Feature selection method and apparatus, and pattern discrimination method and apparatus

Patent metadata

Field	Value
Publication number	US-9697441-B2
Application number	US-201213451817-A
Country	US
Kind code	B2
Filing date	Apr 20, 2012
Priority date	Apr 20, 2011
Publication date	Jul 4, 2017
Grant date	Jul 4, 2017

Abstract

A feature selection apparatus, which selects features to be used to discriminate an object by a discriminator using learning data including the object, extracts a plurality of partial data from the learning data, and obtains discrimination values obtained by controlling the discriminator to process the plurality of extracted partial data as features of the plurality of partial data. The feature selection apparatus evaluates the obtained features based on discrimination degrees on a discrimination space defined by the discriminator, and selects features to be used to discriminate the object from a plurality of features obtained in association with the plurality of partial data based on an evaluation result.

First claim

What is claimed is:

1. A method comprising:
preparing a discriminator to be used for discriminating partial data;
extracting a plurality of partial data from learning data;
extracting at least one local data in each of the plurality of partial data;
classifying the plurality of partial data by the discriminator to obtain a discrimination degree related to a difference between the at least one local data and a decision boundary of the discriminator for each of the plurality of partial data;
evaluating a usefulness degree of the extracted partial data based on the discrimination degree;
selecting at least a portion of the plurality of partial data to be used to be associated with the discriminator from the plurality of partial data based on the usefulness degree; and
recording information of the selected partial data as an example to the discriminator.

2. The method according to claim 1 , wherein a plurality of discriminators are prepared by setting different discrimination rules in the discriminator,
wherein at least one local data is extracted for each of the plurality of discriminators based on the set different discrimination rules.

3. The method according to claim 2 , further comprising evaluating each of the plurality of discriminators based on the usefulness degrees obtained by the plurality of discriminators, and selecting a predetermined number of discriminators from the plurality of discriminators in descending order of evaluation result.

4. The method according to claim 1 , wherein the discriminator includes a plurality of element discriminators, each of which outputs an output value as a result of classification, and
wherein the discriminator obtains, as a discrimination value, a sequence in which a plurality of output values from the plurality of element discriminators are arranged.

5. The method according to claim 4 , wherein in obtaining the discrimination degree,
an individual discrimination degree is calculated for each of the plurality of output values using a classification space defined by the plurality of element discriminators, and
a sum total of the individual discrimination degrees is calculated as the discrimination degree.

6. The method according to claim 4 , wherein the usefulness degree is obtained based on the discrimination degree and uniqueness of the discrimination value.

7. The method according to claim 4 , wherein the plurality of element discriminators calculate intermediate processing results by applying predetermined processing to a partial data extracted from the learning data, and the discrimination degree is obtained by calculating a distance between an origin of a discrimination space and a point indicating the intermediate processing results in the discrimination space, wherein the discrimination space is a space which expresses the intermediate processing results.

8. The method according to claim 1 , wherein the discriminator is a linear binary discriminator which outputs a discrimination value based on an inner product of a base vector and an input vector generated based on the partial data extracted from the learning data,
wherein, the discrimination value is obtained based on an average value of a plurality of inner product values obtained by applying the discriminator to the plurality of partial data, and
wherein, in obtaining the discrimination degree, a distribution of the plurality of inner product values is estimated on a discrimination space defined by the discriminator, and the discrimination degree is calculated by calculating a relationship between a boundary where a discrimination value changes and the average value on the discrimination space based on the estimated distribution.

9. The method according to claim 1 , further comprising associating a relative position of the selected partial data with respect to a predetermined position in the learning data.

10. The method according to claim 9 , wherein the predetermined position corresponds to a position of an object to be detected.

11. The method according to claim 9 , wherein the learning data includes an object to be detected, and the relative position of the selected partial data includes information indicating an orientation of the object in the learning data.

12. The method according to claim 1 , wherein the usefulness degree is calculated based on discrimination values and the discrimination degrees obtained by applying the discriminator to the partial data.

13. The method according to claim 1 , wherein the learning data is a learning image, and the partial data is a partial image, and the local data comprises pixel values extracted from the partial image.

14. The method according to claim 13 , wherein the discrimination degree is a value related to a distance between the pixel value and a discrimination boundary line.

15. The method according to claim 1 , wherein the usefulness degree is obtained based on the discrimination degree and a number of partial data that provides the same discrimination result.

16. A pattern discrimination method for discriminating an object in data, the method comprising:
obtaining discrimination values by applying a discriminator, which has been trained using partial data selected by a training method, to a plurality of partial data obtained from the data; and
classifying the data to a predetermined class based on the obtained discrimination values,
wherein the training method comprises:
extracting a plurality of partial data from a learning data;
extracting at least one local data in each of the plurality of partial data;
classifying the plurality of partial data by the discriminator to obtain a discrimination degree related to a difference between the at least one local data and a decision boundary of the discriminator for each of the plurality of partial data;
evaluating a usefulness degree of the extracted partial data based on the discrimination degree;
selecting at least a portion of the plurality of partial data to be used to be associated with the discriminator from the plurality of partial data based on the usefulness degree; and
recording information of the selected partial data as an example to the discriminator.

17. An apparatus comprising:
a processor; and
a memory storing instructions that, when executed by the processor, cause the processor to:
prepare a discriminator to be used for discriminating partial data;
extract a plurality of partial data from a learning data;
extract at least one local data in each of the plurality of partial data;
classify the plurality of partial data by the discriminator to obtain a discrimination degree related to a difference between the at least one local data and a decision boundary of the discriminator for each of the plurality of partial data;
evaluate a usefulness degree of the extracted partial data based on the discrimination degree;
select at least a portion of the plurality of partial data to be used to be associated with the discriminator from the plurality of partial data based on the usefulness degree; and
record information of the selected partial data as an example to the discriminator.

18. A pattern discrimination apparatus for discriminating an object in data, the pattern discrimination apparatus comprising:
a first processor; and
a first memory storing instructions that, when executed by the first processor, cause the first processor to:
obtain discrimination values by applying a discriminator, which has been trained using partial data selected by a training method, to a plurality of partial data obtained from the

Assignees

• Mitarai Yusuke
• Matsugu Masakazu
• Canon Kk

Inventors

• Mitarai Yusuke
• Matsugu Masakazu

Classifications

• G06F18/2115 Primary

  by evaluating different subsets according to an optimisation criterion, e.g. class separability, forward selection or backward elimination · CPC title

• G06F18/245

  relating to the decision surface · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06K9/6231 Primary

  Physics · mapped topic

• G06K9/6285

  Physics · mapped topic