Methods and logic for autonomous generation of ensemble classifiers, and systems incorporating ensemble classifiers

• US 8,582,871 B2
• Filed: 10/06/2010
• Issued: 11/12/2013
• Est. Priority Date: 10/06/2009
• Status: Expired due to Fees

• Current Assignee: Wright State University
• Original Assignee: Wright State University
• Inventors: Skipper, Julie A., Ganapathy, Priya
• Primary Examiner(s): AHMED, SAMIR ANWAR
• Application Number:
  US12/898,916
• Publication Number:
  US 20110081073A1
• Time in Patent Office: 1,133 Days
• Field of Search: 706/12, 706/14, 706/15, 706/45, 706/48, 706/56
• US Class Current: 382/159
• CPC Class Codes: G06F 18/217 Validation; Performance evaluation; Active pattern learning techniques
• CPC Class Codes: G06F 18/2415 based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data
• CPC Class Codes: G06F 18/257 Belief theory, e.g. Dempster-Shafer
• CPC Class Codes: G06F 18/285 Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
• CPC Class Codes: G06V 10/255 Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/776 Validation; Performance evaluation
• CPC Class Codes: G06V 10/809 of classification results, e.g. where the classifiers operate on the same input data
• CPC Class Codes: G06V 10/814 using belief theory, e.g. Dempster-Shafer
• CPC Class Codes: G06V 10/87 using selection of the recognition techniques, e.g. of a classifier in a multiple classifier system

Abstract

In one embodiment, a method for generating an ensemble classifier may include transforming multidimensional training data into a plurality of response planes. Each of the response planes includes a set of confidence scores. The response planes are transformed into a plurality of binary response planes. Each of the binary response planes include a set of binary scores corresponding to one of the confidence scores. Combinations of the binary response planes are transformed into sets of diversity metrics according to a diversity measure. A metric is selected from the sets of diversity metrics. A predicted performance of a child combination of the recognition algorithms corresponding to the combinations is generated. The predicted performance is based at least in part upon the metrics. Parent recognition algorithms are selected from the recognition algorithms based at least in part upon the predicted performance. The ensemble classifier is generated and includes the parent recognition algorithms.

Claims

1. A method for generating an ensemble classifier, the method comprising:

transforming, automatically with a processor, multidimensional training data into a plurality of response planes according to a plurality of recognition algorithms, wherein each of the response planes comprise a set of confidence scores;

transforming, automatically with the processor, the response planes into a plurality of binary response planes, wherein each of the binary response planes comprise a set of binary scores corresponding to one of the confidence scores;

transforming, automatically with the processor, a first combination of the binary response planes into a first set of diversity metrics according to a first diversity measure;

transforming, automatically with the processor, a second combination of the binary response planes into a second set of diversity metrics according to a second diversity measure;

selecting a first metric from the first set of diversity metrics;

selecting a second metric from the second set of diversity metrics;

generating, automatically with the processor, a predicted performance of a child combination of the recognition algorithms corresponding to the first combination and the second combination, wherein the predicted performance is based at least in part upon the first metric and the second metric;

selecting parent recognition algorithms from the recognition algorithms based at least in part upon the predicted performance; and

generating the ensemble classifier, wherein the ensemble classifier comprises the parent recognition algorithms.

2. The method of claim 1, further comprising:

identifying a target square object within the multidimensional training data, wherein the multidimensional training data is two-dimensional;

dividing the multidimensional training data into grid squares;

generating a true negative template image, the true negative template image comprising the grid squares excluding the target square object; and

generating a true positive template image, the true positive template image comprising the target square object.

3. The method of claim 2, further comprising:

identifying additional target square objects within the multidimensional training data; and

optimizing, automatically with the processor, a grid square dimension, wherein a starting point for optimizing is an average dimension of the target square object and additional target square objects.

4. The method of claim 2, further comprising calculating a sensitivity, wherein the sensitivity is based at least in part upon an overlap between one of the binary response planes and the true positive template image.

5. The method of claim 2, further comprising calculating a specificity, wherein the specificity is based at least in part upon a total number of the grid squares and an overlap between one of the binary response planes and the true negative template image.

6. The method of claim 1, further comprising applying weights to the recognition algorithms of the ensemble classifier, wherein the ensemble classifier is less dependent upon an underlying fusion scheme.

7. The method of claim 6, wherein the weights are exponential values and the weights are applied automatically by the processor.

8. The method of claim 1, further comprising quantifying a cost function of the ensemble classifier.

9. The method of claim 8, wherein the cost function is based upon a false negative, a false positive, a target likelihood, a non-target likelihood, or a combination thereof.

10. The method of claim 1, wherein the predicted performance is an output from a quadratic model.

11. The method of claim 1, wherein the first diversity measure is a double fault dependency metric.

12. The method of claim 11, wherein the first metric is a minimum of the first set of diversity metrics.

13. The method of claim 12, further comprising storing a threshold value in a memory, wherein one of the confidence scores is transformed to one of the binary scores by a thresholding operation based at least in part upon the threshold value and the threshold value corresponds to the first metric.

14. The method of claim 1, wherein:

the first metric is selected by an optimizer executed by the processor;

the first combination of the binary response planes comprises a first binary response plane and a second binary response plane;

one of the response planes is transformed to the first binary response plane based at least in part upon a first threshold value;

one of the response planes is transformed to the second binary response plane based at least in part upon a second threshold value; and

a starting threshold utilized by the optimizer is equal to an average of the first threshold value and the second threshold value.

15. The method of claim 1, wherein:

wherein the first metric is selected by an optimizer executed by the processor;

the first combination of the binary response planes comprises a first binary response plane and a second binary response plane;

one of the response planes is transformed to the first binary response plane based at least in part upon a first threshold value;

one of the response planes is transformed to the second binary response plane based at least in part upon a second threshold value; and

a starting threshold utilized by the optimizer is equal to a product of the first threshold value and the second threshold value.

16. The method of claim 1, wherein:

the first metric is selected by an optimizer executed by the processor;

the first combination of the binary response planes comprises a first binary response plane and a second binary response plane;

one of the response planes is transformed to the first binary response plane based at least in part upon a first threshold value;

one of the response planes is transformed to the second binary response plane based at least in part upon a second threshold value; and

a starting threshold utilized by the optimizer is equal to a ratio of the first threshold value and the second threshold value, such that a largest of the first threshold value and the second threshold value is a denominator.

17. The method of claim 1, further comprising ranking the response planes according to an area under a receiver-operator characteristic curve.