METHOD AND APPARATUS FOR DETECTING EYES IN FACE REGION

• US 20080118113A1
• Filed: 06/11/2007
• Published: 05/22/2008
• Est. Priority Date: 11/21/2006
• Status: Active Grant

• Current Assignee: Electronics and Telecommunications Research Institute
• Original Assignee: Electronics and Telecommunications Research Institute
• Inventors: CHUNG, Yun Su, JUNG, Sung Uk, YOO, Jang Hee, CHUNG, Kyo Il
• Granted Patent: US 7,929,734 B2
• Time in Patent Office: Days
• US Class Current: 382/117
• CPC Class Codes: G06V 40/193 Preprocessing; Feature extraction

Abstract

Provided are a method and an apparatus for accurately detecting positions of eyes in an input face image. The method includes extracting a symmetric face region from a face image based on magnitude and phase of gradient in each pixel of the face image; detecting available eye positions based on brightness information of the extracted face region; verifying regions around each of the detected eye positions with a classifier which determines whether an input image is an eye image using information obtained by supervised learning on sample images. According to the present invention, faulty eye detection is prevented by removing obscuring elements such as illumination, glasses and hair from an input face image. Also, eye coordinates can be extracted more accurately using an eye classifier.

Claims

1. A method of detecting eyes comprising:

(a) extracting a symmetric face region from a face image based on magnitude and phase of gradient in each pixel of the face image;

(b) detecting available eye positions based on brightness information of the extracted face region;

(c) verifying regions around each of the detected eye positions with a classifier which determines whether an input image is an eye image using information obtained by supervised learning on sample images.

2. The method of detecting eyes of claim 1, wherein operation (a) comprises:

(a1) calculating magnitude and phase of gradient in each pixel of the face image;

(a2) creating a certain-sized window smaller than the face image and comparing an magnitude and phase of gradient in each pixel, which are calculated with reference to a center pixel in the window, with a critical value, and marking pixels which have greater magnitude and phases of gradient than the critical value as symmetric points; and

(a3) applying a symmetric filter obtained from the face image in which the symmetric points are marked to the face image and extracting the symmetric face region from the face image.

3. The method of detecting eyes of claim 2, the method further comprising (a4) removing illumination effects from the face image prior to operation (a1),

wherein operations (a1), (a2) and (a3) are performed on the face image from which illumination effects have been removed.

4. The method of detecting eyes of claim 2, wherein operation (a3) comprises:

obtaining the symmetric filter by Gaussian filtering the face image in which the symmetric points are marked;

performing symmetric filtering by applying the symmetric filter in one-to-one correspondence with each pixel of the face image; and

removing high-frequency components by Gaussian filtering the face image in which the symmetric filtering is performed.

5. The method of detecting eyes of claim 1, wherein operation (b) comprises:

(b1) extracting left and right eye bands by dividing the extracted face image; and

(b2) extracting pixels having low brightness in each of the eye bands as available eye positions.

6. The method of detecting eyes of claim 5, wherein operation (b2) comprises:

inserting arbitrary sample points into the eye bands; and

tracing the movements of the sample points which are converged into pixels having low brightness in the eye bands.

7. The method of detecting eyes of claim 1, wherein operation (c) comprises:

(c1) extracting regions around each of the detected eye positions;

(c2) calculating values of each of the regions by inputting the regions to a classifier which has information on eye properties obtained from supervised learning of various samples of human eye or non-eye images; and

(c3) determining eye positions which have maximum values in the regions as eyes.

8. The method of detecting eyes of claim 7, the method further comprising (c4) verifying the eye positions by performing operations (c1), (c2) and (c3) repeatedly while modifying the size of the face image in which the eye positions are detected in operation (b).

9. An apparatus for detecting eyes comprising:

an image processing unit which extracts a symmetric face region from a face image based on magnitude and phase of gradient in each pixel of the face image;

an eye position detection unit which detects available eye positions based on brightness information of the extracted face region;

an eye position verification unit which verifies regions around each of the detected eye positions with a classifier which determines whether an input image is an eye image using information obtained by supervised learning on sample images.

10. The apparatus for detecting eyes of claim 9, wherein the image processing unit comprises:

an edge calculation unit which calculates magnitude and phase of gradient in each pixel of the face image;

a symmetric point marking unit which creates a certain-sized window smaller than the face image and compares magnitude and phase of gradient in each pixel, which are calculated with reference to a center pixel in the window, with a critical value, and marks pixels which have greater magnitude and phase of gradient than the critical value as symmetric points; and

a symmetric region extraction unit which applies a symmetric filter obtained from the face image in which the symmetric points are marked to the face image and extracts the symmetric face region from the face image.

11. The apparatus for detecting eyes of claim 10, the apparatus further comprising an illumination processing unit which removes illumination effects from the face image prior to the calculating of the magnitude and phase of gradient in each pixel of the face image.

12. The apparatus for detecting eyes of claim 10, wherein the symmetric region extraction unit comprises:

a symmetric filter obtaining unit which obtains the symmetric filter by Gaussian filtering the face image in which the symmetric points are marked;

a filtering unit which performs symmetric filtering by applying the symmetric filter in one-to-one correspondence with each pixel of the face image, and removes high-frequency components by Gaussian filtering the face image in which the symmetric filtering is performed.

13. The apparatus for detecting eyes of claim 9, wherein the eye position detection unit comprises:

an eye band extraction unit which extracts left and right eye bands by dividing the extracted face image; and

an eye position extraction unit which extracts pixels having low brightness in each of the eye bands as available eye positions.

14. The apparatus for detecting eyes of claim 13, wherein the eye position extraction unit extracts eye positions by inserting arbitrary sample points into the eye bands and tracing the movements of the sample points which are converged into pixels having low brightness in the eye bands.

15. The apparatus for detecting eyes of claim 9, the eye position verification unit comprising:

an eye position region extraction unit which extracts regions around each of the detected eye positions;

a classification unit which calculates values of each of the regions by inputting the regions to a classifier which has information on eye properties obtained from supervised learning of various samples of human eye or non-eye images; and

an eye determination unit which determines eye positions which have maximum values in the regions as eyes.

16. The apparatus for detecting eyes of claim 15, the apparatus further comprising an image size modification unit which modifies the size of the face image in which the eye positions are detected.