This application claims priority under 35 U.S.C. §119(a) to a Korean Patent Application filed in the Korean Intellectual Property Office on May 9, 2007 and assigned Serial No. 2007-44981, the disclosure of which is incorporated herein by reference.
1. Field of the Invention
The present invention relates generally to a system and method for verifying the face of a user using a light mask, and in particular, to a system and method for recognizing an image input through a camera and verifying a registered user in a robotic environment.
2. Description of the Related Art
In the prior art, there are many technologies for recognizing and verifying a face. Conventional face verification technologies are mainly related to a process for dividing a facial image into blocks, extracting features from respective blocks, and creating Gaussian Mixture Models (GMMs). The division of a facial image into blocks is shown in
An aspect of the present invention is to address at least the above problems and/or disadvantages and to provide at least the advantages described below. Accordingly, an aspect of the present invention is to provide a face verification system and method that extract features from an entire facial area, unlike a conventional method of dividing a facial area into blocks and extracting facial features from the blocks, thus sufficiently taking into account the entire facial area. In detail, facial features are extracted from the entire facial area, as shown in
Another aspect of the present invention is to provide a face verification system and method, which increase the amount of data through linear transformation that applies light masks to a face, in order to cope with situations in which the amount of data extracted through feature extraction from the entire facial area is decreased. In detail, as shown in
According to one aspect of the present invention, there is provided a system for verifying a face of a user using a light mask. The system includes a facial feature extraction unit for extracting a facial feature vector from a specific facial image, a non-user Gaussian Mixture Model (GMM) configuration unit for generating a non-user GMM from a non-user facial image stored in a non-user database (DB), a user GMM configuration unit for generating a user GMM by applying light masks to a user facial image stored in a user DB, a log-likelihood value calculation unit for inputting the facial feature vector both to the non-user GMM and to the user GMM, thus calculating log-likelihood values, and a user verification unit for comparing the calculated log-likelihood values with a predetermined threshold, thus verifying whether the specific facial image is a facial image of the user.
According to another aspect of the present invention, there is provided a method of verifying a face of a user using a light mask. The method includes extracting a facial feature vector from a specific facial image, calculating log-likelihood values by inputting the facial feature vector to a Gaussian Mixture Model (GMM) configuration unit for storing a non-user GMM and a user GMM, storing a non-user GMM, generated from a non-user facial image stored in a non-user Database (DB) required to calculate log-likelihood values, and a user GMM, generated by applying light masks to a user facial image stored in a user DB, comparing the calculated log-likelihood values with a predetermined threshold, thus verifying whether the specific facial image is a facial image of the user.
The above and other aspects, features and advantages of the present invention will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:
Preferred embodiments of the present invention will now be described in detail with reference to the annexed drawings. In the drawings, the same or similar elements are denoted by the same reference numerals even though they are depicted in different drawings. In the following description, a detailed description of known functions and configurations incorporated herein has been omitted for clarity and conciseness.
The facial feature extraction unit 320 performs preprocessing on the specific facial image received from the image input unit 310, and thereafter extracts a facial feature vector from the preprocessed image. The facial feature vector can be extracted using an entire facial area as a unit for facial feature extraction.
The non-user GMM configuration unit 400 generates and stores a non-user GMM for the non-user facial image of each non-user registered in a non-user DB. The user GMM configuration unit 500 generates and stores a user GMM for the user facial image of each user registered in a user DB. In particular, the user GMM configuration unit 500 generates facial images illuminated by a plurality of lighting devices by applying light masks to the facial feature vector extracted from the facial feature extraction unit 320. Therefore, the amount of data about the facial feature vector increases, and verification probability can be increased when user verification is performed. Construction related to the generation of a typical GMM will be described later.
The log-likelihood value calculation unit 330 calculates log-likelihood values by inputting the facial feature vector, extracted from the facial feature extraction unit 320, both to the non-user GMM and to the user GMM. Equation [1] indicates a procedure for calculating log-likelihood values. When there is a plurality of users, a plurality of resultant values for the user GMM is obtained. In this case, the largest log-likelihood value is taken and calculated.
z=log(p(X|SFamily))−log(p(X|SGeneral)) [1]
The user verification unit 340 compares the value z, obtained by the log-likelihood value calculation unit 330, with an actual threshold, thus verifying whether the face in the image received from the image input unit 310 is the user's face. When the value z is equal to or greater than the threshold, the specific facial image is determined to be the user's face, whereas, when the value z is less than the threshold, the specific facial image is determined to be the non-user's face.
