KR100545559B1 - Method for recognizing a face using Haar-like feature/LDA and apparatus thereof - Google Patents

Abstract

The present invention relates to a face recognition apparatus and method, and more particularly, to a face recognition apparatus and method using a similar feature / linear discriminant analysis. According to the present invention, a face recognition apparatus using a haar feature / linear discriminant analysis is a device for recognizing a face using a haar-like feature and linear discriminant analysis (LDA). A detector for generating face image information, a partial image information extractor for extracting partial image information to be analyzed from the face image information, a like feature extractor for extracting like feature information from the partial image information, and the extracted A linear discriminant analyzing unit extracting final feature information of the face by performing linear discriminant analysis based on like feature information, a feature information comparison and determination unit comparing the final feature information with previously stored final feature information to determine whether the same person is the same; A storage unit storing final feature information corresponding to a face of at least one or more recognition objects; T determination may include an output section for outputting the result of the determination portion.

Haar-like feature, linear discriminant analysis (LDA), face, recognition.

Description

Method for recognizing a face using Haar-like feature / LDA and apparatus

1 shows a prototype of Haar-like feature information.

2 is a flow chart illustrating a method of extracting feature information performed by AdaBoost.

3 is a system configuration including a face recognition device according to a preferred embodiment of the present invention.

Figure 4 is a block diagram of a face recognition device according to an embodiment of the present invention.

5 is a diagram for extracting partial image information according to a preferred embodiment of the present invention.

6 is a diagram for extracting feature information from partial image information according to a preferred embodiment of the present invention.

7 is a data diagram used in a face recognition experiment according to a preferred embodiment of the present invention.

8 is a diagram illustrating a recognition rate extracted in a face recognition experiment according to a preferred embodiment of the present invention.

9 is a diagram illustrating a recognition rate extracted in a face recognition experiment according to another exemplary embodiment of the present invention.

10 is a flowchart illustrating a face recognition method according to a first preferred embodiment of the present invention.

11 is a flowchart illustrating a face recognition method according to a second preferred embodiment of the present invention.

<Explanation of symbols for main parts of the drawings>

410: input unit

421: Partial image information extraction unit

423: Like feature extractor

425: Adamboost extracting unit

427: linear discriminant analysis unit

430: feature information comparison determination unit

440: storage unit

450: output unit

The present invention relates to a face recognition apparatus and a method, and more particularly, to a face recognition apparatus and method using a haar-like feature and LDA.

In image processing and pattern recognition, a human face becomes an important factor for visual discrimination and identification, and image processing methods are classified into a face detection process and a face recognition process. First, face detection is a process of pre-processing face recognition as distinguishing a face from a background by grasping a location of a face, ie, an area where a face is displayed, in an image signal. Face detection can be applied to various fields such as biometrics, video conferencing, unattended surveillance, human computer interaction (HCI), and the like. Next, face recognition is a process in which a computer recognizes whether the same person is the closest to the way a human recognizes a person along with voice recognition technology.

Face recognition methods include geometric feature-based, template-based, and model-based methods. Statistical methods such as LDA (Linear Discriminant Analysis) are mainly used. The technique using the LDA technique is a method of recognizing a face by projecting the face images onto a base plane that can be best separated. That is, by using the information about the change existing on the face of one person and the information about the change existing on the face with the other person, each person makes a discriminating expression that can express their information well. The LDA technique usually combines the PCA to reduce the number of feature values and increase the robustness rather than directly accessing the face image value. PCA (Principal Component Analysis) is a statistical method applied to the face recognition field, and it is a method of analyzing common characteristics that are most important in many face images and defining them as feature information representing each face.

However, in case of using PCA, speed is slowed down because excessive computation is required in extracting the basis vector for face recognition. In recent years, face recognition technology often requires real-time processing, and for this purpose, a face recognition method requiring a small amount of computation while expressing an input pattern well is required. In addition, there is a need for a method for recognizing a face regardless of the influence of ambient lighting.

