, the disturbance value is removed, and then the two-dimensional two-dimensional principal component analysis algorithm is used to reduce the data dimension while using the two directions of the two- And the recognition rate can be improved.
In addition, according to the present invention, a polynomial radial basis function based neural network is used, but a fuzzy C-means (FCM) clustering algorithm is used, and the degree of polynomial, the number of clusters, By optimizing, the linear decision boundary in the output space can be represented by the nonlinear decision boundary, and the recognition performance is improved by fast learning convergence in the optimized parameter.

Description

FIELD OF THE INVENTION [0001] The present invention relates to a two-way two-dimensional principal component analysis algorithm and a nighttime facial recognition method using a polynomial radial basis function based neural network. [0002]

More particularly, the present invention relates to a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using a polynomial radial basis function-based neural network.

Recently, a biometric information authentication system for recognizing and authenticating biometric information such as a fingerprint and an iris face has been developed. Among them, face recognition does not require contact unlike other biometric information recognition methods, thereby giving convenience to the user. Conventionally, an image is acquired by a CCD camera (see Patent Application No. 10-2006-0119515, etc.). However, since the illuminance is low at nighttime without illumination, recognition is difficult to perform.

In addition, for facial recognition, it is necessary to recognize facial patterns to extract and identify features. Among various methods for pattern recognition, a neural network based classifier is one of the most used classifiers in pattern recognition systems of various fields because of its superior ability of learning ability and generalization ability. The neural network-based classifier is a multi-layered structure with an input layer, a plurality of hidden layers, and an output layer, and is learned using a slope descent method. Many classifiers based on existing neural networks have connection weights that are made up of constant terms, and connection weights composed of constant terms are expressed as linear discernment functions in the output space through linear combination with the activation function of hidden neurons. 3, pp. 697-710 (2002), "Magnetic Resonance Imaging (MRI), SQ Wu, JW Lu, and HL Toh," Face Recognition with RBF Neural Networks, "IEEE Trans. Neural Networks, Vol. .). This may cause the network to create a linear decision boundary (hyperplane) within the output space and have a linear characteristic.

On the other hand, principal component analysis is based on the statistical characteristics of vector representation, and expresses M vectors of statistically changed N dimensions as eigenvectors by covariance matrix. The principal component analysis method is widely used as a practical method for simply expressing the dimensions of different spaces. However, in the pattern recognition of facial data composed of high dimensional data, conversion in one direction has a problem of lowering the recognition rate.

The present invention has been proposed in order to solve the above-mentioned problems of the previously proposed methods. By acquiring an image through a night vision camera, a two-way two-dimensional principal component analysis algorithm capable of recognizing faces even in a light- The object of the present invention is to provide a nighttime face recognition method using a polynomial radial basis function based neural network.

In addition, the present invention reduces the data size while using two directions of the two-dimensional image intact by using the two-way two-dimensional principal component analysis algorithm after detecting the face region in the obtained image data and removing the disturbance value, Another object of the present invention is to provide a two-way two-dimensional principal component analysis algorithm with improved recognition rate and a nighttime face recognition method using an optimal polynomial radial basis function-based neural network.

In addition, the present invention uses a polynomial radial basis function-based neural network, uses a FCM (Fuzzy C-means) clustering algorithm, optimizes the degree of polynomials, the number of clusters and the fuzzy coefficient using a difference evolution algorithm Two-way two-dimensional principal component analysis algorithm and optimal polynomial radial basis function-based neural network, in which the linear decision boundary in the output space can be represented by a nonlinear decision boundary and the recognition performance is improved by fast learning convergence in the optimized parameter Another object of the present invention is to provide a nighttime facial recognition method.

According to an aspect of the present invention, there is provided a nighttime face recognition method using a two-way two-dimensional principal component analysis algorithm and an optimal polynomial radial basis function based neural network,

A method of recognizing a face in an environment without illumination,

(1) acquiring an image using a night vision camera;

(2) a data preprocessing step of detecting a face region in the image data obtained in the step (1) and removing a disturbance value;

(3) The data preprocessed in the above step (2) is subjected to a two-dimensional 2D principal component analysis (hereinafter referred to as '2D (2D) PCA' Reducing the data dimension while using both directions intact; And

(4) a step of recognizing a face in the image data in which the data dimension is reduced by using a polynomial-based radial basis functional neural network (hereinafter referred to as " pRBFNNs " It is a feature of the configuration.

