System and Method for Using Contextual Features to Improve Face Recognition in Digital Images

• US 20130121584A1
• Filed: 08/16/2010
• Published: 05/16/2013
• Est. Priority Date: 09/18/2009
• Status: Active Grant

• Current Assignee: Adobe Inc.
• Original Assignee: Adobe Systems Incorporated (Adobe Inc.)
• Inventors: Bourdev, Lubomir D., Xu, Jiang
• Granted Patent: US 8,503,739 B2
• Time in Patent Office: Days
• US Class Current: 382/190
• CPC Class Codes: G06V 40/171 Local features and components; Facial parts eye characteristics G06V40/18; Occluding parts, e.g. glasses; Geometrical relationships
• CPC Class Codes: G06V 40/179 metadata assisted face recognition

Abstract

Various embodiments of a system and methods for using contextual features to improve face recognition in digital images are described. A face recognition system may semi-automatically label faces and/or search for faces in images. Faces may be detected in the images and visual and non-visual features may be determined for each face. The determined features may be facial and/or contextual features. Features may also be determined from labels assigned to the faces. Pair-wise comparisons of the faces, using the determined features, may determine distances between pairs of faces. Some features may be used to determine the relevance of other features. Dependent on the calculated distances, similar faces may be grouped and suggested labels for faces may be provided. The system may be scalable, processing faces in groups to limit the use of system resources to a certain amount, regardless of the size of a digital image collection.

Claims

1. A method for facial recognition in digital images, comprising:

determining, for each of a plurality of faces in a plurality of digital images, visual and non-visual features;

activating a subset of the plurality of faces, wherein the subset comprises a predetermined number of active faces, wherein at least one of the plurality of faces remains un-active;

comparing the visual and non-visual features for a plurality of pairs of the active faces; and

dependent on said comparing, calculating a distance for each one of the plurality of pairs of active faces, wherein the distance corresponds to a measure of similarity between faces in a pair.

2. The method of claim 1, further comprising:

deactivating at least one of the active faces and activating at least one of the un-active faces to create one or more new active faces; and

performing said comparing and said calculating on the one or more new active faces.

3. The method of claim 1, wherein said calculating the distance for each one of the plurality of pairs of active faces comprises calculating a facial distance between the respective pair of active faces and calculating a clothes distance between the respective pair of active faces, and wherein the distance for the respective pair of active faces is dependent on the facial distance and the clothes distance.

4. The method of claim 1, further comprising:

forming, dependent on said calculating, a group of similar faces, wherein the group is a subset of the set of active faces;

selecting, dependent on said calculating, a representative face from the group of similar faces;

receiving a label corresponding to the representative face; and

determining, from the received label, one or more features for the representative face.

5. The method of claim 4, wherein said determining visual features for each one of the plurality of faces comprises determining a race and an age for each one of the plurality of faces, and further comprising:

determining, dependent on the received label, a race for the representative face and an age for the representative face;

comparing the race for the representative face to the race for each one of the active faces;

comparing the age for the representative face to the age for each one of the active faces; and

dependent on said comparing, calculating a race distance between the representative face and each one of the active faces and calculating an age distance between the representative face and each one of the active faces.

6. The method of claim 4, wherein said determining one or more features for the representative face comprises a trained gender classifier determining, dependent on the received label, a gender for the representative face.

7. The method of claim 6, wherein said determining visual features for each one of the plurality of faces comprises determining a gender for each one of the plurality of faces, and further comprising:

comparing the gender for the representative face to the gender for each one of the active faces; and

dependent on said comparing, calculating a gender distance between the representative face and each one of the active faces.

8. A system, comprising:

a memory; and

one or more processors coupled to the memory, wherein the memory comprises program instructions executable by the one or more processors to implement a face recognition module configured to;

determine, for each of a plurality of faces in a plurality of digital images, visual and non-visual features;

activate a subset of the plurality of faces, wherein the subset comprises a predetermined number of active faces, wherein at least one of the plurality of faces remains un-active;

compare the visual and non-visual features for a plurality of pairs of the active faces; and

dependent on the comparison, calculate a distance for each one of the plurality of pairs of active faces, wherein the distance corresponds to a measure of similarity between faces in a pair.