The Harar-like feature Viola first used in face detection in February 2001 is used as a simple yet efficient recognizer. It is characterized in that the operation is performed using the feature information without performing the operation with the pixel itself. In other words, one feature information is obtained by adding a value of pixels in an individual area to obtain a simple area sum and multiplying the values by weights to calculate the sum.

1 is a diagram showing a prototype of haar-like feature information. Referring to FIG. 1, prototypes 110, 115, 120, and 125 representing feature information of edges in an image, prototypes 130, 135, 140, 145, 150, 155, 160, representing feature information of lines in an image, 165 is shown.

Prototype 110 representing the feature information of the corner may be 43,200 cases when the size is expanded by maintaining the same ratio as other positions in the 24 * 24 window. Characteristic information can be obtained by the following formula.

가중치

weight

RecSum (r _i ): width of rectangle r _i

This is an example of an extended set of Haar-like Features for Rapid Object, published by Rainer Lienhart and Jochen Mayd in September 2002. The detailed description of this is omitted since it is described in the article "Detection".

2 is a flowchart illustrating a method of extracting feature information performed by AdaBoost.

Boost is an algorithm that combines weak selection criteria to produce high accuracy selection criteria when deriving results from data. The AdaBoost learning algorithm is the best known Boost algorithm and has the advantage of being simple and efficient. Referring to FIG. 2, the Adaboost learning algorithm groups the extracted feature information 210 into stages 220, 230, and 240. Features that are determined to be inadequate at each step are classified into data 250 that does not meet the selection criteria, and feature information critical to face recognition is extracted during each step.

Hereinafter, an apparatus for recognizing a face using the above-mentioned like feature information and the Adaboost learning algorithm will be described.

Accordingly, in order to solve the above problems, an object of the present invention is to provide a face recognition apparatus and method using the like feature and linear discriminant analysis (LDA) to reduce the time required for face recognition.

Another object of the present invention is to provide a face recognition apparatus and method for reducing the time required for face recognition while minimizing the influence of ambient lighting.

Another object of the present invention is to increase the face recognition rate by setting a specific part in a face image and generating feature information on the part, and not extracting information on a part having a lot of variability.

Other objects of the present invention will become more apparent through the preferred embodiments described below.

In order to achieve the above objects, according to an aspect of the present invention, a device for recognizing a face of a person using a haar-like feature and linear discriminant analysis (LDA) can be provided.

According to a preferred embodiment, an apparatus for recognizing a face of a person using haar-like features and linear discriminant analysis (LDA) may include a detector configured to detect a face of a person and generate face image information. A partial image information extracting unit for extracting partial image information to be analyzed from the image information, a like feature extracting unit for extracting like feature information from the partial image information, and linear discriminant analysis based on the extracted like feature information A linear discriminant analyzer configured to extract final feature information of the face, a feature information comparison determiner configured to compare the final feature information with previously stored final feature information, and determine whether the person is the same person; A storage unit for storing the final feature information, an output unit for outputting the determination result of the feature information comparison determination unit It may include.

Here, the detection unit may detect a face using the like feature information, and extract feature information of the face image information using an AdaBoost learning algorithm based on the extracted like feature information. The apparatus may further include a boost extracting unit, and the linear discrimination analyzing unit may perform linear discrimination analysis based on the feature information extracted by the adamboost extracting unit. In addition, the partial image information may be image information excluding a hair and a neck portion of a human face.

In addition, the partial image information extractor performs Fast Lighting Correction (FLC) on the extracted partial image information. The FLC normalizes the brightness component by subtracting the average from the brightness value of the image and dividing by the variance. The image of the extracted partial image information may be divided into left and right halves, and the image information of the left image and the image information of the right image may be overlapped with each other and multiplied by the weight of the information of the original image. The partial image information extractor may extract the partial image information based on the eye position of the face image.

In order to achieve the above objects, according to an aspect of the present invention, it is possible to propose a method for recognizing a face of a person using a haar-like feature and linear discriminant analysis (LDA).

