Weak hypothesis generation apparatus and method, learning apparatus and method, detection apparatus and method, facial expression learning apparatus and method, facial expression recognition apparatus and method, and robot apparatus

• US 20050102246A1
• Filed: 06/17/2004
• Published: 05/12/2005
• Est. Priority Date: 07/24/2003
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.), University of California San Diego (University of California)
• Original Assignee: Sony Corporation (Sony Group Corp.), University of California San Diego (University of California)
• Inventors: Sabe, Kohtaro, Kawamoto, Kenta, Carlson, Eric C., Hidai, Kenichi, Hershey, John, Movellan, Javier R., Fasel, Ian R., Bartlett, Marian S., Littlewort, Gwendolen C., Susskind, Josh
• Granted Patent: US 7,379,568 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06N 3/004 Artificial life, i.e. computing arrangements simulating life
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 40/165 using facial parts and geometric relationships
• CPC Class Codes: G06V 40/175 Static expression

Abstract

A facial expression recognition system that uses a face detection apparatus realizing efficient learning and high-speed detection processing based on ensemble learning when detecting an area representing a detection target and that is robust against shifts of face position included in images and capable of highly accurate expression recognition, and a learning method for the system, are provided. When learning data to be used by the face detection apparatus by Adaboost, processing to select high-performance weak hypotheses from all weak hypotheses, then generate new weak hypotheses from these high-performance weak hypotheses on the basis of statistical characteristics, and select one weak hypothesis having the highest discrimination performance from these weak hypotheses, is repeated to sequentially generate a weak hypothesis, and a final hypothesis is thus acquired. In detection, using an abort threshold value that has been learned in advance, whether provided data can be obviously judged as a non-face is determined every time one weak hypothesis outputs the result of discrimination. If it can be judged so, processing is aborted. A predetermined Gabor filter is selected from the detected face image by an Adaboost technique, and a support vector for only a feature quantity extracted by the selected filter is learned, thus performing expression recognition.

Claims

1. A weak hypothesis generation apparatus for generating a weak hypothesis for estimating whether provided data is a detection target or not by using a data set including plural learning samples, each of which has been labeled as a detection target or non-detection target, the weak hypothesis generation apparatus comprising:

a selection unit for selecting a part of plural hypotheses and selecting one or plural weak hypotheses having higher estimation performance than others with respect to the data set of the selected part of the hypotheses, as high-performance weak hypotheses;

a new weak hypothesis generation unit for generating one or more new weak hypotheses formed by adding a predetermined modification to the high-performance weak hypotheses, as new weak hypotheses; and

a weak hypothesis selection unit for selecting one weak hypothesis having the highest estimation performance with respect to the data set, from the high-performance weak hypotheses and the new weak hypotheses.

2. The weak hypothesis generation apparatus as claimed in claim 1, wherein the new weak hypothesis generation unit generates new weak hypotheses from the high-performance weak hypotheses on the basis of statistical characteristics of the detection target.

3. The weak hypothesis generation apparatus as claimed in claim 1, wherein the learning samples are images,

the weak hypotheses are adapted for extracting features from a predetermined area in a provided image and estimating whether the provided image is a detection target or not, on the basis of the extracted features, and

the new weak hypothesis generation unit generates the new weak hypotheses for extracting the features from a new area shifted from the predetermined area in the high-performance weak hypotheses by a predetermined quantity.

4. The weak hypothesis generation apparatus as claimed in claim 1, wherein the learning samples are images,

the weak hypotheses are adapted for extracting features from a predetermined area in a provided image and estimating whether the provided image is a detection target or not, on the basis of the extracted features, and

the new weak hypothesis generation unit generates the new weak hypotheses for extracting the features from a new area acquired by converting the scale of the predetermined area in the high-performance weak hypotheses to a predetermined size.

