Training an image processing neural network without human selection of features

• US 9,317,779 B2
• Filed: 04/05/2013
• Issued: 04/19/2016
• Est. Priority Date: 04/06/2012
• Status: Active Grant

• Current Assignee: Brigham Young University (The Church of Jesus Christ of Latter-Day Saints)
• Original Assignee: Brigham Young University (The Church of Jesus Christ of Latter-Day Saints)
• Inventors: Lillywhite, Kirt Dwayne, Lee, Dah-Jye
• Primary Examiner(s): Sherali, Ishrat I
• Application Number:
  US13/857,909
• Publication Number:
  US 20130266214A1
• Time in Patent Office: 1,110 Days
• Field of Search: 382/156, 382/159
• US Class Current: 1/1
• CPC Class Codes: G06F 18/21 Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation

Abstract

A method for training an image processing neural network without human selection of features may include providing a training set of images labeled with two or more classifications, providing an image processing toolbox with image transforms that can be applied to the training set, generating a random set of feature extraction pipelines, where each feature extraction pipeline includes a sequence of image transforms randomly selected from the image processing toolbox and randomly selected control parameters associated with the sequence of image transforms. The method may also include coupling a first stage classifier to an output of each feature extraction pipeline and executing a genetic algorithm to conduct genetic modification of each feature extraction pipeline and train each first stage classifier on the training set, and coupling a second stage classifier to each of the first stage classifiers in order to increase classification accuracy.

Claims

1. A method for training an image processing neural network without human selection of features, the method comprising:

providing a training set comprising images labeled with two or more classifications;

providing an image processing toolbox comprising a plurality of image transforms;

generating a random set of feature extraction pipelines, each feature extraction pipeline comprising a sequence of image transforms randomly selected from the image processing toolbox and randomly selected control parameters associated with the sequence of image transforms;

coupling a first stage classifier to an output of each feature extraction pipeline; and

executing a genetic algorithm to conduct genetic modification of each feature extraction pipeline and train each first stage classifier on the training set.

2. The method of claim 1, further comprising coupling each first stage classifier to a second stage classifier and training the second stage classifier.

3. The method of claim 2, wherein training the second stage classifier prunes the feature extraction pipelines to a selected set of feature extraction pipelines.

4. The method of claim 3, further comprising processing images using the selected set of feature extraction pipelines, the first stage classifiers coupled thereto, and the second stage classifier.

5. The method of claim 1, wherein genetic modification comprises pipeline splicing.

6. The method of claim 1, wherein genetic modification comprises mutation of control parameters.

7. The method of claim 1, wherein the first stage classifier is a perceptron.

8. The method of claim 1, wherein the second stage classifier is trained by adaptive boosting.

9. The method of claim 1, further comprising determining control parameters values for an image transform that do not result in processing errors and limiting the randomly selected control parameters and subsequent genetic modifications to those values.

10. An apparatus for processing images, the apparatus comprising:

a storage device configured to store a training set comprising images labeled with two or more classifications;

a processor configured to execute an image processing toolbox comprising a plurality of image transforms;

a plurality of feature extraction pipelines, each feature extraction pipeline comprising a sequence of image transforms; and

a corresponding plurality of first stage classifiers coupled to the plurality of feature extraction pipelines;

wherein the plurality of feature extraction pipelines were generated by;

generating a random set of feature extraction pipelines, each feature extraction pipeline comprising a sequence of image transforms randomly selected from the image processing toolbox and randomly selected control parameters associated with the sequence of image transforms,

coupling a first stage classifier to an output of each feature extraction pipeline, and

executing a genetic algorithm to conduct genetic modification of each feature extraction pipeline and train each first stage classifier on the training set.

11. The apparatus of claim 10, further comprising a second stage classifier coupled to each first stage classifier.

12. The apparatus of claim 10, wherein the second stage classifier is trained by adaptive boosting.

13. The apparatus of claim 10, wherein genetic modification comprises pipeline splicing.

14. The apparatus of claim 10, wherein genetic modification comprises mutation of control parameters.

15. The apparatus of claim 10, wherein the first stage classifier is a perceptron.

16. An apparatus for training an image processing neural network without human selection of features, the apparatus comprising:

a storage device configured to store a training set comprising images labeled with two or more classifications;

a processor configured to execute an image processing toolbox comprising a plurality of image transforms;

the processor configured to generate a random set of feature extraction pipelines, each feature extraction pipeline comprising a sequence of image transforms randomly selected from the image processing toolbox and randomly selected control parameters associated with the sequence of image transforms;

the processor configured to couple a first stage classifier to an output of each feature extraction pipeline; and

the processor configured to execute a genetic algorithm to conduct genetic modification of each feature extraction pipeline and train each first stage classifier on the training set.

17. The apparatus of claim 16, wherein the processor is configured to couple each first stage classifier to a second stage classifier and train the second stage classifier.

18. The apparatus of claim 17, wherein the processor trains the second stage classifier by adaptive boosting.

19. The apparatus of claim 16, wherein genetic modification comprises pipeline splicing or mutation of control parameters.

20. The apparatus of claim 16, wherein the first stage classifier is a perceptron.