Decoy-based matching system for facial recognition

Patent metadata

Field	Value
Publication number	US-9977950-B2
Application number	US-201615008186-A
Country	US
Kind code	B2
Filing date	Jan 27, 2016
Priority date	Jan 27, 2016
Publication date	May 22, 2018
Grant date	May 22, 2018

Abstract

Techniques are provided for facial recognition using decoy-based matching of facial image features. An example method may include comparing extracted facial features of an input image, provided for recognition, to facial features of each of one or more images in a gallery of known faces, to select a closest gallery image. The method may also include calculating a first distance between the input image and the selected gallery image. The method may further include comparing the facial features of the input image to facial features of each of one or more images in a set of decoy faces, to select a closest decoy image and calculating a second distance between the input image and the selected decoy image. The method may further include recognizing a match between the input image and the selected gallery image based on a comparison of the first distance and the second distance.

First claim

What is claimed is:

1. A processor-implemented method for facial recognition, the method comprising:
comparing, by a processor, extracted facial features of an input image, provided for recognition, to facial features of each of one or more images in a gallery of known faces, to select a closest gallery image, wherein the known faces are associated with potential recognition matches to the input image;
calculating, by the processor, a first distance between the input image and the selected gallery image;
comparing, by the processor, the facial features of the input image to facial features of each of one or more images in a set of decoy faces, to select a closest decoy image, wherein the decoy faces are associated with persons that should not generate recognition matches to the input image;
calculating, by the processor, a second distance between the input image and the selected decoy image; and
recognizing, by the processor, a match between the input image and the selected gallery image based on a comparison of the first distance and the second distance, wherein the recognizing includes
rejecting the match if the second distance is less than the first distance,
accepting the match if the first distance is less than a first selected threshold value, and
accepting the match if a ratio of the first distance to the second distance is less than a second selected threshold value, even if the first distance is greater than the first selected threshold value.

2. The method of claim 1 , wherein the selecting of the closest gallery image further comprises searching for a minimum distance between facial features of the input image and facial features of each of the one or more gallery images; and the selecting of the closest decoy image further comprises searching for a minimum distance between facial features of the input image and facial features of each of the one or more decoy images.

3. The method of claim 1 , further comprising generating the set of decoy faces by:
collecting decoy facial images;
clustering the decoy facial images into groups of similar faces based on distances between facial features associated with the decoy faces;
selecting a representative face from each the groups; and
compiling the representative faces into the set of decoy faces.

4. The method of claim 1 , wherein the facial features of the input image are extracted based on at least one of Local Binary Pattern Classification (LBP), Modified Census Transformation (MCT), Principal Component Analysis (PCA), Shift Invariant Feature Transformation (SIFT), Speeded Up Robust Feature Detection (SURF) or a Learning-Based Descriptor.

5. The method of claim 1 , wherein the calculations of the first distance and the second distance are based on at least one of Euclidean Distance, Manhattan Distance, Squared Euclidean Distance, Angel-Based Distance, Correlation Coefficient-Based Distance, Mahalanobis Distance or Chi-Squared Distance.

6. A system for facial recognition, the system comprising:
a feature extraction circuit to extract facial features of an input image provided for recognition;
a gallery ranking circuit to compare the extracted facial features of the input image to facial features of each of one or more images in a gallery of known faces, to select a closest gallery image, wherein the known faces are associated with potential recognition matches to the input image; and to calculate a first distance between the input image and the selected gallery image;
a decoy ranking circuit to compare the extracted facial features of the input image to facial features of each of one or more images in a set of decoy faces, to select a closest decoy image, wherein the decoy faces are associated with persons that should not generate recognition matches to the input image; and to calculate a second distance between the input image and the selected decoy image; and
a verification circuit to verify a match between the input image and the selected gallery image based on a comparison of the first distance and the second distance, wherein the verification includes
rejecting the match if second distance is less than the first distance,
verifying the match if the first distance is less than a first selected threshold value, and
verifying the match if a ratio of the first distance to the second distance is less than a second selected threshold value, even if the first distance is greater than the first selected threshold value.

7. The system of claim 6 , wherein the gallery ranking circuit is further to select the closest gallery image based on a search for a minimum distance between facial features of the input image and facial features of each of the one or more gallery images; and the decoy ranking circuit is further to select the closest decoy image based on a search for a minimum distance between facial features of the input image and facial features of each of the one or more decoy images.

8. The system of claim 6 , further comprising a decoy set generation module to:
collect decoy facial images;
cluster the decoy facial images into groups of similar faces based on distances between facial features associated with the decoy faces;
select a representative face from each the groups; and
compile the representative faces into the set of decoy faces.

9. The system of claim 6 , wherein the feature extraction circuit is further to extract facial features of the input image based on at least one of Local Binary Pattern Classification (LBP), Modified Census Transformation (MCT), Principal Component Analysis (PCA), Shift Invariant Feature Transformation (SIFT), Speeded Up Robust Feature Detection (SURF) or a Learning-Based Descriptor.

10. The system of claim 6 , wherein the calculations of the first distance and the second distance are based on at least one of Euclidean Distance, Manhattan Distance, Squared Euclidean Distance, Angel-Based Distance, Correlation Coefficient-Based Distance, Mahalanobis Distance or Chi-Squared Distance.

11. At least one non-transitory computer readable storage medium having instructions encoded thereon that, when executed by one or more processors, result in the following operations for facial recognition, the operations comprising:
comparing extracted facial features of an input image, provided for recognition, to facial features of each of one or more images in a gallery of known faces, to select a closest gallery image, wherein the known faces are associated with potential recognition matches to the input image;
calculating a first distance between the input image and the selected gallery image;
comparing the facial features of the input image to facial features of each of one or more images in a set of decoy faces, to select a closest decoy image, wherein the decoy faces are associated with persons that should not generate recognition matches to the input image;
calculating a second distance between the input image and the selected decoy image; and
recognizing a match between the input image and the selected gallery image based on a comparison of the first distance and the second distance, wherein the recognizing includes
rejecting the match if the second distance is less than the first distance,
accepting the match if the first distance is less than a first selected threshold value, and
accepting the match if a ratio of the first distance to the second distance is less than a second selected threshold value, even if the first distance is greater than the first selected threshold value.

12. The computer readable storage medium of claim 11 , wherein the selecting of the closest gallery image further comprises searching for a minimum distance between facial features of

Assignees

• Intel Corp

Inventors

• Lee Hyungsoo
• Cheon Yeongjae
• Lee Sumin
• Park Minje

Classifications

• G06V40/172

  Classification, e.g. identification · CPC title

• G06F16/5838 Primary

  using colour · CPC title

• G06F18/22

  Matching criteria, e.g. proximity measures · CPC title

• G06F17/30256

  Physics · mapped topic

• G06K9/6215

  Physics · mapped topic