5. The weak hypothesis generation apparatus as claimed in claim 1, wherein the learning samples are images,

the weak hypotheses are adapted for extracting features from a predetermined area in a provided image and estimating whether the provided image is a detection target or not, on the basis of the extracted features, and

the new weak hypothesis generation unit generates the new weak hypotheses for extracting the features from a new area acquired by converting the predetermined area in the high-performance weak hypotheses utilizing symmetry of the detection target and/or the new area and the predetermined area before the conversion.

6. The weak hypothesis generation apparatus as claimed in claim 1, wherein data weighting is set for each learning sample of the data set, and

estimation performance with respect to the data set is calculated on the basis of the data weighting set for each learning sample of the data set.

7. The weak hypothesis generation apparatus as claimed in claim 6, further comprising a data weighting update unit for updating the data weighting of each of the learning samples on the basis of the estimation performance with respect to the data set, of the weak hypothesis selected by the weak hypothesis selection unit.

8. The weak hypothesis generation apparatus as claimed in claim 7, wherein every time the data weighting is updated by the data weighting update unit, processing to generate a weak hypothesis by selecting one of the plural weak hypothesis is repeated.

9. The weak hypothesis generation apparatus as claimed in claim 7, wherein the data weighting update unit updates the data weighting in such a manner that the data weighting of a learning sample on which the estimation value outputted by the weak hypothesis is incorrect becomes relatively larger than the data weighting of a learning sample on which the estimation value is correct.

10. The weak hypothesis generation apparatus as claimed in claim 7, wherein the weak hypothesis deterministically outputs a result of estimation with respect to provided data.

11. The weak hypothesis generation apparatus as claimed in claim 7, wherein the weak hypothesis probabilistically outputs a result of estimation with respect to provided data.

12. The weak hypothesis generation apparatus as claimed in claim 6, wherein the data set includes a variable density image representing the detection target and a variable density image representing the non-detection target, and

the weak hypothesis estimates whether a variable density image provided as an input is a detection target or not on the basis of the difference between the sum of brightness values in one or plural rectangular boxes and the sum of brightness values in the other rectangular boxes in a group of two or more rectangular boxes contained in the variable density image.

13. A weak hypothesis generation method for generating a weak hypothesis for estimating whether provided data is a detection target or not by using a data set including plural learning samples, each of which has been labeled as a detection target or non-detection target, the weak hypothesis generation method comprising:

a selection step of selecting a part of plural hypotheses and selecting one or plural weak hypotheses having higher estimation performance than others with respect to the data set of the selected part of the hypotheses, as high-performance weak hypotheses;

a new weak hypothesis generation step of generating one or more new weak hypotheses formed by adding a predetermined modification to the high-performance weak hypotheses, as new weak hypotheses; and

a weak hypothesis selection step of selecting one weak hypothesis having the highest estimation performance with respect to the data set, from the high-performance weak hypotheses and the new weak hypotheses.

14. A weak hypothesis generation apparatus for generating a weak hypothesis for estimating whether provided data is a detection target or not by using a data set including plural learning samples each of which has been labeled as a detection target or non-detection target, the weak hypothesis generation apparatus comprising:

a selection unit for selecting a part of plural weak hypotheses;

a new weak hypothesis generation unit for generating one or more new weak hypotheses formed by adding a predetermined modification to the part of the weak hypotheses selected by the selection unit, as new weak hypotheses; and

a weak hypothesis selection unit for selecting one weak hypothesis having the highest estimation performance with respect to the data set, from the part of the weak hypotheses selected by the selection unit and the new weak hypotheses.

15. A weak hypothesis generation method for generating a weak hypothesis for estimating whether provided data is a detection target or not by using a data set including plural learning samples each of which has been labeled as a detection target or non-detection target, the weak hypothesis generation method comprising:

a selection step of selecting a part of plural weak hypotheses;

a new weak hypothesis generation step of generating one or more new weak hypotheses formed by adding a predetermined modification to the part of the weak hypotheses selected at the selection step, as new weak hypotheses; and

a weak hypothesis selection step of selecting one weak hypothesis having the highest estimation performance with respect to the data set, from the part of the weak hypotheses selected at the selection step and the new weak hypotheses.

