KR20100089360A - Method and apparatus for face recognition using boosting method - Google Patents

Abstract

PURPOSE: A method and an apparatus for recognizing a face by using a boosting method are provide to keep the consistence of a learning method by using a feature vector dimension reduction, a feature learning and a boosting method. CONSTITUTION: Through the representative feature vector of a representative feature vector extractor(141) and the feature vectors of a feature vector extractor(131), a binary feature vector extractor(142) extracts the binary feature vectors. The binary feature vectors indicates the deviation between the feature vectors, and a learning and vector dimension reduction unit(151) performs learning according a boosting method for the feature of the binary feature vector. Through the dominant feature, the learning and vector dimensional reduction unit configures a reduce dimension representative feature vector.

Description

Method and apparatus for face recognition using boosting technique {METHOD AND APPARATUS FOR FACE RECOGNITION USING BOOSTING METHOD}

Embodiments of the present invention relate to a face recognition method and apparatus.

The recognition performance of the face recognition system depends on the learning method used in the feature vector dimension reduction process and feature learning process.

When the feature vector dimension is reduced, various learning techniques can be used. Recently, the boosting technique has been widely used in recognition of excellent learning performance.

In the conventional face recognition method, the boosting technique can only be used for the feature vector dimension reduction process. This is because boosting techniques designed for binary-class classification problems cannot be directly applied to multi-class classification problems such as face recognition.

Therefore, in the existing face recognition method, separate learning methods such as linear discriminant analysis (LDA), support vector machine (SVM), and back-propagation (BP) were used in the feature learning and recognition process.

However, considering that the boosting technique has better learning and recognition ability than the learning technique, and that the consistency of the learning technique can maximize the learning and recognition performance throughout the face recognition system, It is reasonable to use the same boosting technique used in the process in the feature learning and recognition phase.

On the other hand, in the conventional face recognition method, since overlapping learning by different learning techniques is inevitable in the feature vector dimension reduction process and feature learning and recognition process, the configuration and implementation of the entire face recognition system is complicated and the learning for face recognition is required. It was a long time.

One embodiment of the present invention, in the face recognition apparatus and method using a boosting technique in the process of feature vector dimension reduction, feature learning, recognition, etc. of the boosting technique having excellent learning and recognition performance The purpose of the present invention is to increase the utilization and maintain the consistency of the learning technique, improve the face recognition rate, simplify the configuration of the face recognition device, increase the device efficiency, and reduce the learning time for face recognition.

In accordance with an aspect of the present invention, a face recognition apparatus includes: a feature vector extractor configured to extract feature vectors from registered images; A representative feature vector extracting unit for extracting a representative feature vector that is a vector obtained by calculating an average of the feature vectors; A binary feature vector extracting unit for extracting a binary feature vector representing a deviation between feature vectors of the images using the feature vector and the representative feature vector; And perform learning on the features on the binary feature vector according to a boosting technique, extracting a dominant feature based on a learning performance result of the feature, and reducing the dimension of the representative feature vector. Learning and constituting a reduced dimension representative feature vector.

In this case, a reduced dimensional feature vector extracting unit extracts a reduced dimensional feature vector corresponding to a dominant feature of the reduced dimensional representative feature vector from the input image; An identification vector extracting unit for extracting an identification vector used to recognize an identity of the input image using the reduced dimension representative feature vector and the reduced dimension feature vector; And a recognition unit that classifies the identification vectors using the boosting technique on which the learning is performed and recognizes the input image identity.

In this case, the binary feature vector includes an internal deviation feature vector representing a deviation between feature vectors of images of the same object and an external deviation feature vector representing a deviation between feature vectors of images of different objects. The dimension reduction unit assigns a larger learning weight to the binary feature vector by a boosting learning technique, so that the larger the learning error is, and learns each feature on the binary feature vector according to the learning weight, and based on the learning result of the features. The dominant feature is extracted, and the learning and vector dimension reduction unit determines whether the dominant feature is less than the selected number, and when the dominant feature is less than the selected number, By repeating the feature learning procedure, the dominant feature A can be added to the extraction, it constitutes the representative characteristic of the dimension of the reduced-dimensional vector reduction representative feature vector.

A face recognition method according to an embodiment of the present invention comprises the steps of extracting feature vectors from the registered images; Extracting a representative feature vector that is a vector from which the average of the feature vectors is calculated; Extracting a binary feature vector representing a deviation between feature vectors of the images using the feature vector and the representative feature vector; Performing learning according to a boosting technique on the features on the binary feature vector; And extracting a dominant feature having a high learning performance result based on the learning performance result of the feature, and constructing a reduced dimension representative feature vector having a reduced dimension of the representative feature vector.