According to a preferred embodiment, a method of recognizing a face of a person using haar-like features and linear discriminant analysis (LDA) may include detecting a face of a person and generating face image information. Extracting partial image information to be analyzed from the information, extracting like feature information from the partial image information, and performing linear discriminant analysis based on the extracted like feature information to obtain a final feature of the face image information The method may include extracting the information and comparing the extracted final feature information with previously stored final feature information to determine whether they are the same person.

Here, in the step of detecting the face of the person and generating face image information, the face may be detected using the like feature information, and using the AdaBoost learning algorithm based on the extracted like feature information. The method may further include extracting feature information of the face image information, and performing linear discriminant analysis based on the extracted feature information.

The FLC may further include performing fast lighting correction (FLC) on the extracted partial image information, wherein the FLC normalizes the brightness component by subtracting the average from the brightness value of the image and dividing by the variance. The method may further include dividing the extracted partial image information into left and right halves and overlapping the image information of the left image and the image information of the right image by multiplying the original image information by a weight.

Hereinafter, preferred embodiments of a face recognition apparatus and method using a similar feature / linear discriminant analysis according to the present invention will be described in detail with reference to the accompanying drawings. The same or corresponding components are given the same reference numerals and redundant description thereof will be omitted.

Overall composition

3 is a system configuration diagram including a face recognition apparatus according to a preferred embodiment of the present invention. Referring to FIG. 3, a face 310 of a person to be recognized, a face recognition device 320, a camera 330 for capturing a face image of a person, an output unit 340 indicating a recognition result, and a user button 350. This is shown.

The feature information of the face 310 of the person to be recognized is previously stored in a database interworking with the face recognition apparatus 320. When the feature information is extracted by the face recognition apparatus according to the present invention, the feature information is stored in advance. Compare to and determine whether they are the same. The camera 330 photographs the face 310 of a person and provides image information to detect and recognize the face. The output unit 340 displays the recognition result, and preferably outputs the face image received from the camera 330 to the outside. The user button 350 includes a plurality of buttons for the user's convenience. For example, the user button 350 may perform a function such as a reset button and a retry.

4 is a block diagram of a face recognition apparatus according to an exemplary embodiment of the present invention. Referring to FIG. 4, the detector 410, the feature information extractor 420, the partial image information extractor 421, the like feature extractor 423, the adaboost extractor 425, and the linear discriminant analyzer ( 427, the feature information comparison determination unit 430, the storage unit 440, and the output unit 450 are illustrated.

The detector 410 is connected to the partial image information extractor 421, receives a face image of a recognition target photographed from a camera, detects a face from a background, and generates face image information to generate a partial image information extractor 420. To transmit. The detector 410 may detect the face by using the like feature information based on the face image. The feature information extractor 420 is divided into a partial image information extractor 421, a like feature extractor 423, an adamboost extractor 425, and a linear discriminant analyzer 427.

The partial image information extracting unit 421 is connected to the detecting unit 410, and extracts partial image information to be analyzed due to low variability and strong features among the facial image information generated by the detecting unit 410. That is, since human hair, neck, etc. are not highly variable or characteristic, the part except for this is extracted from the photographed face image.

The like feature extractor 423 is connected to the partial image information extractor 421, and extracts feature information using only a simple addition operation as described above. This is considerably less than the PCA and requires a certain amount of time to extract the feature values regardless of the training set. As described above, the adamboost extractor 425 extracts feature information decisive for face recognition during each step. In each step, deterministic feature information on the face information queried based on the face image information of the training set is extracted for face recognition.

The linear discriminant analysis unit 427 is connected to the adaboost extractor 425 or the like feature extractor 423, and forms a discriminant that allows each person to express their information well. This best distinguishable segment is created and the pattern is judged by the feature values projected onto it. The extracted like feature information is used as an input of linear discriminant analysis (LDA), and the feature information is trained to extract a basis vector of the linear discriminant analysis. The extracted basis vector is used to obtain feature information of the white feature / linear discriminant analysis of face images in the data set.