16. A learning apparatus for learning data to be used by a detection apparatus, the detection apparatus being adapted for judging whether provided data is a detection target or not by using a data set including plural learning samples each of which has data weighting set thereon and has been labeled as a detection target or non-detection target, the learning apparatus comprising:

a weak hypothesis generation unit for generating a weak hypothesis for estimating whether provided data is a detection target or not;

a reliability calculation unit for calculating reliability of the weak hypothesis on the basis of the result of estimation of the weak hypothesis generated by the weak hypothesis generation unit with respect to the data set; and

a data weighting update unit for updating the data weighting in such a manner that the data weighting of a learning sample on which estimation by the weak hypothesis generated by the weak hypothesis generation unit is incorrect becomes relatively larger than the data weighting of a learning sample on which estimation is correct;

wherein the weak hypothesis generation unit a selection unit for selecting a part of plural weak hypotheses and selecting one or plural weak hypotheses having higher estimation performance with respect to the data set from the selected part of the weak hypotheses, as high-performance weak hypotheses, a new weak hypothesis generation unit for generating one or more new weak hypotheses by adding a predetermined modification to the high-performance weak hypotheses, and a weak hypothesis selection unit for selecting one weak hypothesis having the highest estimation performance with respect to the data set from the high-performance weak hypotheses and the new weak hypotheses, the weak hypothesis generation unit repeating processing to generate a weak hypothesis every time the data weighting is updated by the data weighting update unit.

17. A learning apparatus for learning data to be used by a detection apparatus, the detection apparatus being adapted for judging whether provided data is a detection target or not by using a data set including plural learning samples each of which has been labeled as a detection target or non-detection target, the learning apparatus comprising:

a weak hypothesis selection unit for repeating processing to select one weak hypothesis from plural weak hypotheses for estimating whether provided data is a detection target or not;

a reliability calculation unit for, every time a weak hypothesis is selected by the weak hypothesis selection unit, calculating reliability of the weak hypothesis on the basis of the result of estimation of the selected weak hypothesis with respect to the data set; and

a threshold value learning unit for calculating and adding the product of the result of estimation of the weak hypothesis with respect to the data set and the reliability of the weak hypothesis every time a weak hypothesis is selected by the weak hypothesis selection unit, and learning an abort threshold value for aborting the processing by the detection apparatus to judge whether the provided data is a detection target or not on the basis of the result of the addition.

18. The learning apparatus as claimed in claim 17, further comprising a data weighting update unit for, every time the weak hypothesis is selected, updating the data weighting in such a manner that the data weighting of a learning sample on which estimation by the selected weak hypothesis is incorrect becomes relatively larger than the data weighting of a learning sample on which estimation is correct, and

the weak hypothesis selection unit repeats processing to select the weak hypothesis every time the data weighting is updated.

19. The learning apparatus as claimed in claim 17, wherein the threshold value learning unit learns the abort threshold value on the basis of the result of addition of the product of the result of estimation of the weak hypothesis with respect to positive data labeled as the detection target, of the data set, and the reliability of the weak hypothesis, calculated and added every time a weak hypothesis is selected by the weak hypothesis selection unit.

20. The learning apparatus as claimed in claim 19, wherein the threshold value learning unit stores a smaller one of a minimum value of the result of addition with respect to the positive data and a discrimination boundary value, as the abort threshold value, every time the weak hypothesis is selected.

21. The learning apparatus as claimed in claim 17, wherein the threshold value learning unit learns the abort threshold value on the basis of the result of addition of the product of the result of estimation of the weak hypothesis with respect to negative data labeled as the non-detection target, of the data set, and the reliability of the weak hypothesis, calculated and added every time a weak hypothesis is selected by the weak hypothesis selection unit.