Extracting a reduced dimension feature vector corresponding to a dominant feature of the reduced dimension representative feature vector from the input image; Extracting an identification vector used to recognize an identity of the input image using the reduced dimension representative feature vector and the reduced dimension feature vector; And classifying the identification vectors by using the boosting technique on which the learning is performed to recognize the input image identity.

According to an embodiment of the present invention, in the face recognition apparatus and its method, a boosting technique having excellent learning and recognition performance by using a boosting technique in the process of feature vector dimension reduction, feature learning, and recognition, etc. By increasing the utilization of the technique and maintaining the consistency of the learning technique, it is possible to improve the face recognition rate, simplify the configuration of the face recognition device, increase the device efficiency, and reduce the learning time for face recognition.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention is not limited or limited by the embodiments. Like reference numerals in the drawings denote like elements.

1 is a block diagram of a face recognition apparatus according to an embodiment of the present invention. A configuration of a face recognition apparatus according to an embodiment of the present invention will be described with reference to FIG. 1.

As shown in FIG. 1, the facial recognition apparatus according to an exemplary embodiment of the present invention may include a feature extractor 130, a binary feature vector and an identification vector extractor 140, and a boosting learner and recognizer 150. ).

The feature extractor 130 may include a feature vector extractor 131 and a reduced dimension feature vector extractor 132.

The feature vector extractor 131 extracts one feature vector from each of the plurality of face images 110 registered for learning.

The feature vector extractor 131 and the reduced dimension feature vector extractor 132 may use the same feature extraction method for feature extraction, and a feature extraction method such as a Gabor filter may be used.

The binary feature vector and identification vector extractor 140 may include a representative feature vector extractor 141, a binary feature vector extractor 142, and an identification vector extractor 143.

The representative feature vector extractor 141 receives the feature vectors extracted for learning from the feature vector extractor 131 and extracts the representative feature vector.

In this case, the representative feature vector is for extracting a binary feature vector and an identification vector, and can be obtained as an average of feature vectors obtained from a face image of the same person.

That is, the representative feature vector refers to a feature vector capable of canceling the deviation caused by the direction or brightness of the various face images of a specific person, brightness, facial expression, etc., and expressing common and main features of the particular person.

For example, if I face images representing U people are registered, I feature vectors may be extracted through the feature vector extractor 131, and U may be extracted through the feature vector extractor 141. Two representative feature vectors will be calculated.

라 하면, 인물 u에 대한 대표 특징 벡터

는 다음의 수학식 1에 의해 구할 수 있다.K feature vectors representing a particular person u (u = 1,2, ..., U)

Suppose that the representative feature vector for person u

Can be obtained by the following equation (1).

The representative feature vector obtained through Equation 1 has the same dimension as the feature vector obtained by the feature vector extractor 131.

The binary feature vector extractor 142 receives the feature vector of the representative feature vector and the registered images from the representative feature vector extractor 141 and extracts the binary feature vector for learning by using the feature vector.

The binary feature vector is divided into an internal deviation feature vector and an external deviation feature vector. The internal deviation feature vector is a vector representing a deviation of feature vectors of the same person, and the external deviation feature vector is a vector representing a difference of feature vectors of different people.

, 인물 u에 대한 대표 특징 벡터를

라 할 때, 임의의 내부 편차 특징 벡터

는 다음의 수학식 2를 통하여 구할 수 있다.K feature vectors representing a particular person u (u = 1,2, ..., U)

, Representative feature vector for person u

Is an arbitrary internal deviation feature vector.

Can be obtained from Equation 2 below.

, ,

,,

는 다음의 수학식 3을 이용하여 구할 수 있다.Similarly any external deviation feature vector

Can be obtained using Equation 3 below.

,

,

,

,

,

,

,

,

The boosting learner and recognizer 150 may include a learner and vector dimension reducer 151 and a recognizer 152.

The learning and vector dimension reduction unit 151 receives a binary feature vector and a representative feature vector from the binary feature vector extractor 142, and obtains a predetermined dominant feature through boosting learning on the binary feature vector. By selecting them, the dimension of the representative feature vector is reduced to obtain a reduced dimension representative feature vector.

Through this process, dimension reduction and feature learning for the feature vector are performed at the same time.

The reduced dimensional feature vector extractor 132 extracts one reduced dimensional feature vector from the input face image 120. In this case, the reduced dimensional feature vector extractor 132 extracts only the dominant features selected through the learning and vector dimensional reducer 151. Therefore, the reduced dimensional feature vector extracted by the reduced dimensional feature vector extractor 132 has a reduced dimension corresponding to the number of selected dominant features.

를 이용하여 U개의 식별 벡터

를 다음과 같이 추출한다.The identification vector extractor 143 is a reduced dimension feature vector f received from the reduced dimension feature vector extractor 132 and U reduced dimension representative feature vectors received from the learning and vector dimension reducer 151.