The feature information comparison determination unit 430 is connected to the linear discriminant analysis unit 427 and the storage unit 440, and uses the Euclidean distance between coefficients of the similarity feature / linear discriminant analysis, and has the smallest distance. The face image is determined to be the face of the same person as the queried image.

Euclidean distance is as follows.

: 유클리디언 거리

Euclidean Street

: 질의한 얼굴 이미지의 i번째 특징값

: I th feature value of the queried face image

: 데이터 셋 얼굴 이미지의 i번째 특징값

: I th feature value of data set face image

The storage unit 440 is connected to the feature information comparison determination unit 430, and extracts and stores the coefficients of the like feature / linear discriminant analysis from the face image in advance, and the output unit 450 is photographed by the input unit 410. The screen outputs one face image, and outputs the content determined by the feature information comparison determination unit 430.

FIG. 5 is a diagram for extracting partial image information according to a preferred embodiment of the present invention, and FIG. 6 is a diagram for extracting feature information from partial image information according to a preferred embodiment of the present invention. 5 and 6, faces 510 and 610 of a person to be recognized, an extracted partial image 520, and extracted like feature information 620 are illustrated.

The extracted partial image 520 is extracted based on the eye position of the face image, excluding an image such as a hair having a high variability or a neck without a feature. Since the data set previously stored in the storage unit is normalized with respect to the eye position, an easy comparison between the feature information can be made.

The extracted like feature information 620 represents the feature information of the edge and the feature information of the line, and the feature information may be extended up to 24 * 24 in the X and Y directions in the window. In addition, since the feature information changes freely in various cases that can be located in the window, a lot of feature information can be obtained. Referring to FIG. 6, line feature information 620 is illustrated.

Concrete experiment

In the above, the general configuration of the face recognition apparatus and method using the Like Like feature / linear discriminant analysis has been described. Hereinafter, the face recognition apparatus and method using the Like Like feature / linear discriminant analysis according to the present invention will be described with reference to the accompanying drawings. This will be described based on specific experiments performed. In this example, the results of three experiments will be described. The first experiment is about the recognition rate of PCA / linear discriminant analysis and like feature / linear discriminant analysis according to the change of feature information number. The second experiment is about the recognition rate according to the number of like feature information. The third experiment is about the training time and the time when the face image is retrieved from the database.

7 is a diagram illustrating the configuration of data used in a face recognition experiment according to a preferred embodiment of the present invention.

The experiment was performed by randomly extracting 300 people from 2000 data sets. Referring to FIG. 7, five pieces of image information are stored for each person, and a picture including a change in expression, a change in pose, and a change in lighting is included. The images used in this experiment are 46 * 56 gray images, each of which is normalized to the eye position. Through the training process, the basis vector of linear discriminant analysis was extracted and feature information was extracted from the test set. The test set extracts feature information of three video images for each person, stores them in a database, and makes a query using the remaining two video images. When the person with the smallest Euclidean distance is selected as the first priority The face recognition rate was examined by checking whether the same person was the same person.

8 is a diagram illustrating a recognition rate extracted in a face recognition experiment according to a preferred embodiment of the present invention.

This is a result of comparing the face recognition rate change when the number of feature information of the like feature / linear discriminant analysis and the number of feature information of the PCA / linear discriminant analysis are equally changed. Referring to FIG. 8, it can be seen that as the number of feature information used for face recognition increases, the recognition rate also increases. In addition, when using more than 55 pieces of feature information, it can be seen that a recognition value of about 98% can be obtained. Here, a number of prototypes of howe features have been extracted as feature information that best expresses faces through the Adaboost learning algorithm. These results show that the similarity recognition / face discrimination analysis shows similar facial recognition rate as compared with the PCA / linear discrimination analysis.

9 is a diagram illustrating a recognition rate extracted in a face recognition experiment according to another exemplary embodiment of the present invention.

