Face recognition by dividing an image and evaluating a similarity vector with a support vector machine

• US 7,873,189 B2
• Filed: 09/06/2006
• Issued: 01/18/2011
• Est. Priority Date: 11/26/2005
• Status: Expired due to Fees

• Current Assignee: Electronics and Telecommunications Research Institute
• Original Assignee: Electronics and Telecommunications Research Institute
• Inventors: Jee, Hyung Keun, Chung, Yun Su, Chung, Kyo Il, Yoo, Jang Hee
• Primary Examiner(s): Werner; Brian P
• Assistant Examiner(s): DRENNAN, BARRY T
• Application Number:
  US11/516,090
• Publication Number:
  US 20070122009A1
• Time in Patent Office: 1,595 Days
• Field of Search: 382/115, 382/117, 382/118, 382/155, 382/159, 382/160, 382/224, 382/228, 706/10, 706/12, 706/14
• US Class Current: 382/118
• CPC Class Codes: G06V 40/16 Human faces, e.g. facial parts, sketches or expressions

Abstract

A face recognition and apparatus are provided. According to the method, an SVM classifier is created through machine learning on the basis of a degree of similarity of a divided facial image, and a facial image to be authenticated is normalized to a predetermined size using a center between two eyes. The normalized image is divided into more than one image in horizontal and vertical directions, respectively. Next, predetermined characteristic vectors from the divided images are extracted and a similarity vector based on a degree of similarity with respect to a registered characteristic vector is created. The similarity vector is input to the SVM classifier, so that authentication is performed.

Claims

1. A face recognition method comprising:

creating an SVM (support vector machine) classifier through machine learning on the basis of a degree of similarity between stored facial images in a database;

normalizing a measured facial image to a predetermined size and subsequently dividing the normalized measured facial image along horizontal and vertical directions into at least six sub-divided images of the normalized measured facial image, wherein dividing the normalized measured facial image comprises;

dividing the normalized measured facial image into three equal regions horizontally which respectively represent an eye region, a nose region, and a mouth region and which each have a height equal to ⅔

the distance between the eyes; and

dividing the normalized measured facial image into three equal regions vertically which respectively represent a left eye region, a nose region, and a right eye region and which each have a width equal to ⅔

the distance between the eyes;

extracting characteristic vectors from each of the sub-divided images of the normalized measured facial image and creating a similarity vector based on a degree of similiarity between registered characteristic vectors associated with at least one of the stored facial images and that of the extracted characteristic vectors of the normalized measured facial image, wherein the extracted and registered characteristic vectors are obtained by applying PCA (principal component analysis) to reduce the dimensions of data of images, and then applying LDA (linear discriminant analysis), wherein six extracted characteristic vectors are created from six sub-divided images of the normalized measured facial image; and

inputting the similarity vector to the SVM classifier to perform authentication of the measured facial image.

2. The method of claim 1, wherein creating of the SVM classifier comprises:

normalizing a stored facial image of the stored facial images to the predetermined size, dividing along horizontal and vertical directions the normalized stored facial image into a plurality of sub-divided images of the normalized stored facial image, and extracting characteristic vectors from the sub-divided images of the normalized stored facial image;

comparing the characteristic vectors from the normalized stored facial image against registered characteristic vectors of other divided images of the stored facial images to measure the degree of similarity; and

normalizing the measured degree of similarity of the compared characteristic vectors between −

1 and +1.

3. The method of claim 1, wherein six extracted characteristic vectors are created from six sub-divided images of the normalized measured facial image.

4. The method of claim 1, wherein normalizing of the measured facial image comprises:

rotating the measured facial image using the slope of a line connecting centers of two eyes to correct an overall slope of the measured facial image; and

clipping the measured facial image using the distance between the two eyes as a reference and sampling the clipped facial image into the predetermined size to normalize the measured facial image.

5. The method of claim 1, wherein extracting of the characteristic vectors comprises:

projecting the sub-divided images onto a PCA transform matrix to reduce the dimensions of the sub-divided images;

projecting the sub-divided images onto an LDA transform matrix to extract the extracted characteristic vectors; and

comparing the extracted characteristic vectors with registered characteristic vectors which are registered in advance to measure the degree of similarity, and to create the similarity vector.

6. A face recognition apparatus having a processor and memory comprising:

a support vector machine (SVM) configured to create an SVM classifier through machine learning on the basis of a degree of similarity between stored facial images in a database;

an image division unit configured to normalize a measured facial image to a predetermined size and subsequently divide the normalized measured facial image along horizontal and vertical directions into a plurality of at least six sub-divided images of the normalized measured facial image, wherein the image division unit comprises;

a normalizer configured to rotate the measured facial image using the slope of a line connecting centers of two eyes to correct an overall slope of the measured facial image, to clip the measured facial image using the distance between the two eyes as a reference and to sample the clipped facial image into the predetermined size to normalize the measured facial image;

a horizontal division configured to divide the normalized measured facial image into three equal regions horizontally which respectively represent an eye region, a nose region, and a mouth region and which each have a height equal to ⅔

the distance between the eyes; and

a vertical division part configured to divide the normalized measured facial image into three equal regions vertically which respectively represent a left eye region, a nose region, and a right eye region and which each have a width equal to ⅔

the distance between the eyes;

a similarity vector creating unit configured to extract characteristic vectors from each of the sub-divided images of the normalized measured facial image and to create a similarity vector based on a degree of similiarity between registered characteristic vectors associated with at least one of the stored facial images and that of the extracted characteristic vectors of the normalized measured facial image; and

an authentication unit configured to input the similarity vector to the SVM classifier to perform authentication of the measured facial image.

7. The apparatus of claim 6, wherein the SVM comprises:

a first vector creator configured to normalize a stored facial image of the stored facial images to the predetermined size, to divide along horizontal and vertical directions the normalized stored facial image into a plurality of sub-divided images of the normalized stored facial image, and to extract characteristic vectors from the sub-divided images of the normalized stored facial image;

a second vector creator configured to compare the characteristic vectors from the normalized stored facial image against registered characteristic vectors of other divided images of the stored facial images to measure the degree of similarity and to normalize the measured degree of similarity of the compared characteristic vectors between −

1 and +1; and

an SVM learning part configured to designate a 1 into the similarity vector when comparing the stored facial image against itself, and to designate a −

1 into the similarity vector when comparing the stored facial image against another stored facial image of the stored facial images to learn and create the SVM classifier.

8. The apparatus of claim 6, wherein the similarity vector creating unit comprises:

a PCA (principal component analysis) part configured to project the sub-divided images onto a PCA transform matrix to reduce the dimensions of the sub-divided images;

an LDA (linear discriminant analysis) part configured to project the sub-divided images onto an LDA transform matrix to extract the extracted characteristic vectors; and

a similarity vector output part comparing the characteristic vectors with a characteristic vector registered in advance to measure a degree of similarity, and creating the similarity vector using a result of the measurement configured to compare the extracted characteristic vectors with registered characteristic vectors which are registered in advance to measure the degree of similarity, and to create the similarity vector.