Preferably, the step (2)

(2-1) detecting a face region using a Haar-like feauture and an Ada-boost algorithm; And

(2-2) removing the disturbance value using a histogram equalization technique.

More preferably, the step (2-2)

(2-2-1) generating a histogram of the face region image detected in the step (2-1);

(2-2-2) normalizing the histogram generated in the step (2-2-1) using the following equation, calculating a normalized cumulative value, and deriving a correction value; And

(G: image brightness, N: image size, H (i): cumulative normalization value, h (i): correction value)

(2-2-3) mapping a new output applying the correction value corrected in the step (2-2-2) to the pixel position of the face region image detected in the step (2-1) have.

Preferably, the step (2)

ASM (Active Shape Model) can be used to separate face area and background.

Preferably, the step (3)

(3-1) deriving an average image value of samples of a learning image represented by a matrix;

(3-2) deriving a covariance matrix of the image sequence of the learning image samples using the average image value derived in the step (3-1);

(3-3) deriving an eigenvalue of the covariance matrix of the image sequence derived in the step (3-2) and an eigenvector corresponding to the eigenvalue by analyzing the eigenvalue;

(3-4) For each eigenvalue derived in step (3-3), d eigenvalues are selected in descending order of eigenvalues, a transformation matrix having an eigenvector corresponding to the selected eigenvalue is generated, and data d Deriving a vector of image columns with reduced dimensions;

(3-5) deriving a covariance matrix of the image row of the learning image samples using the average image value derived in the step (3-1);

(3-6) deriving an eigenvalue of the covariance matrix of the image row derived in the step (3-5) and an eigenvector corresponding thereto, through eigenvalue analysis;

(3-7) For each eigenvalue derived in step (3-6), d eigenvalues are selected in descending order of eigenvalues, a transformation matrix having an eigenvector corresponding to the selected eigenvalue is generated, and data d Deriving a vector of image rows with reduced dimensions; And

(3-8) deriving the entire image for face recognition using the image column and the vector of the image row generated in the step (3-4) and the step (3-7).

Preferably, in said step (4), said pRBFNNs comprise

(Fuzzy C-means) clustering algorithm is used as the active function of the conditional part, and the connection weight of the conclusion part can be extended to a polynomial function.

More preferably, the step (4)

And optimizing the degree of the polynomial, the number of clusters and the fuzzy coefficient using a Differential Evolution (DE) algorithm.

According to the two-way two-dimensional principal component analysis algorithm proposed by the present invention and the nighttime facial recognition method using the optimal polynomial radial basis function based neural network, it is possible to recognize the face even in the environment without illumination by acquiring the image through the night vision camera Do.

Further, according to the present invention, the face area is detected in the obtained image data, the disturbance value is removed, and then the two-dimensional two-dimensional principal component analysis algorithm is used to reduce the data dimension while using the two directions of the two- And the recognition rate can be improved.

In addition, according to the present invention, a polynomial radial basis function based neural network is used, but a fuzzy C-means (FCM) clustering algorithm is used, and the degree of polynomial, the number of clusters, By optimizing, the linear decision boundary in the output space can be represented by the nonlinear decision boundary, and the recognition performance is improved by fast learning convergence in the optimized parameter.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flowchart illustrating a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using an optimal polynomial radial basis function-based neural network according to an embodiment of the present invention.
FIG. 2 illustrates a concrete flow of step S200 in a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using an optimal polynomial radial basis function-based neural network according to an embodiment of the present invention.
FIG. 3 is a diagram showing a feature extraction method of the Hal-like feature. FIG.
4 is a diagram illustrating a process of extracting an integral image feature.
FIG. 5 illustrates a histogram smoothing process. FIG.
FIG. 6 illustrates a specific flow of step S300 in a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using an optimal polynomial radial basis function based neural network according to an embodiment of the present invention.
FIG. 7 is a diagram showing a process of dimensionally reducing by using a (2D) PCA; FIG.
8 shows a conventional radial basis functional neural network;
9 is a diagram illustrating a polynomial radial basis function based neural network proposed in the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. In the following detailed description of the preferred embodiments of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. The same or similar reference numerals are used throughout the drawings for portions having similar functions and functions.

In addition, in the entire specification, when a part is referred to as being 'connected' to another part, it may be referred to as 'indirectly connected' not only with 'directly connected' . Also, to "include" an element means that it may include other elements, rather than excluding other elements, unless specifically stated otherwise.

