Grouping face images using statistic distribution estimate

Patent metadata

Field	Value
Publication number	US-2016171284-A1
Application number	US-201615050189-A
Country	US
Kind code	A1
Filing date	Feb 22, 2016
Priority date	May 8, 2014
Publication date	Jun 16, 2016
Grant date	—

Abstract

A computer-implemented method for sorting face images of different individuals into different groups includes obtaining face images comprising faces of unknown individuals by a computer processor; calculating similarity functions between pairs of face images by the computer processor; joining face images that have values of the similarity functions above a predetermined threshold into a hypothetical face group, wherein the face images in the hypothetical face group hypothetically belong to a same person; conducting non-negative matrix factorization on values of the similarity functions in the hypothetical face group to test truthfulness of the hypothetical face group; and identifying the hypothetical face group as a true face group if a percentage of the associated similarity functions being true is above a threshold based on the non-negative matrix factorization.

First claim

What is claimed is:

1 . A computer-implemented method for recognizing face images, comprising:
storing face models or face images of a known person as training faces in a computer storage; joining a group of testing face images with a group of training faces that belong to the known person to form a joint face group; calculating similarity functions, by a computer processor, between pairs of testing face images and training faces in the joint face group; conducting non-negative matrix factorization on values of the similarity functions in the joint face group to test truthfulness of the joint face group; and identifying the testing face images to belong to the known persons if a percentage of the associated similarity functions being true in the joint face group is above a threshold based on the non-negative matrix factorization.

2 . The computer-implemented method of claim 1 , further comprising:
merging the testing face images with the training faces of the known person to form a new set of training faces for the known person.

3 . The computer-implemented method of claim 1 , wherein the step of conducting non-negative matrix factorization comprises:
forming a non-negative matrix using values of similarity functions between all different pairs of testing face images and training faces in the joint face group, wherein the non-negative matrix factorization is conducted over the non-negative matrix.

4 . The computer-implemented method of claim 1 , wherein the similarity functions in the joint face group are described in a similarity distribution function, wherein the step of conducting non-negative matrix factorization outputs a True similarity distribution function and a False similarity distribution function.

5 . The computer-implemented method of claim 4 , wherein the step of identifying the testing face images to belong to the known persons comprises:
comparing the similarity distribution function to the True similarity distribution function and the False similarity distribution function.

6 . The computer-implemented method of claim 1 , further comprising:
detecting testing faces in images; and cropping portions of the images to produce the testing face images comprising the testing faces of the unknown individuals.

Assignees

• Shutterfly Inc

Inventors

• Sandler Roman
• Kenis Alexander M

Classifications

• G06V10/763

  Non-hierarchical techniques, e.g. based on statistics of modelling distributions · CPC title

• G06V40/161

  Detection; Localisation; Normalisation · CPC title

• G06V40/172 Primary

  Classification, e.g. identification · CPC title

• G06F18/22

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

• G06F18/24

  Classification techniques · CPC title