U identification vectors using

Extract as

,

,

와 동일한 축소된 차원을 갖는다.The identification vector includes only the features corresponding to the dominant feature selected by the learning and vector dimensional reducer 151, and thus the reduced dimensional feature vector f and the reduced dimensional representative feature vector.

Has the same reduced dimension as.

By extending one reduced dimensional feature vector into a plurality of identification vectors as described above, it is possible to use boosting learned about the binary feature vector as it is in the recognition process.

를 전달받고, 학습 및 벡터 차원 축소부(151)에서 학습된 부스팅(Boosting) 장치를 이용하여 식별 벡터들을 분류함으로써 입력이미지의 신원을 인식한다. 이때, 임의의 식별 벡터

에 대한 부스팅(Boosting) 분류결과를 통해

가 내부 대표 편차 특징 벡터일 확률을 얻을 수 있다.The recognition unit 152 may extract U identification vectors extracted through the identification vector extraction unit 143.

The identification of the input image is recognized by classifying the identification vectors using a boosting device learned by the learning and vector dimension reduction unit 151. Where any identification vector

Boosting classification results for

It is possible to obtain the probability that is an internal representative deviation feature vector.

이 가장 큰 확률을 갖고 상 기 확률이 정해진 값보다 크다면 입력이미지는 등록된 전체 U명의 인물 중 인물 m을 표현하는 이미지로서 인식된다. 그러나, 상기 확률이 정해진 값 이하라면 입력이미지는 등록되지 않은 인물로 인식할 수 있다.If the identification vector

If this probability has the greatest probability and the probability is greater than the predetermined value, the input image is recognized as an image representing the person m among all registered U persons. However, if the probability is less than or equal to the predetermined value, the input image may be recognized as an unregistered person.

2 is a flowchart illustrating a face recognition method according to an embodiment of the present invention.

The feature vector extractor extracts feature vectors from the registered images (S210).

Thereafter, the representative feature vector extracting unit extracts a representative feature vector that is a vector obtained by calculating an average of the feature vectors (S220).

Thereafter, the binary feature vector extracting unit extracts a binary feature vector representing a deviation between feature vectors of the images by using the feature vector and the representative feature vector (S230).

In this case, the binary feature vector may include an internal deviation feature vector representing a deviation between feature vectors of images of the same object and an external deviation feature vector representing a deviation between feature vectors of images of different objects.

At this time, the extraction of the binary feature vector simplifies the problem by converting the multi-class problem of face recognition into a binary-class problem, enables the boosting technique, and improves the learning effect. It aims to improve face recognition rate by maximizing the selection of more discriminating features.

Subsequently, a learning and vector dimension reduction unit learns the features on the binary feature vector according to a boosting technique, and based on the learning performance result of the feature, the dominant feature having excellent learning performance results. ) Is extracted to form a reduced dimension representative feature vector in which the dimension of the representative feature vector is reduced (S240).

In this case, the learning and vector dimension reduction unit gives a larger learning weight to the binary feature vector by a boosting learning technique, and gives a larger learning weight and learns each feature on the binary feature vector according to the learning weight. Based on the learning result, the dominant feature excellent in the learning performance result may be extracted.

In this case, the learning and vector dimension reduction unit determine whether the dominant feature falls below a predetermined number, and if the dominant feature falls below a predetermined number, repeats the feature learning procedure by the boosting learning technique. The feature may be further extracted to form a reduced dimension representative feature vector having a reduced dimension of the representative feature vector.

Thereafter, the reduced dimensional feature vector extractor may extract the reduced dimensional feature vector corresponding to the dominant feature of the reduced dimensional representative feature vector from the input image (S250).

Thereafter, the identification vector extracting unit may extract the identification vector used to recognize the identity of the input image by using the reduced dimension representative feature vector and the reduced dimension feature vector (S260).

Extraction of the identification vector extends one reduced dimensional feature vector into multiple identification vectors to solve the face recognition problem, which is a multi-class problem, with boosting on the binary feature vector. In addition, boosting completed in the step S230 may be used as it is in subsequent recognition processes without additional learning techniques and learning processes.

Therefore, according to the extraction of the identification vector according to an embodiment of the present invention, by increasing the utilization of boosting with excellent learning and recognition performance, and maintaining the consistency of the learning technique used for reducing the feature vector dimension, feature learning, and recognition, The face recognition rate can be improved, the face recognition device can be simplified and the efficiency can be improved, and the learning time can be shortened.

Thereafter, the recognition unit may classify the identification vectors using the boosting technique on which the learning is performed to recognize the input image identity (S270).

The recognition unit may recognize the input image by classifying the identification vectors by using boosting learned in operation S140 and determining a degree of matching with the registered images.

3 is a flowchart illustrating a method of reducing a learning and representative feature vector dimension according to an embodiment of the present invention.