FIG. 1 is a flowchart illustrating a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using an optimal polynomial radial basis function-based neural network according to an embodiment of the present invention. As shown in FIG. 1, a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using an optimal polynomial radial basis function-based neural network according to an embodiment of the present invention are implemented using a night vision camera A data preprocessing step (S200) for performing a process of acquiring an image (S100), detecting a face area in the image data acquired in step S100 and removing a disturbance value, a preprocessing step (S200) (S300) of reducing a data dimension while using two directions of a two-dimensional image intact by using a 2-dimensional principal component analysis (hereinafter referred to as '2D (2D) PCA') algorithm, and Using Polynomial-based Radial Basis Function Neural Networks (pRBFNNs), the data dimension is reduced to a reduced image size Including a step (S400) of recognizing a face from emitter it can be implemented. Step S100 is performed through the night vision camera and steps S200 to S400 are performed by the user terminal or the night preprocessing module (step S200), the data dimension reduction module (step S300) and the face recognition module ), Respectively.

In step S100, an image can be acquired using a night vision camera. Night Vision is a method of using thermal imaging technology, in which an infrared camera senses the thermal energy of an object and reproduces it as a monochrome image. The conventional two - dimensional face recognition system using the CCD camera has difficulty in recognition because it is difficult to extract faces at night. According to the present invention, a night vision camera capable of acquiring a nighttime facial image is used in place of the conventional CCD camera to compensate for such a disadvantage, so that an image similar to a daytime facial image can be obtained.

In step S200, data preprocessing such as a process of detecting a face area and removing a disturbance value from the image data acquired in step S100 may be performed. FIG. 2 is a diagram illustrating a specific flow of step S200 in a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using an optimal polynomial radial basis function-based neural network according to an embodiment of the present invention. 2, in operation S200, a face region is detected using a Haar-like feature and an Ada-boost algorithm S210) and removing a disturbance value using a histogram equalization technique (S220). The AdaBoost algorithm detects only the face region by analyzing the feature data of the background and face in the night image, and the histogram smoothing can improve the quality in the night image and compensate the distortion caused by illumination.

Specifically, step S210 is a process of detecting a face region using a Hal-like feature and an Adaboost algorithm. First, the face region can be detected through the Hal-like feature and the integral image. The Hal-like property is calculated by P. Viola and M. Jones using simple sums. Because it can operate at high speed, it is suitable for face recognizer that recognizes in real time. 3 is a diagram illustrating a feature extraction method of the Hal-like feature. As shown in FIG. 3, an edge feature, a line feature, and diagonal features can be extracted, and the sum of the areas can be calculated by adding pixel values in the feature region And the difference of the sum of each area is calculated to extract the Hal-like characteristic. 4 is a diagram illustrating an extraction process of an integral image feature. As shown in FIG. 4, it is possible to quickly calculate using all the accumulated values for all the pixels in the image. The AdaBoost algorithm has a hierarchical structure approaching step by step with algorithms proposed by Y. Freund and R.Shapire. This structure linearly combines the n weak classifiers by eliminating the error step by step and using the probability distribution. Finally, a strong classifier with high detection capability is created.

In step S220, the disturbance value may be removed using the histogram smoothing technique. The disturbance value may exist in the face image data acquired using the night vision camera. In the present invention, a histogram can be used to eliminate such disturbance values. FIG. 5 is a diagram illustrating a histogram smoothing process. As shown in FIG. 5, in order to minimize the disturbance, the distribution values of the business cards of the obtained face image are made uniform, so that the histogram biased toward one side can be made to have a uniform distribution.

Preferably, step S220 includes generating a histogram of the face area image detected in step S210 (S221), normalizing the histogram generated in step S221 using the following equation (1), calculating a normalized cumulative value, And a step S223 of mapping a new output applying the correction value corrected in step S222 to the pixel position of the face area image detected in step S210.

(G: image brightness, N: image size, H (i): cumulative normalization value, h (i): correction value)

Meanwhile, according to another embodiment of the present invention, in step S200, an ASM (Active Shape Model) can be used to separate the background and the face. ASM is one of the methods that are used for extracting feature points of face by extracting feature points using statistical model. The ASM uses the feature points from the training data to obtain the average shape to generate the shape model. After resizing, rotating and averaging the shape to match the average again, the average of the deformed shapes is obtained again. Repeat the process until the difference between the previous average and the newly obtained average is no more than a certain value. When a new face image comes in using the average shape thus obtained, the face region is detected by comparing with the average shape. This is an embodiment of the face area detection method and is not limited thereto.