9. The system of claim 8, wherein the face recognition module is further configured to:

deactivate at least one of the active faces and activate at least one of the un-active faces to create one or more new active faces; and

perform said comparing and said calculating on the one or more new active faces.

10. The system of claim 8, wherein said calculating the distance for each one of the plurality of pairs of active faces comprises calculating a facial distance between the respective pair of active faces and calculating a clothes distance between the respective pair of active faces, and wherein the distance for the respective pair of active faces is dependent on the facial distance and the clothes distance.

11. The system of claim 8, wherein the face recognition module is further configured to:

form, dependent on said calculating, a group of similar faces, wherein the group is a subset of the set of active faces;

select, dependent on said calculating, a representative face from the group of similar faces;

receive a label corresponding to the representative face; and

determine, from the received label, one or more features for the representative face.

12. The system of claim 11, wherein said determining one or more features for the representative face comprises a trained gender classifier determining, dependent on the received label, a gender for the representative face.

13. The system of claim 12, wherein said determining visual features for each one of the plurality of faces comprises determining a gender for each one of the plurality of faces, and wherein the face recognition module is further configured to:

compare the gender for the representative face to the gender for each one of the active faces; and

dependent on the comparison, calculate a gender distance between the representative face and each one of the active faces.

14. A non-transitory computer-readable storage medium storing program instructions executable on a computer to implement a face recognition module configured to:

determine, for each of a plurality of faces in a plurality of digital images, visual and non-visual features;

activate a subset of the plurality of faces, wherein the subset comprises a predetermined number of active faces, wherein at least one of the plurality of faces remains un-active;

compare the visual and non-visual features for a plurality of pairs of the active faces; and

dependent on the comparison, calculate a distance for each of the plurality of pairs of active faces, wherein the distance corresponds to a measure of similarity between faces in a pair.

15. The non-transitory computer-readable storage medium of claim 14, wherein the face recognition module is further configured to:

deactivate at least one of the active faces and activating at least one of the un-active faces to create one or more new active faces; and

perform said comparing and said calculating on the one or more new active faces.

16. The non-transitory computer-readable storage medium of claim 14, wherein said calculating the distance for each one of the plurality of pairs of active faces comprises calculating a facial distance between the respective pair of active faces and calculating a clothes distance between the respective pair of active faces, and wherein the distance for the respective pair of active faces is dependent on the facial distance and the clothes distance.

17. The non-transitory computer-readable storage medium of claim 14, wherein the face recognition module is further configured to:

form, dependent on said calculating, a group of similar faces, wherein the group is a subset of the set of active faces;

select, dependent on said calculating, a representative face from the group of similar faces;

receive a label corresponding to the representative face; and

determine, from the received label, one or more features for the representative face.

18. The non-transitory computer-readable storage medium of claim 17, wherein said determining one or more features for the representative face comprises a trained gender classifier determining, dependent on the received label, a gender for the representative face.

19. The non-transitory computer-readable storage medium of claim 18, wherein said determining visual features for each one of the plurality of faces comprises determining a gender for each one of the plurality of faces, wherein the face recognition module is further configured to:

compare the gender for the representative face to the gender for each one of the active faces; and

dependent on the comparison, calculate a gender distance between the representative face and each one of the active faces.

20. (canceled)

21. A method for facial recognition in digital images, comprising:

detecting a plurality of faces in a plurality of digital images;

determine, for each face detected in the plurality of digital images, visual or non-visual features;

create, based on the determined visual or non-visual features of each face detected in the plurality of digital images, a feature vector representing the determined visual or non-visual features of each face;

activating a subset of the plurality of faces that includes a predetermined number of active faces, at least one of the plurality of faces remaining un-active; and

calculating, based on the feature vectors of the active faces, a distance for each one of a plurality of pairs of the active faces, the distance corresponding to a measure of similarity between the active faces in a pair.