22. The learning apparatus as claimed in claim 21, wherein the threshold value learning unit stores a larger one of a maximum value of the result of calculation with respect to the negative data and a discrimination boundary value, as the abort threshold value, every time the weak hypothesis is selected.

23. A learning method for learning data to be used by a detection apparatus, the detection apparatus being adapted for judging whether provided data is a detection target or not by using a data set including plural learning samples each of which has been labeled as a detection target or non-detection target, the learning method comprising:

a weak hypothesis selection step of repeating processing to select one weak hypothesis from plural weak hypotheses for estimating whether provided data is a detection target or not;

a reliability calculation step of, every time a weak hypothesis is selected at the weak hypothesis selection step, calculating reliability of the weak hypothesis on the basis of the result of estimation of the selected weak hypothesis with respect to the data set; and

a threshold value learning step of calculating and adding the product of the result of estimation of the weak hypothesis with respect to the data set and the reliability of the weak hypothesis every time a weak hypothesis is selected at the weak hypothesis selection step, and learning an abort threshold value for aborting the processing by the detection apparatus to judge whether the provided data is a detection target or not on the basis of the result of the addition.

24. A detection apparatus for detecting a detection target by discriminating whether provided data is a detection target or not, the detection apparatus comprising:

an estimation result output unit including plural weak hypotheses; and

a discrimination unit for discriminating whether the provided data is a detection target or not on the basis of the result of output of the estimation result output unit;

wherein the estimation result output unit estimates and outputs whether the provided data is a detection target or not for each weak hypothesis on the basis of a feature quantity that has been learned in advance, and

the discrimination unit has an abort unit for adding the product of the result of estimation of a weak hypothesis and reliability that has been learned in advance on the basis of estimation performance of the weak hypothesis, every time one hypothesis outputs the result of estimation, and deciding whether or not to abort processing by the estimation result output unit on the basis of the result of the addition.

25. The detection apparatus as claimed in claim 24, wherein the weak hypotheses are sequentially generated by ensemble learning using a data set including plural learning samples labeled as a detection target or non-detection target,

the estimation result output unit outputs the result of estimation in order of the generation of the plural weak hypotheses,

the abort unit decides whether or not to abort estimation processing by the next weak hypothesis on the basis of the result of comparison between the result of addition and an abort threshold value, every time one weak hypothesis outputs the result of estimation, and

the abort threshold value is learned in advance on the basis of the result of addition of the product of the result of estimation with respect to the data set by the weak hypotheses sequentially generated by ensemble learning and reliability of the weak hypotheses every time a weak hypothesis is generated.

26. The detection apparatus as claimed in claim 25, wherein the abort threshold value is a smaller value of a minimum value of the result of addition acquired from the result of estimation with respect to positive data labeled as the detection target, of the data set, and a discrimination boundary value, and

the abort unit aborts processing by the estimation result output unit when the result of addition is smaller than the abort threshold value.

27. The detection apparatus as claimed in claim 25, wherein the abort threshold value is a larger value of a maximum value of the result of addition acquired from the result of estimation with respect to negative data labeled as the non-detection target, of the data set, and a discrimination boundary value, and

the abort unit aborts processing by the estimation result output unit when the result of addition is larger than the abort threshold value.

28. The detection apparatus as claimed in claim 25, wherein each learning sample of the data set has data weighting,

the data weighting is sequentially updated in such a manner that the data weighting of a learning sample on which the result of estimation outputted by the sequentially generated weak hypothesis is incorrect becomes relatively larger than the data weighting of a learning sample on which the result of estimation is correct, and

the reliability is learned on the basis of the result of estimation on each learning sample with the data weighting set thereon, every time the weak hypothesis is generated.

29. The detection apparatus as claimed in claim 24, wherein the detection target is a face image.

30. The detection apparatus as claimed in claim 29, further comprising:

a face feature extraction unit for extracting face features by filtering a face image detected as the detection target by the discrimination unit, using a Gabor filter; and

an expression recognition unit for recognizing the expression of the provided face image on the basis of the face features.

