LEARNING DEVICE, LEARNING METHOD, AND PROGRAM

• US 20120300980A1
• Filed: 05/14/2012
• Published: 11/29/2012
• Est. Priority Date: 05/23/2011
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: YOKONO, Jun
• Granted Patent: US 9,053,358 B2
• Time in Patent Office: Days
• US Class Current: 382/103
• CPC Class Codes: G06F 18/2148 characterised by the process organisation or structure, e.g. boosting cascade
• CPC Class Codes: G06F 18/231 Hierarchical techniques, i.e. dividing or merging pattern sets so as to obtain a dendrogram
• CPC Class Codes: G06F 18/2413 based on distances to training or reference patterns
• CPC Class Codes: G06V 10/7625 Hierarchical techniques, i.e. dividing or merging patterns to obtain a tree-like representation; Dendograms
• CPC Class Codes: G06V 10/7747 Organisation of the process, e.g. bagging or boosting
• CPC Class Codes: G06V 40/113 Recognition of static hand signs

Abstract

Disclosed is a learning device. A feature-quantity calculation unit extracts a feature quantity from each feature point of a learning image. An acquisition unit acquires a classifier already obtained by learning as a transfer classifier. A classifier generation unit substitutes feature quantities into weak classifiers constituting the transfer classifier, calculates error rates of the weak classifiers on the basis of classification results of the weak classifiers and a weight of the learning image, and iterates a process of selecting a weak classifier of which the error rate is minimized a plurality of times. In addition, the classifier generation unit generates a classifier for detecting a detection target by linearly coupling a plurality of selected weak classifiers.

Claims

1. A learning device comprising:

a feature-quantity extraction unit for extracting a feature quantity from a feature point of a learning image with respect to each of a plurality of learning images including a learning image including a detection target and a learning image not including the detection target;

a weak-classification calculation unit for calculating a classification result of the detection target according to a weak classifier for every learning image by substituting the feature quantity corresponding to the weak classifier into the weak classifier with respect to each of a plurality of weak classifiers constituting a transfer classifier, which is a classifier for detecting the detection target obtained by statistical learning; and

a classifier generation unit for generating the classifier for detecting the detection target using the weak classifier selected from the plurality of weak classifiers on the basis of the classification result.

2. The learning device according to claim 1, further comprising:

a weight setting unit for setting a weight of the learning image based on the classification result; and

an error-rate calculation unit for calculating an error rate of the weak classifier based on the classification result of each learning image according to the weak classifier and the weight,

wherein the classifier generation unit selects the weak classifier based on the error rate.

3. The learning device according to claim 2, wherein the classifier generated by the classifier generation unit is used for multi-class object recognition.

4. The learning device according to claim 3, wherein:

the classifier generated by the classifier generation unit is a classifier constituting a classifier of a tree structure, and

the transfer classifier is a classifier constituting a leaf of the classifier of the tree structure.

5. A learning method for use in a learning device including a feature-quantity extraction unit for extracting a feature quantity from a feature point of a learning image with respect to each of a plurality of learning images including a learning image including a detection target and a learning image not including the detection target, a weak-classification calculation unit for calculating a classification result of the detection target according to a weak classifier for every learning image by substituting the feature quantity corresponding to the weak classifier into the weak classifier with respect to each of a plurality of weak classifiers constituting a transfer classifier, which is a classifier for detecting the detection target obtained by statistical learning, and a classifier generation unit for generating the classifier for detecting the detection target using the weak classifier selected from the plurality of weak classifiers on the basis of the classification result, the learning method comprising:

extracting, by the feature-quantity extraction unit, the feature quantity from the learning image;

calculating, by the weak-classification calculation unit, the classification result; and

generating, by the classifier generation unit, the classifier.

6. A program for causing a computer to execute:

extracting a feature quantity from a feature point of a learning image with respect to each of a plurality of learning images including a learning image including a detection target and a learning image not including the detection target;

calculating a classification result of the detection target according to a weak classifier for every learning image by substituting the feature quantity corresponding to the weak classifier into the weak classifier with respect to each of a plurality of weak classifiers constituting a transfer classifier, which is a classifier for detecting the detection target obtained by statistical learning; and

generating the classifier for detecting the detection target using the weak classifier selected from the plurality of weak classifiers on the basis of the classification result.

7. A learning device comprising:

a feature-quantity extraction unit for extracting a feature quantity from a feature point of a learning image with respect to each of a plurality of learning images including a learning image including a detection target and a learning image not including the detection target;

a weak-classifier setting unit for generating a weak classifier based on the feature quantity corresponding to a transfer weak classifier constituting a transfer classifier, which is a classifier for detecting the detection target obtained by statistical learning, among feature quantities extracted from the learning image and the learning image;

a weak-classification calculation unit for calculating a classification result of the detection target according to the weak classifier for every learning image by substituting the feature quantity corresponding to the weak classifier into the weak classifier; and

a classifier generation unit for generating the classifier for detecting the detection target using the weak classifier selected from the plurality of weak classifiers on the basis of the classification result.

8. The learning device according to claim 7, further comprising:

a weight setting unit for setting a weight of the learning image based on the classification result; and

an error-rate calculation unit for calculating an error rate of the weak classifier based on the classification result of each learning image according to the weak classifier and the weight,

wherein the classifier generation unit selects the weak classifier based on the error rate.

9. The learning device according to claim 8, wherein the classifier generated by the classifier generation unit is used for multi-class object recognition.

10. The learning device according to claim 9, wherein:

the classifier generated by the classifier generation unit is a classifier constituting a classifier of a tree structure, and

the transfer classifier is a classifier constituting a leaf of the classifier of the tree structure.

11. A learning method for use in a learning device including a feature-quantity extraction unit for extracting a feature quantity from a feature point of a learning image with respect to each of a plurality of learning images including a learning image including a detection target and a learning image not including the detection target, a weak-classifier setting unit for generating a weak classifier based on the feature quantity corresponding to a transfer weak classifier constituting a transfer classifier, which is a classifier for detecting the detection target obtained by statistical learning, among feature quantities extracted from the learning image and the learning image, a weak-classification calculation unit for calculating a classification result of the detection target according to the weak classifier for every learning image by substituting the feature quantity corresponding to the weak classifier into the weak classifier, and a classifier generation unit for generating the classifier for detecting the detection target using the weak classifier selected from the plurality of weak classifiers on the basis of the classification result, the learning method comprising:

extracting, by the feature-quantity extraction unit, the feature quantity from the learning image;

generating, by the weak-classifier setting unit, the weak classifier;

calculating, by the weak-classification calculation unit, the classification result; and

generating, by the classifier generation unit, the classifier.

12. A program for causing a computer to execute:

extracting a feature quantity from a feature point of a learning image with respect to each of a plurality of learning images including a learning image including a detection target and a learning image not including the detection target;

generating a weak classifier based on the feature quantity corresponding to a transfer weak classifier constituting a transfer classifier, which is a classifier for detecting the detection target obtained by statistical learning, among feature quantities extracted from the learning image and the learning image;

calculating a classification result of the detection target according to the weak classifier for every learning image by substituting the feature quantity corresponding to the weak classifier into the weak classifier; and

generating the classifier for detecting the detection target using the weak classifier selected from the plurality of weak classifiers on the basis of the classification result.