The training and vector dimension reduction unit equally weights all binary feature vectors, calculates the error of learning each binary feature vector for the dominant feature each time a dominant feature is selected, and The larger vector may be given a relatively larger learning weight so that more learning can be made (S310).

Thereafter, the learning and vector dimension reduction unit performs feature learning using the binary feature vectors (S320).

In this case, the learning and vector dimension reduction unit may use a learning method of boosting technique as a learning method, and boosting such as AdaBoost, MutualBoost, and Segment-Boost. Techniques can be used.

Thereafter, the learning and vector dimension reduction unit selects a dominant feature according to the learning result (S330). In this case, the learning and vector dimension reduction unit may select a feature having the best learning result as a dominant feature.

Thereafter, the learning and vector dimension reduction unit determines whether the number of the selected dominant features has reached a predetermined required quantity (S340), and when the required quantity is not reached, starting from the first learning weight adjustment step (S310). Again, additional governing features may be selected.

Thereafter, the learning and vector dimension reduction unit removes the remaining features except for the selected dominant features and constructs a reduced dimension representative feature vector including only the dominant features (S350). Therefore, the reduced dimension representative feature vector has a dimension corresponding to the selected total number of dominant features.

In the present invention as described above has been described by the specific embodiments, such as specific components and limited embodiments and drawings, but this is provided to help a more general understanding of the present invention, the present invention is not limited to the above embodiments. For those skilled in the art, various modifications and variations are possible from these descriptions. Therefore, the spirit of the present invention should not be limited to the described embodiments, and all of the equivalents or equivalents of the claims as well as the claims to be described later will belong to the scope of the present invention. .

1 is a block diagram of a face recognition apparatus according to an embodiment of the present invention.

2 is a flowchart illustrating a face recognition method according to an embodiment of the present invention.

3 is a flowchart illustrating a method of reducing a learning and representative feature vector dimension according to an embodiment of the present invention.

<Explanation of symbols for the main parts of the drawings>

130: feature extraction unit

131: feature vector extracting unit

132: reduced dimension feature vector extractor

140: binary feature vector and identification vector extractor

141: representative feature vector extracting unit

142: binary feature vector extractor

143: identification vector extraction unit

150: Boosting Learning and Recognition

151: Lesson and vector dimension reduction

152: recognition unit

Claims (5)

A feature vector extracting unit to extract feature vectors from the registered images; A representative feature vector extracting unit for extracting a representative feature vector that is a vector obtained by calculating an average of the feature vectors; A binary feature vector extracting unit for extracting a binary feature vector representing a deviation between feature vectors of the images using the feature vector and the representative feature vector; And Learning the features of the binary feature vector according to a boosting technique, extracting a dominant feature based on a learning performance result of the features, and dimensioning the representative feature vector. Learn to construct this reduced reduced dimension representative feature vector and reduce the vector dimension Face recognition device comprising a. The method of claim 1, A reduced dimensional feature vector extracting unit for extracting a reduced dimensional feature vector corresponding to a dominant feature of the reduced dimensional representative feature vector from the input image; An identification vector extracting unit for extracting an identification vector used to recognize an identity of the input image using the reduced dimension representative feature vector and the reduced dimension feature vector; And Recognizing unit for classifying the identification vectors by using the boosting technique on which the learning is performed to recognize the input image identity Face recognition device further comprising. The method of claim 1, The binary feature vector is An internal deviation feature vector representing a deviation between feature vectors of images of the same object and an external deviation feature vector representing a deviation between feature vectors of images of different objects, The learning and vector dimension reduction unit, By a boosting learning technique, a larger learning weight is given to the binary feature vector as a learning error increases, and each feature on the binary feature vector is learned according to the learning weight, and the dominant feature is based on the learning result of the features. extract the dominant feature, The learning and vector dimension reduction unit, Determining whether the dominant feature falls below the selected number and if the dominant feature falls below the selected number further extracts the dominant feature by repeating the feature learning procedure by the boosting learning technique, And a face recognition apparatus constituting a reduced dimension representative feature vector having reduced dimensions of the representative feature vector. Extracting feature vectors from registered images; Extracting a representative feature vector that is a vector from which the average of the feature vectors is calculated; Extracting a binary feature vector representing a deviation between feature vectors of the images using the feature vector and the representative feature vector; Performing learning according to a boosting technique on the features on the binary feature vector; And Extracting a dominant feature based on a learning performance result of the feature to construct a reduced dimension representative feature vector having a reduced dimension of the representative feature vector; Face recognition method comprising a. The method of claim 4, wherein Extracting a reduced dimensional feature vector corresponding to a dominant feature of the reduced dimensional representative feature vector from the input image; Extracting an identification vector used to recognize an identity of the input image using the reduced dimension representative feature vector and the reduced dimension feature vector; And Classifying the identification vectors by using the boosting technique on which the learning is performed, and recognizing the input image identity. Face recognition method further comprising.

Images