This is a result of measuring the change in the recognition rate of the face while changing the number of the like feature information by the like feature / linear discriminant analysis. At this time, the number of feature information extracted by linear discriminant analysis was fixed to 50 pieces. Referring to FIG. 9, when 50 or more of the number of like feature information is used, the face recognition rate is constant at 98%. However, when less than 50 are used, the face recognition rate decreases in a straight line. Therefore, in order to increase the face recognition rate in face recognition, more than 50 number of similar feature information should be used.

In addition, the following table compares the training speeds in the PCA / linear discriminant analysis and the like feature / linear discriminant analysis.

Face recognition algorithm Training speed PCA / Linear Discrimination Analysis 8 minutes 25 seconds How Like Feature / Linear Discrimination Analysis 37.5 seconds

Table 1.Comparison of Training Speeds in PCA / Linear Discrimination Analysis and Like Feature / Linear Discrimination Analysis

In this experiment, we used 83 feature information for both algorithms and showed the same face recognition rate. Referring to Table 1, it can be seen that the training speed is about 13 times or more when the PCA / linear discriminant analysis is used than when the similar feature / linear discriminant analysis is used. This is because the white feature information can be obtained quickly by simply adding, while the PCA requires complex calculations in the process of matrix calculation of large dimensions to obtain the basis vector. This training speed is an important factor when face recognition is required in real time.

In addition, the following table compares the search speeds in PCA / linear discriminant analysis and white feature / linear discriminant analysis.

Face recognition algorithm Search speed PCA / Linear Discrimination Analysis 1181 ms How Like Feature / Linear Discrimination Analysis 93 ms

[Table 2] Comparison of Search Speeds in PCA / Linear Discrimination Analysis and Like Feature / Linear Discrimination Analysis

In this experiment, when feature information is extracted by PCA / linear discriminant analysis and like feature / linear discriminant analysis, it is compared with facial feature information stored in the database and the time taken to extract the image of the same face as the queried face is determined. It is a comparison. Referring to Table 2, it can be seen that when the PCA / linear discriminant analysis is used, the search speed is about 13 times higher than when the same feature / linear discriminant analysis is used. In this experiment, 100 people were randomly extracted, and each person had 5 pictures. Among them, 3 face pictures were extracted and stored in a database, and the remaining 2 face pictures were queried. Since the basis vector obtained through PCA / linear discriminant analysis is a matrix multiplied by the number of feature information extracted from the horizontal, vertical and linear discriminant analysis of the image, a complex calculation must be performed when the input image is received. However, in the case of the howe feature / linear discriminant analysis, since the number of the feature feature information is multiplied by the number of feature information extracted by the linear discriminant analysis, the retrieval speed is faster than that by the PCA / linear discriminant analysis.

Face recognition procedure

10 is a flowchart illustrating a face recognition method according to a first embodiment of the present invention.

In operation S1010, the queried face image is received, the face of the person is detected by distinguishing it from the background, and the face image information is generated and transmitted to the partial image information extracting unit. In operation S1020, the partial image information excluding the part where the feature of the face is clearly displayed, that is, the hair and the neck, is extracted from the received facial image image and transmitted to the like feature information extracting unit.

In step S1030, the like feature information is extracted from the partial image information, and in step S1040, the stronger face image feature information is extracted from the like feature information using an Adaboost learning algorithm, and in step S1050, the extracted face image The final feature information is extracted by performing linear discriminant analysis based on the feature information.

In step S1060, the final feature information extracted through the linear discriminant analysis and the final feature information previously stored in the storage unit are searched and compared to determine whether they are the same person.

11 is a flowchart illustrating a face recognition method according to a second preferred embodiment of the present invention. The following description focuses on the differences from the first embodiment.