In step S300, the data preprocessed in step S200 can be reduced in size using the bi-directional two-dimensional principal component analysis algorithm. Principal Component Analysis (PCA) is a method based on the statistical properties of vector representation, and expresses M vectors of N dimensions with statistical changes by eigenvectors by covariance matrix. Principal component analysis is a practical method of simply expressing the dimensions of different spaces. Since facial image data is high-dimensional data, there is a disadvantage in that the recognition speed is slowed down due to a large amount of data when used for pattern recognition. In the present invention, the dimension of the data is reduced using the (2D) PCA algorithm to improve the recognition speed. The general PCA method reduces the dimension in one direction. In the present invention, however, it is possible to reduce the dimension using two directions of the two-dimensional image using the (2D) PCA algorithm. By adopting this method, the recognition rate can be further increased as compared with the PCA method.

FIG. 6 is a diagram illustrating a concrete flow of step S300 in a two-way two-dimensional principal component analysis algorithm and a nighttime face recognition method using an optimal polynomial radial basis function-based neural network according to an embodiment of the present invention. As shown in FIG. 6, according to an embodiment of the present invention, step S300 includes deriving an average image value of samples of a learning image represented by a matrix (S310), calculating an average image value derived in step S310 Deriving a covariance matrix of the image sequence of the learning image samples at step S320, deriving eigenvalues of the covariance matrix of the image sequence derived at step S320 and the eigenvectors corresponding thereto at step S330 through eigenvalue analysis at step S330, Selecting d eigenvalues in descending order of the eigenvalues with respect to the eigenvalues derived from the eigenvalues derived from the eigenvalues obtained in step S340 , Deriving a covariance matrix of the image row of the learning image samples using the average image value derived in step S310 (S350) (S360) of deriving the eigenvalues of the covariance matrix of the image row derived in step S350 and the eigenvectors corresponding thereto, d eigenvalues are selected in descending order of the eigenvalues for the eigenvalues derived in step S360, Generating a transform matrix having an eigenvector corresponding to the selected eigenvalue, deriving a vector of image rows in which the data dimension is reduced by d (S370), and deriving a vector of image columns and image rows generated in steps S340 and S370 And deriving an entire image for face recognition using the captured image (S380). Step S310 to step S380 are steps of (2D) PCA, as shown in FIG. 6, steps S320 to S340 are steps of deriving vectors of image columns, and steps S350 to S370 are steps of deriving vectors of image rows , The vector of the image sequence may be derived first, or the vector of the image sequence may be derived first. FIG. 7 is a view showing a process of dimensionally reducing by (2D) PCA. As shown in Fig. 7, when the (2D) PCA is used, the dimension is reduced using the two directions of the two-dimensional image intact. Hereinafter, steps S310 to S380 will be described in more detail.

행렬(n = 1, 2, …, M)로, 학습 이미지의 샘플들의 평균 이미지를

로 나타낸다고 할 때, 행렬로 표현되는 학습 이미지의 샘플들의 평균 이미지값(

)은, 하기 수학식 2로 표현될 수 있다(단계 S310).The number of total learning image data is M,

With the matrix (n = 1, 2, ..., M), the average image of samples of the training image

The average image value of the samples of the learning image represented by the matrix (

) Can be expressed by the following equation (2) (step S310).

The covariance matrix RG _t of the image sequence of the learning image sets can be expressed by the following equation (3) (step S320).

의 고유값

과 이에 대응하는 고유벡터

을 계산할 수 있다(하기 수학식 4 참고, 단계 S330).Through eigenvalue analysis

Eigenvalue of

And the corresponding eigenvectors

(See Equation 4 below, step S330).

(T: transpose, λ: eigenvalue, u: eigenvector r: number of columns)

에 대해 고유값이 큰 순서대로 d개의 고유치

를 선택하고, 선택한 고유값에 대응되는 고유벡터를 가지는 변환행렬

을 생성할 수 있다(단계 S340).

는, d만큼 차원을 줄인 이미지 열의 벡터를 의미한다.
The eigenvalues obtained in step S330

D < / RTI > eigenvalues < RTI ID = 0.0 >

And a transformation matrix having an eigenvector corresponding to the selected eigenvalue

(Step S340).

