System and method for improved classification

Patent metadata

Field	Value
Publication number	US-9639780-B2
Application number	US-34158708-A
Country	US
Kind code	B2
Filing date	Dec 22, 2008
Priority date	Dec 22, 2008
Publication date	May 2, 2017
Grant date	May 2, 2017

Abstract

A system and method for improved classification. A first classifier is trained using a first process running on at least one computing device using a first set of training images relating to a class of images. A set of additional images are selected using the first classifier from a source of additional images accessible to the computing device. The first set of training images and the set of additional images are merged using the computing device to create a second set of training images. A second classifier is trained using a second process running on the computing device using the second set of training images. A set of unclassified images are classified using the second classifier thereby creating a set of classified images. The first classifier and the second classifier employ different classification methods.

First claim

We claim:

1. A method comprising:
training, using a first process running on a computing device, a first classifier using a first set of training images relating to a categorical class of images, said first classifier being a type associated with said categorical class;
selecting, using the first classifier on the computing device, a set of additional images from a source of additional images accessible to the computing device;
merging, using the computing device, the first set of training images and the set of additional images to create a second set of training images;
training, using a second process running on the computing device, a second classifier using the second set of training images, said second classifier being a type different from said first classifier and based upon a categorical class of said second set of training images, the first process and the second process being different classification methods that are based on the type of each classifier;
classifying, using the computing device, a set of unclassified images using the second classifier thereby creating a set of classified images.

2. The method of claim 1 wherein the set of additional images selected by the first classifier uses a selection method that provides a high confidence that the selected set of additional images are similar to the training set.

3. The method of claim 2 wherein said selection comprises a nearest neighbor search.

4. The method of claim 1 further comprising:
filtering, using the computing device, the additional images using a filtering method that increases the likelihood that the images will contain a higher percentage of images of the image class, wherein the filtered additional images are used by the selecting step to select the set of additional images.

5. The method of claim 1 wherein the training images, the additional images, and the unclassified images are preprocessed using a preprocessing method to place the images to a format that increases the efficiency of the first classifier and the second classifier.

6. The method of claim 5 wherein the preprocessing method reduces each image to a small vector of dimensions using a shift invariant sparse feature extractor.

7. The method of claim 1 wherein the classified image are stored on a classified image database accessible to an online service which provides image search services to end users.

8. The method of claim 7 wherein the online service is a web-based image search service, wherein when a user submits an image search request to the image search service requesting images relating to the class of images, the image search service uses the classified image database to identify images to be returned to the end user for display on an end user system.

9. The method of claim 1 wherein the set of additional images and the set of unclassified images are selected from a plurality of image sources accessible via the Internet.

10. The method of claim 7 wherein the classified image database comprises an index of images accessible via the Internet from a plurality of image sources.

11. The method of claim 1 wherein the method is adapted to classify video files.

12. The method of claim 1 wherein the method is adapted to classify web pages.

13. The method of claim 1 wherein the method is adapted to classify documents.

14. The method of claim 1 wherein the computing device comprises at least two computing devices wherein the first process and the second process run on different computing devices.

15. A system comprising:
a processor;
a non-transitory storage medium for tangibly storing thereon program logic for execution by the processor, the program logic comprising:
logic executed by the processor for training, using a first process, a first classifier using a first set of training images relating to a categorical class of images, said first classifier being a type associated with said categorical class;
logic executed by the processor for selecting, using the first classifier, a set of additional images from a source of additional images;
logic executed by the processor for merging the first set of training images and the set of additional images to create a second set of training images;
logic executed by the processor for training, using a second process, a second classifier using the second set of training images, said second classifier being a type different from said first classifier and based upon a categorical class of said second set of training images, the first process and the second process being different classification methods that are based on the type of each classifier;
logic executed by the processor for classifying, a set of unclassified images using the second classifier thereby creating a set of classified images.

16. The system of claim 15 wherein the set of additional images selected by the first classifier uses a selection method that provides a high confidence that the selected set of additional images are similar to the training set.

17. The system of claim 14 wherein said selection comprises a nearest neighbor search.

18. The system of claim 15 logic executed by the processor for filtering the additional images using a filtering method that increases the likelihood that the images will contain a higher percentage of images of the image class.

19. The method of claim 15 logic executed by the processor for preprocessing the training images, the additional images, and the unclassified images using a preprocessing method to place the images to a format that increases the efficiency of the first classifier and the second classifier.

20. The method of claim 19 wherein the preprocessing method reduces each image to a small vector of dimensions using a shift invariant sparse feature extractor.

21. The method of claim 15 wherein the system is adapted to classify video files.

22. The method of claim 15 wherein the system is adapted to classify web pages.

23. The method of claim 15 wherein the system is adapted to classify documents.

24. The system of claim 15 wherein the set of classified images are stored on a classified image database is accessible, to an online service which provides image search services to end users.

25. The system of claim 15 wherein the computing device processor comprises at least two processors computing devices wherein the first classifier and the second classifier run on different processors.

26. A non-transitory computer-readable storage medium tangibly storing thereon computer-executable instructions, that when executed by a processor associated with a computing device, perform a method comprising:
training, via a first process, a first classifier using a first set of training images relating to a categorical class of images, said first classifier being a type associated with said categorical class;
selecting, using the first classifier, a set of additional images from a source of additional images accessible to the computing device;
merging the first set of training images and the set of additional images to create a second set of training images;
training, via a second process, a second classifier using the second set of training images, said second classifier being a type different from said first classifier and based upon a categorical class of said second set of training images, the first process and the second process being different classification methods that are base

Assignees

• Ranzato Marc Aurelio
• Weinberger Kilian Quirin
• Hoerster Eva
• Slaney Malcolm
• Excalibur Ip Llc

Inventors

• Ranzato Marc Aurelio
• Weinberger Kilian Quirin
• Hoerster Eva
• Slaney Malcolm

Classifications

• G06F18/214 Primary

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06K9/6256 Primary

  Physics · mapped topic