Clustering images

Patent metadata

Field	Value
Publication number	US-9317783-B2
Application number	US-201013701632-A
Country	US
Kind code	B2
Filing date	Jun 1, 2010
Priority date	Jun 1, 2010
Publication date	Apr 19, 2016
Grant date	Apr 19, 2016

Abstract

A method and system for clustering images, comprising detecting faces in respective images to generate face data, generating a first cluster of images using the face data, the first cluster comprising images with a common face and representing a corresponding person, generating a second cluster of images using clothes data representing a set of clothing signatures, the second cluster comprising images with a common set of clothes, and using the first and second clusters to link an image of the person to the clusters in order to generate a third cluster for the person.

First claim

What is claimed is:

1. A method for clustering images, comprising:
detecting faces in respective images to generate face data;
generating a first cluster of images using the face data, the first cluster comprising images with a common face and representing a corresponding person;
generating a set of clothing signatures in which regions of the face data corresponding to the faces and hair are disregarded;
generating a second cluster of images using clothes data representing the set of clothing signatures, the second cluster comprising images with a common set of clothes, the second cluster separately clustering the images as compared to the first cluster according to clothing worn by people in the images irrespective of the people wearing the clothing; and
using the first and second clusters to link an image of the person to the clusters in order to generate a third cluster for the person,
wherein a clothing signature is generated by:
using the face data, defining respective facial regions for identified faces, and, on the basis of the facial regions, defining respective corresponding regions for hair and clothing;
using image pixel data within the facial and hair regions, generating a measure of the skin and hair tones for the person; and
using the measure to discard those regions of the image corresponding to skin and hair within the clothing region in order to provide a clothing mask representing a region of clothing for the person.

2. The method as claimed in claim 1 , further comprising:
segmenting the facial, clothing and hair regions into relatively smaller regions; and discarding any of the smaller regions representing background or overlapping areas.

3. The method as claimed in claim 1 , further comprising:
generating a distance matrix in which each item of the matrix represents a dissimilarity between two pieces of clothing.

4. The method as claimed in claimed 3 , further comprising:
using the distance matrix, grouping clothes pieces into multiple clothing clusters using average-link hierarchical clustering.

5. An image processing system for clustering a set of images comprising:
a processor;
a non-transitory computer-readable medium storing computer code executable by the processor to:
detect faces in respective images in the set to generate face data;
generate a first cluster of images using the face data, the first cluster comprising images with a common face and representing a corresponding person;
generate a set of clothing signatures in which regions of the face data corresponding to the faces and hair are disregarded;
generate a second cluster of images using clothes data representing the set of clothing signatures, the second cluster comprising images with a common set of clothes, the second cluster separately clustering the images as compared to the first cluster according to clothing worn by people in the images irrespective of the people wearing the clothing; and
use the first and second clusters to link an image of the person to the clusters in order to generate a third cluster for the person,

wherein a clothing signature is generated by:
using the face data defining respective facial regions for identified faces, and, on the basis of the facial regions, defining respective corresponding regions for hair and clothing;
using image pixel data within the facial and hair regions, generating a measure of the skin and hair tones for the person; and
using the measure to discard those regions of the image corresponding to skin and hair within the clothing region in order to provide a clothing mask representing a region of clothing for the person.

6. A non-transitory computer-readable data storage medium storing computer-executable code executable by a processor to perform a method comprising:
detecting faces in respective images to generate face data;
generating a first cluster of images using the face data, the first cluster comprising images with a common face and representing a corresponding person;
generating a second cluster of images using clothes data representing a set of clothing signatures, the second cluster comprising images with a common set of clothes, the second cluster separately clustering the images as compared to the first cluster according to clothing worn by people in the images irrespective of the people wearing the clothing; and
using the first and second clusters to link an image of the person to the clusters in order to generate a third cluster for the person,
wherein a clothing signature is generated by:
using the face data, defining respective facial regions for identified faces, and, on the basis of the facial regions, defining respective corresponding regions for hair and clothing;
using image pixel data within the facial and hair regions, generating a measure of the skin and hair tones for the person; and
using the measure to discard those regions of the image corresponding to skin and hair within the clothing region in order to provide a clothing mask representing a region of clothing for the person.

7. The non-transitory computer-readable data storage medium of claim 6 , wherein the method further comprises:
segmenting the facial, clothing and hair regions into relatively smaller regions; and
discarding any of the smaller regions representing background or overlapping areas.

8. The non-transitory computer-readable data storage medium of claim 6 , wherein generating a measure of the skin and hair tones comprises clustering pixels in the skin and hair regions using a Gaussian Mixture Model.

9. The non-transitory computer-readable data storage medium of claim 6 , wherein generating a clothing signature comprising clustering pixels in the clothing segment using a Gaussian Mixture Model.

10. The non-transitory computer-readable data storage medium of claim 6 , wherein the Gaussian Mixture Model uses a pixel value in the CIELAB colour space.

11. The non-transitory computer-readable data storage medium of claim 6 , wherein the method further comprises:
generating a distance matrix in which each item of the matrix represents a dissimilarity between two pieces of clothing.

12. The non-transitory computer-readable data storage medium of claim 11 , wherein the method further comprises:
using the distance matrix, grouping clothes pieces into multiple clothing clusters using average-link hierarchical clustering.

Assignees

• Zhang Wei
• Zhang Tong
• Tretter Daniel R
• Hewlett Packard Development Co

Inventors

• Zhang Wei
• Zhang Tong
• Tretter Daniel R

Classifications

• G06F18/24

  Classification techniques · CPC title

• G06V40/162 Primary

  using pixel segmentation or colour matching · CPC title

• G06K9/00362

  Physics · mapped topic

• G06K9/00234

  Physics · mapped topic

• G06K9/6267 Primary

  Physics · mapped topic