Means a vector of image columns whose dimension is reduced by d.

On the other hand, the covariance matrix LG _t of the image rows of the learning image sets can be expressed by the following equation (5) (step S350).

의 고유값

과 이에 대응하는 고유벡터 을 계산할 수 있다(하기 수학식 6 참고, 단계 S360).Through eigenvalue analysis

Eigenvalue of

And the corresponding eigenvector can be calculated (see Equation (6), step S360).

(T: transpose, λ: eigenvalue, u: eigenvector, c: number of rows)

고유값이 큰 순서대로 d개의 고유치

를 선택하고, 선택한 고유값에 대응되는 고유벡터를 가지는 변환행렬

을 생성할 수 있다(단계 S370).

는, d만큼 차원을 줄인 이미지 행의 벡터를 의미한다.
The eigenvalues obtained in step S360

The d eigenvalues in descending order of eigenvalues

And a transformation matrix having an eigenvector corresponding to the selected eigenvalue

(Step S370).

Means a vector of image rows that have been reduced in dimension by d.

In step S380, the entire image for face recognition can be derived using the vector of the image sequence generated in step S340 and the vector of the image row generated in step S370. The equation for deriving the entire image for face recognition can be represented by the following equation (7).

In step S400, the face can be recognized in the image data in which the data dimension is reduced by using the polynomial radial basis function-based neural network (pRBFNNs). The pRBFNNs proposed in the present invention can operate as three functional modules in combination with existing neural networks. That is, the Fuzzy C-means (FCM) clustering algorithm is used as an active function of the conditional part, and the connection weight of the conclusion part is extended to a polynomial function .

Unlike the conventional RBFNN, the pRBFNNs proposed in the present invention are extended from the constant term to the polynomial of the concatenation weight. And the conditional activity function uses FCM clustering instead of the Gaussian function used in the conventional RBFNN. The pRBFNNs pattern classifier uses the preprocessed input image data in the learning and recognition phase. FIG. 8 is a diagram illustrating a conventional radial basis function network, and FIG. 9 is a diagram illustrating a polynomial radial basis function based neural network proposed in the present invention. As shown in FIG. 8, the conventional radial basis functional neural network is composed of an input layer, a hidden layer, and an output layer, and functions as a functional module separated into a conditional part, a conclusion part, and a reasoning part.

The conditional part means a hidden layer of a general neural network. The number of nodes in the hidden layer is determined by the user and there are as many active functions as the number of nodes. In general, the Gaussian function is used as the active function of the conditional part, and the constant part is used as the conclusion part. The final output of the model is expressed as a constant term of conditional and conclusion.

9, the condition part of the pRBFNNs pattern classifier proposed in the present invention uses a fuzzy C-means clustering algorithm instead of the Gaussian function used in the conventional RBFNN, and FCM clustering is included in each cluster The degree of membership of the data is output as a fuzzy set and represents a radial type of active function, so that the function of the Gaussian function can be substituted. In the conclusion part, it has one of the three forms of linear equation, quadratic equation and modified quadratic equation except for constant term as shown in Table 1 below.

Type Polynomial Type Linear

Quadratic

Modified
Quadratic

Clustering algorithms are algorithms used in the classification of data, which classify data by reference to similar patterns and forms in the data to define the belonging group of the object. In the present invention, the FCM clustering algorithm can be used to measure the degree of affiliation of each data for clustering on the basis of distance and to classify characteristics of data. FCM clustering is a data classification algorithm in which each data point belonging to one cluster is listed in order of the degree of belonging to the cluster. By using the FCM clustering algorithm instead of the Gaussian activation function in the conditional part of the pattern classifier, the characteristics of the input data can be better reflected. The number of clusters in the FCM clustering algorithm is equal to the number of hidden layer nodes, the value of the membership matrix composed of fuzzy sets is equal to the fitness value by the Gaussian activation function, and the fuzzification coefficient of FCM clustering is the same as the distribution constant of the Gaussian function. The process is as follows. First, the number of clusters and the fuzzification coefficient are selected, and the belonging matrix U (r) is initialized (see Equation (8)).

(U ^(r) : corresponding member matrix, c: number of clusters, n: number of data)

Secondly, the center value for each cluster is obtained based on the values (see Equation (9) and (10) below).