In operation S1130, the partial image extracting unit performs a preprocessing process to minimize the influence of light on the extracted partial image information. The preprocessing process is largely divided into two types, and may be independently performed. However, the preprocessing process may be performed continuously to increase the recognition rate. The first is a process for minimizing the influence of the illumination of the entire face, which is a query object, and performing Fast Lightning Correction (FLC) on the face image information. This FLC normalizes the brightness components by subtracting the mean from the brightness values of the image and dividing the variance. This can be expressed as an expression:

This is an example of an extended set of Haar-like Features for Rapid Object, published by Rainer Lienhart and Jochen Mayd in September 2002. The detailed description of this is omitted since it is described in the article "Detection".

In addition, the second step is to divide the image in half and add the left and right images to each other by multiplying the original image by multiplying the weights to minimize the change in illumination from the left or right. By using the preprocessing as described above, the recognition rate of the queried face can be improved by minimizing the influence of lighting.

The present invention is not limited to the above embodiments, and many variations are possible by those skilled in the art within the spirit of the present invention.

As described above, the apparatus and method for face recognition using the likeness / linear discriminant analysis according to the present invention can reduce the time required for face recognition.

In addition, another effect of the present invention can increase the face recognition rate by minimizing the influence of the ambient light to recognize the face.

Another effect of the present invention is to set a specific region on the face image and generate feature information on the region, thereby increasing the face recognition rate by not extracting information on the region having high variability.

Claims (12)

An apparatus for recognizing a face of a person using haar-like features and linear discriminant analysis (LDA); A detector for detecting a face of a person and generating face image information; A partial image information extracting unit extracting partial image information to be analyzed from the face image information; A like feature extractor extracting like feature information from the partial image information; A linear discriminant analyzer configured to extract linear feature information of the face by performing linear discriminant analysis based on the extracted like feature information; A feature information comparison determiner which determines whether the same person is the same by comparing the last feature information with previously stored last feature information; A storage unit which stores final feature information corresponding to at least one face of the recognition object; And And an output unit configured to output a determination result of the feature information comparison determination unit. The method of claim 1; The detection unit detects a face by using like feature information. The method of claim 1; The apparatus may further include an adamboost extracting unit extracting feature information of the face image information by using an AdaBoost learning algorithm based on the extracted like feature information. And the linear discriminant analyzing unit performs linear discriminant analysis on the basis of the feature information extracted by the adabost extracting unit. The method of claim 1; The partial image information is a face recognition apparatus, characterized in that the image information excluding the hair and neck portion of the human face. The method of claim 1; The partial image information extraction unit Performing Fast Lighting Correction (FLC) on the extracted partial image information, wherein the FLC normalizes the brightness components by subtracting the average from the brightness values of the image and dividing by variance. The method of claim 1; The partial image information extraction unit And dividing the extracted partial image information into left and right halves so as to overlap the image information of the left image and the image information of the right image by multiplying the original image information and multiplying the information of the original image by weight. The method of claim 1; The partial image information extraction unit And extracting the partial image information based on the eye position of the face image. A method for recognizing a face of a person using haar-like features and linear discriminant analysis (LDA); Detecting a face of a person and generating face image information; Extracting partial image information to be analyzed from the face image information; Extracting like feature information from the partial image information; Extracting final feature information of the face image information by performing linear discriminant analysis based on the extracted like feature information; And And comparing the extracted final feature information with previously stored final feature information to determine whether the same person is the same person. The method of claim 8, Detecting a face by using like feature information in the step of detecting the face of the person and generating face image information. The method of claim 8, The method may further include extracting feature information of the face image information by using an AdaBoost learning algorithm based on the extracted like feature information. And performing linear discriminant analysis based on the extracted feature information. The method of claim 8, Performing Fast Lighting Correction (FLC) on the extracted partial image information, further comprising: normalizing the brightness component by subtracting the average from the brightness value of the image and dividing by the variance. Face recognition method. The method of claim 8, And dividing the extracted partial image information into left and right halves, overlapping the image information of the left image and the image information of the right image, and multiplying and adding the multiplying weights to the information of the original image.

Images