A procedure actually performed by the non-user GMM configuration unit 400 is represented by the following equations. When the number of facial images stored in the non-user DB 410 is N and is represented by X=[x1, . . . , xN], data about the facial images is modeled in the form of a Gaussian Mixture Model (GMM). Typically, Gaussian probability density is obtained by the following Equation [2] when the dimension of facial image data is D and facial image data is x.
A plurality of Gaussian probability densities forms a single probability model. This probability model is designated as a GMM, which can be represented by the following Equation [3], where M is the number of mixtures.
In this model, parameters can be summarized as follows: the mean μj, covariance Σj (or σj), and weight P(j) (or ωj) of each mixture component j. These parameters are learned using given facial image data x. A learning method is performed to find parameters for allowing a mixture model to most satisfactorily represent the facial image data x, that is, parameters for maximizing a log-likelihood value. A log-likelihood function to be maximized is represented by the following Equation [4], where λ={μj, Σj, ωj|j=1, . . . , M}.
In order to find the parameters for maximizing the log-likelihood function of Equation [4], the parameters are initialized using K-means, and then an Expectation-Maximization algorithm (hereinafter ‘EM algorithm’) is used. By means of this algorithm, respective data items are divided into several groups. This clustering procedure is performed according to the following process.
1. M facial image data items are arbitrarily selected and designated as μ of respective groups.
2. The following procedure is repeated until convergence is achieved.
(1) Euclidian distance ∥xi−μj∥2 for each facial image data item i is obtained and is classified as the closest group.
(2) Values μ of respective groups are calculated again using the classified data.
The EM algorithm is executed using the values μ of respective groups, obtained by performing initialization, as initial values, and the parameters are obtained. The EM algorithm is divided into an E-step and an M-step. In E-step, sufficient statistics are predicted, and, in M-step, the parameters are predicted on the basis of the sufficient statistics. Theses E-step and M-step are described below. First, sufficient statistics for each facial image data item, given in E-step, are represented by the following Equation [5].
The parameters are obtained in M-step on the basis of the sufficient statistics obtained in E-step, as in Equation [6].
The EM algorithm is repeated until satisfactory convergence is achieved. Consequently, the parameters {ωj, μj, σj|j=1 . . . M} of the probability model for representing a non-user GMM can be obtained through the EM algorithm.
1. Occupation probability γm(xi) for each mixture m is obtained in Equation [7].
2. A user GMM is generated from the non-user GMM on the basis of the occupation probability obtained in this way. The adaptation procedure is represented by the following Equation [8],
where αm is a value required to adjust the weights of the non-user GMM and the registered user GMM and is preset through experiments. The generated user GMM is stored in a user GMM storage unit 550.
A linear transformation calculation unit 650 calculates linear transformations using the differences between the facial feature vectors respectively stored in the normal feature storage unit 630a and the linearly amplified feature storage unit 640b. The linear transformations calculated using the system of
As is apparent from the foregoing description, according to the present invention, facial features are extracted from an entire facial area, without a facial area being divided into blocks and features being extracted from the blocks, thereby performing verification, with the features of the entire facial area sufficiently taken into account.
In addition, light masks are applied to a facial image, so that the amount of data, which is reduced due to the above construction, can be increased, and the verification of a face vulnerable to illumination can be compensated for.
While the invention has been shown and described with reference to a certain preferred embodiment thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Number | Date | Country | Kind |
---|---|---|---|
10-2007-0044981 | May 2007 | KR | national |
Number | Name | Date | Kind |
---|---|---|---|
6160903 | Hamid et al. | Dec 2000 | A |
20010000025 | Darrell et al. | Mar 2001 | A1 |
20030165260 | Kim et al. | Sep 2003 | A1 |
20030225719 | Juang et al. | Dec 2003 | A1 |
20050105780 | Ioffe | May 2005 | A1 |
20060008149 | Tu et al. | Jan 2006 | A1 |
20060015497 | Keating et al. | Jan 2006 | A1 |
20060053014 | Yoshizawa | Mar 2006 | A1 |
20060123241 | Martinian et al. | Jun 2006 | A1 |
20060136205 | Song | Jun 2006 | A1 |
20070052858 | Zhou et al. | Mar 2007 | A1 |
20070174633 | Draper et al. | Jul 2007 | A1 |
20080059183 | Liu et al. | Mar 2008 | A1 |
Number | Date | Country |
---|---|---|
1020030066512 | Aug 2003 | KR |
1020050091558 | Sep 2005 | KR |
Number | Date | Country | |
---|---|---|---|
20080279426 A1 | Nov 2008 | US |