(i: cluster, n: number of data, j: input, v: cluster center)

(i: cluster, n: number of data, j: input, m: fuzzy coefficient, v: cluster center)

Third, the distance between each cluster center and the data is calculated to generate a new member matrix (see Equation (11) and (12) below).

(i: cluster, k: data, l: number of inputs, u ^{(r + 1} )

(i: cluster, k: data, c: number of clusters, m: fuzzification coefficient, u ^{(r +}

Fourth, if the error is within the allowable range, terminate; otherwise, go back to the second step (see Equation 13 below).

(U ^(r) : corresponding member matrix, U ^{(r + 1)} : next member matrix,: error rate)

The conclusion of the polynomial radial basis function neural network expresses each local region separated from the conditional part as a local regression model of the polynomial function and can form the rule after " then " in the following equation (14).

(A: membership function, f _ij (x): polynomial, i: rule, j: output neuron)

In f _ji (x) of equation (14) the subscript for the output neuron j (= 1, ..., s ) f i (x) is omitted is of one of the functions of the three types in the form of Equation 15 to 17 . In other words, the local session model is expressed as a constant term, a linear or quadratic expression.

(i: rule)

(i: rule, n: number of inputs)

(i: rule, n: number of inputs)

If the dimension of the input space is very large, the number of combinations of input variables increases with respect to the second term of the equation (17), so that the amount of computation increases. Therefore, the reduced amount of the quadratic function expressed by equation (18) is used to reduce the amount of computation.

(i: rule, n: number of inputs)

Weighted Least Square Estimation N (WLSE) can be used to estimate the polynomial parameter coefficients. WLSE is an algorithm for estimating the coefficients of regression polynomials and is similar to LSE. The LSE estimates the coefficients so that the sum of the squared errors is minimized, but the WLSE differs in that the weights of the squared errors are multiplied. Since LSE collects polynomial coefficients all at once, it performs global model learning, and when the number of input variables and membership functions increases, the number of fuzzy rules increases exponentially. However, the WLSE is independent of each rule and performs local learning. It has the advantage of improving the interpreting power for each local area, which can reduce the computational load of the computer and can be represented by different types of polynomials. The polynomial radial basis function based neural network proposed in the present invention can improve the performance as a classifier due to the nonlinear coupling between the output of the hidden layer neuron by extending the connection weights of the existing constant term into the first order and the second order.

The proposed polynomial-based RBFNNs structure can be understood as a local session model that expresses the local region as a polynomial and concludes the fuzzy space division through conditional FCM clustering as mentioned above. Inference section obtains final output of network by fuzzy reasoning based on "If-then" fuzzy rule. The input signals are summed by the inference neuron and the result is output to the final output of the output layer neuron (see equation (19)).

(j: data, c: number of rules)

Meanwhile, the step S400 may further include optimizing the degree of the polynomial, the number of clusters and the fuzzification coefficient using a differential evolution (DE) algorithm. The differential evolution algorithm was developed in the course of solving the interpolation problem of Chebychev polynomial curves using vector differences by Price and Storn. Differential evolution algorithms have the advantage that they have excellent convergence to the global optimization and are simple in structure, resulting in shorter computing time than other evolutionary algorithms. The operators used in the differential evolution algorithm mainly use the differences between randomly selected entities. We have optimized the degree of polynomials, the number of clusters and the fuzzy coefficients of the proposed model using a differential evolution algorithm. The procedure of the differential evolution algorithm is as follows.

First, the initial population can be created. Initialize μ objects with random values, and each entity can be composed of n objective variables (see Equation 20 below).

(u: denominator, n: objective variable of the object, t: household)

Second, the objective function of all entities in the group can be evaluated (see Equation 21 below).

(Φ: objective function to be minimized, u: one object of total objective variable)

Thirdly, a mating vector can be created by selecting an object for differential change with respect to all the objects (i = 1, ..., μ), and mated with the mating vector (see Equations 22 and 23 below).

(v: mating vector, i: order of vector for mating, a _r1 , a _r2 , a _r3 : target variable of selected entity for making mating vector)

(x ' _i : mated i-th object, v _i : i-th mating vector, a _i : i-th object)

Fourth, the objective function of all objects can be evaluated. Fifthly, the end condition is checked. If not satisfied, t = t + 1, and the process returns to the third step.

The present invention will be described in more detail with reference to the following experimental examples, but the present invention is not limited in any way by the following experimental examples.

