System and method for face detection through geometric distribution of a non-intensity image property

• US 6,792,135 B1
• Filed: 10/29/1999
• Issued: 09/14/2004
• Est. Priority Date: 10/29/1999
• Status: Expired due to Term

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Toyama, Kentaro
• Primary Examiner(s): Mehta, Bhavesh M.
• Assistant Examiner(s): CARTER, AARON W
• Application Number:
  US09/430,560
• Time in Patent Office: 1,782 Days
• Field of Search: 382/118, 382/209, 382/199, 382/195, 382/191, 382/173, 725/10
• US Class Current: 382/118
• CPC Class Codes: G06V 40/161 Detection; Localisation; Normalisation

Abstract

The present invention is embodied in a system and method for detecting a face within an image using a relational template over a geometric distribution of a non-intensity image property. In general, the system of the present invention includes a hypothesis module for defining a sub-region in which to search for a face, a feature extraction module for extracting image feature values image based on a non-intensity image property, an averaging module for grouping the extracted image feature values into geometrically distributed facial regions, and a relational template module that uses a relational template and facial regions to determine whether a face has been detected. In a preferred embodiment the image property used is edge density, although other suitable properties (such as pixel color) may also be used. The method of the present invention includes performing feature extraction on the image based on an image property (such as edge density), grouping extracted image feature values into facial regions and using a relational template to determine whether a face has been detected.

Claims

1. A system for detecting a face in an image, comprising:

a feature extraction module that extracts features from the image using a non-intensity image property;

a feature template that determines a high-frequency variation around each pixel in the image;

a feature averaging module that groups the features into facial regions;

a relational template module that uses a relational template over a geometric distribution of the non-intensity image property; and

regional values assigned by the relational template module that are based on a relationship between the facial regions and are used to determine whether the face has been detected.

2. The system of claim 1, wherein the non-intensity image property is edge density.

3. The system of claim 1, wherein the feature extraction module extracts features using convolution.

4. The system of claim 3, wherein the feature extraction module further comprises an edge mask for performing edge detection.

5. The system of claim 1, wherein the feature extraction module extracts features using at least one of:

(a) Laplacians;

(b) Gabor wavelets.

6. The system of claim 1, wherein the feature averaging module combines image feature values for each feature.

7. The system of claim 6, wherein the feature averaging module averages the image feature values for each facial region.

8. The system of claim 1, wherein the relationship between the facial regions is based on the non-intensity image property.

9. The system of claim 8, wherein the relational template is a matrix.

10. The system of claim 1, wherein a relational value is determined using regional values.

11. The system of claim 10, wherein the face is detected if the relational value is greater than a threshold value.

12. The system of claim 1, further comprising a preprocessing module capable of at least one of:

(a) resizing the image;

(b) masking the image to filter out unwanted background noise;

(c) performing histogram equalization on the image.

13. The system of claim 1, further comprising a hypothesis module that generates a hypothesis about a location of the face in the image and defines a sub-region within the image based on the hypothesis.

14. The system of claim 13, wherein the hypothesis is based at least partially on the non-intensity image property.

15. The system of claim 1, wherein the facial regions are geometrically distributed in a facial arrangement.

16. A method for detecting a face in an image, comprising:

extracting features from the image using a non-intensity image property;

determining a high-frequency variation around each pixel in the image;

grouping the extracted features into facial regions based on a geometric distribution of the non-intensity image property; and

determining whether a face has been detected in the image by using a relational template over a geometric distribution of the non-intensity property and a relationship between the facial regions.

17. The method of claim 16, wherein the non-intensity image property is edge density.

18. The method of claim 16, wherein the facial regions are geometrically distributed in a facial arrangement.

19. The method of claim 16, wherein the feature template is sensitive to high spatial frequencies.

20. The method of claim 19, wherein extracting further comprises extracting at least one image feature value for each feature.

21. The method of claim 20, wherein the feature template is a texture template that uses edge density.

22. The method of claim 19, wherein extracting further comprises using convolution to convolve the image with the feature template.

23. The method of claim 19, wherein extracting further includes using at least one of:

(a) Laplacians;

(b) Gabor wavelets.

24. The method of claim 20, grouping further comprises combining image feature values for each feature.

25. The method of claim 16, wherein the relationship between the facial regions is based on the non-image property.

26. The method of claim 16, further comprising assigning a regional value for each of the facial regions based on the relationship.

27. The method of claim 26, further comprising defining a relational value based on the regional values.

28. The method of claim 27, wherein determining further comprises detecting the face if the relational value is greater than a threshold value.

29. The method of claim 16, further comprising:

generating a hypothesis about the location of the face within the image; and

defining a sub-region within the image to search for the face.

30. The method of claim 29, wherein the hypothesis is based on at least one of:

(a) a scale;

(b) an aspect ratio;

(c) a location in the image.

31. The method of claim 29, wherein the hypothesis is based at least partially on the non-intensity image property.

32. The method of claim 16, further comprising preprocessing the image.

33. The method of claim 16, wherein preprocessing comprises at least one of:

(a) resizing the image;

(b) filtering out unwanted noise from the image;

(c) performing histogram equalization on the image.

34. A method for detecting a face in an image, comprising:

extracting features from the image using a non-intensity image property and a feature template to determine high-frequency variation around each pixel of the image;

placing the features into facial regions;

using a relational template over a geometric distribution of the non-intensity property to determine a relationship between the facial regions;

defining a relational value based on the relational template; and

detecting a face within the image if the relational value is greater than a threshold value;

wherein the non-intensity image property includes at least one of;

(a) edge density;

(b) pixel color.

35. The method of claim 34, further comprising using information about the detected face in a computer vision system.

36. The method of claim 35, wherein the computer vision system is used to interpret the information about the detected face.

37. The method of claim 35, wherein the computer vision system is used to recognize the detected face.

38. A computer-implemented method for deciding whether a face is present in a image, comprising:

generating an hypothesis as to a location in the image where a face may be present;

defining a sub-region in the image based on the generated hypothesis;

extracting image feature values based on a non-intensity image property for each pixel in the sub-region using a feature template that is sensitive to high frequencies to determine the degree of high-frequency variation around each pixel in the sub-region;

defining facial regions in the sub-region of the image;

combining each of the image feature values in each of the facial regions to obtain a combined feature value for each of the facial regions;

determining a relationship between any two of the facial regions and assigning a regional value based on the relationship;

summing the regional values to obtain a relational value; and

comparing the relational value to a threshold value to decide whether a face is present in the image.

39. The computer-implemented method as set forth in claim 38, wherein the generated hypothesis comprises information about at least one of:

(a) images scales;

(b) aspect ratios;

(c) image locations;

(d) target regions of the image obtained by using vision processing techniques.