Experimental Example  1. Face recognition rate comparison experiment

The data structure of this experiment consisted of 12 members of IC & CI Lab (Intelligent Control and Computer Intelligence Laboratory) of Suwon University and facial data of lab personnel consisted of 3 conditions. As a first condition, 12 face data were acquired using Nightvision in the situation where the light intensity was 0 ~ 5 Lux in the dark state without light. The second is the situation when there is a dim light. The illumination value is 5 ~ 20 Lux and the data is acquired. Finally, the situation is acquired during the day in the same way as we usually acquire. Respectively. In each situation, the data of 12 candidates were composed of a total of 360 sheets of 10 sheets (see Table 2).

Since the acquired data may contain a light source, it may be difficult to identify the image. Therefore, disturbance due to the light source is minimized by using histogram smoothing. We performed a preprocessing process to separate face and background through Harr-like feature and Ada-boost algorithm. And thus the face image obtained is an operation time of the step of reducing the high-dimensional data using a (2D) ² PCA algorithm to Rough because because this consists of high-dimensional data may cause a disadvantage that the time of the operation takes a long Respectively. In order to improve the recognition performance, the parameters are optimized by using the difference evolution algorithm as the optimization algorithm. The parameter values applied to the experiment are shown in Table 3 below.

Parameter Value No. of generations 60 No. of populations 40 Search range Polynomial type 2 to 4 No. of clusters 2 to 5 Fuzzification Coefficient 1.1 to 3.0

Using the reduced data, the recognition rate for daytime and nighttime, PCA and (2D) ² PCA was calculated through the pRBFNNs pattern classifier. Table 4 below shows the results of the recognition rate calculation.

As shown in Table 4, when the (2D) ² PCA method was used, the recognition rate was improved at a time as compared with the PCA method. According to the present invention, it is possible to acquire an image even at night using a night vision camera, a face image is detected using Ada-Boost algorithm, and image distortion can be minimized by using histogram smoothing. The high dimensional images were reduced to low dimensions using the (2D) ² PCA algorithm, face recognition was performed using the intelligent pattern classification model using pRBFNNs, and finally the parameters were optimized using a differential evolution algorithm. In addition, in order to evaluate the performance of the face recognition method proposed by the present invention, the actual face recognition system using IC & CI Lab data was designed, and it was confirmed that the face recognition rate significantly improved at night.

The present invention may be embodied in many other specific forms without departing from the spirit or essential characteristics of the invention.

S100: Step of acquiring an image using a night vision camera
S200: a data preprocessing step of detecting a face area in the image data acquired in step S100 and performing a process of removing a disturbance value
S210: Detecting the face region using Haar-like feaure and Ada-boost algorithm
S220: removing the disturbance value using a histogram equalization technique
S221: a step of generating a histogram of the face area image detected in step S210
S222: normalizing the histogram generated in step S221, calculating a normalized cumulative value, and deriving a correction value
S223: mapping a new output to which the correction value corrected in step S222 is applied to the pixel position of the face region image detected in step S210
S300: The step of reducing the data dimension while using the two directions of the two-dimensional image intact using the bi-directional two-dimensional principal component analysis algorithm at step S200
S310: deriving an average image value of samples of the training image represented by a matrix;
S320: deriving a covariance matrix of the image sequence of the learning image samples using the average image value derived in step S310;
S330: deriving an eigenvalue of the covariance matrix of the image sequence derived in step S320 and an eigenvector corresponding thereto from the eigenvalue analysis;
Step S340: d eigenvalues are selected in descending order of the eigenvalues for the eigenvalues derived in step S330, a transformation matrix having an eigenvector corresponding to the selected eigenvalues is generated, and a vector of the image sequence in which the data dimension is reduced by d is derived ;
S350: deriving a covariance matrix of image rows of the learning image samples using the average image value derived in step S310;
S360: deriving an eigenvalue of the covariance matrix of the image row derived in step S350 and an eigenvector corresponding thereto from the eigenvalue analysis;
In step S370, d eigenvalues are selected in descending order of the eigenvalues for the eigenvalues derived in step S360, a transformation matrix having the eigenvectors corresponding to the selected eigenvalues is generated, and a vector of the image row in which the data dimension is reduced by d ; And
S380: deriving the entire image for face recognition using the image column and the image row vector generated in steps S340 and S370
S400: Step of recognizing a face from image data whose data dimension is reduced by using a polynomial radial basis function-based neural network

Claims (7)

A method of recognizing a face in an environment without illumination,
(1) acquiring an image using a night vision camera;
(2) a data preprocessing step of detecting a face region in the image data obtained in the step (1) and removing a disturbance value;
(3) The data preprocessed in the above step (2) is subjected to two-dimensional two-dimensional principal component analysis (hereinafter referred to as '2D (2D) PCA' Reducing the data dimension while using the direction as it is; And
(4) recognizing a face from the reduced image data by using a polynomial-based radial basis functional neural network (hereinafter, referred to as 'pRBFNNs'),
The step (3)
(3-1) deriving an average image value of samples of a learning image represented by a matrix;
(3-2) deriving a covariance matrix of the image sequence of the learning image samples using the average image value derived in the step (3-1);
(3-3) deriving an eigenvalue of the covariance matrix of the image sequence derived in the step (3-2) and an eigenvector corresponding to the eigenvalue by analyzing the eigenvalue;
(3-4) For each eigenvalue derived in step (3-3), d eigenvalues are selected in descending order of eigenvalues, a transformation matrix having an eigenvector corresponding to the selected eigenvalue is generated, and data d Deriving a vector of image columns with reduced dimensions;
(3-5) deriving a covariance matrix of the image row of the learning image samples using the average image value derived in the step (3-1);
(3-6) deriving an eigenvalue of the covariance matrix of the image row derived in the step (3-5) and an eigenvector corresponding thereto, through eigenvalue analysis;
(3-7) For each eigenvalue derived in step (3-6), d eigenvalues are selected in descending order of eigenvalues, a transformation matrix having an eigenvector corresponding to the selected eigenvalue is generated, and data d Deriving a vector of image rows with reduced dimensions; And
(3-8) deriving an entire image for the face recognition using the image column and the vector of the image row generated in the step (3-4) and the step (3-7)
Wherein the pRBFNNs in step (4)
(Fuzzy C-means) clustering algorithm is used as an active function of the conditional part, and the connection weights of the conclusion part are extended to a polynomial,
The step (4)
Further comprising optimizing the order of the polynomial, the number of clusters, and the fuzzy coefficient using a Differential Evolution (DE) algorithm,
Wherein the step of performing the differential evolution algorithm comprises:
(a) generating an initial group in which each entity consists of n objective variables and initializes u entities with random values;
(b) evaluating an objective function of all entities in the population generated in the step (a) using the following equation;

(Φ: objective function to be minimized, u: one object of total objective variable, t: household)
(c) selecting an entity for a difference change for all the entities (i = 1, ..., μ), creating a mating vector using the following mathematical expression and crossing the mating vector with the mating object;

(v: mating vector, i: order of vector for mating, ar1, ar2, ar3: objective variable of selected object to make mating vector)

(x ' _i : crossed i-th object, vi: i-th crossing vector, ai: i-th object)
(d) evaluating an objective function of all of the entities; and
(e) confirming the termination condition and if not satisfied, returning to step (c) with t = t + 1, characterized in that the two-way two-dimensional principal component analysis algorithm and the polynomial radial basis function based nerve Night Face Recognition Method Using Network.
2. The method of claim 1, wherein step (2)
(2-1) detecting a face region using a Haar-like feauture and an Ada-boost algorithm; And
(2-2) a two-way two-dimensional principal component analysis algorithm and a polynomial radial basis function-based neural network using a histogram equalization technique to remove a disturbance value .
3. The method according to claim 2, wherein the step (2-2)
(2-2-1) generating a histogram of the face region image detected in the step (2-1);
(2-2-2) normalizing the histogram generated in the step (2-2-1) using the following equation, calculating a normalized cumulative value, and deriving a correction value; And

(G: image brightness, N: image size, H (i): cumulative normalization value, h (i): correction value)
(2-2-3) mapping a new output applying the correction value corrected in the step (2-2-2) to the pixel position of the face region image detected in the step (2-1) A two - way two - dimensional principal component analysis algorithm and a nighttime face recognition method using a polynomial radial basis function neural network.
2. The method of claim 1, wherein step (2)
A two-way two-dimensional principal component analysis algorithm and a nighttime facial recognition method using a polynomial radial basis function-based neural network, characterized in that a face region is detected using an ASM (Active Shape Model).
delete delete delete

Images