31. A detection method for detecting a detection target by discriminating whether provided data is a detection target or not, the detection method comprising:

an estimation result output step of estimating and outputting whether the provided data is a detection target or not by each of plural weak hypotheses on the basis of a feature quantity that has been learned in advance; and

a discrimination step of discriminating whether the provided data is a detection target or not on the basis of the result of output from the estimation result output step;

wherein the discrimination step includes an abort step of adding the product of the result of estimation of a weak hypothesis and reliability that has been learned on the basis of estimation performance of the weak hypothesis, every time one hypothesis outputs the result of estimation, and deciding whether or not to abort processing by the plural weak hypotheses on the basis of the result of the addition.

32. A facial expression learning apparatus for learning data to be used by a facial expression recognition apparatus, the facial expression recognition apparatus being adapted for recognizing an expression of a provided face image by using an expression learning data set including plural face images representing specific expressions as recognition targets and plural face images representing expressions different from the specific expressions,

the facial expression learning apparatus comprising an expression learning unit for learning data to be used by the facial expression recognition apparatus, the facial expression recognition apparatus identifying the face images representing the specific expressions from provided face images on the basis of a face feature extracted from the expression learning data set by using a Gabor filter.

33. The facial expression learning apparatus as claimed in claim 32, wherein the expression learning unit learns a support vector for identifying a face image representing the specific expression on the basis of the face feature extracted from the expression learning data set by using the Gabor filter.

34. The facial expression learning apparatus as claimed in claim 32, wherein the expression learning unit has:

a weak hypothesis generation unit for repeating processing to generate a weak hypothesis for estimating whether a provided face image is of the specific expression or not on the basis of the result of filtering by one Gabor filter selected from plural Gabor filters;

a reliability calculation unit for calculating reliability of the weak hypothesis generated by the weak hypothesis generation unit on the basis of estimation performance of the weak hypothesis with respect to the expression learning data set; and

a data weighting update unit for updating data weighting set for the expression learning data set on the basis of the reliability, and

wherein the weak hypothesis generation unit repeats processing to generate the weak hypothesis while selecting one Gabor filter having the highest estimation performance with respect to the expression learning data set every time the data weighting is updated.

35. The facial expression learning apparatus as claimed in claim 32, wherein the expression learning unit has:

a weak hypothesis generation unit for repeating processing to generate a weak hypothesis for estimating whether a provided face image is of the specific expression or not on the basis of the result of filtering by one Gabor filter selected from plural Gabor filters;

a reliability calculation unit for calculating reliability of the weak hypothesis generated by the weak hypothesis generation unit on the basis of estimation performance of the weak hypothesis with respect to the expression learning data set;

a data weighting update unit for updating data weighting set for the expression learning data set on the basis of the reliability; and

a support vector learning unit for learning a support vector for identifying a face image representing the specific expression on the basis of the face feature extracted form the expression learning data set by a predetermined Gabor filter, and

wherein the weak hypothesis generation unit repeats processing to generate the weak hypothesis while selecting one Gabor filter having the highest estimation performance with respect to the expression learning data set every time the data weighting is updated, and

the support vector learning unit extracts the face feature by using the Gabor filter selected by the weak hypothesis generated by the weak hypothesis generation unit, and thus learns the support vector.

36. A facial expression learning method for learning data to be used by a facial expression recognition apparatus, the facial expression recognition apparatus being adapted for recognizing an expression of a provided face image by using an expression learning data set including plural face images representing specific expressions as recognition targets and plural face images representing expressions different from the specific expressions,

the facial expression learning method comprising an expression learning step of learning data to be used by the facial expression recognition apparatus, the facial expression recognition apparatus identifying the face images representing the specific expressions from provided face images on the basis of a face feature extracted from the expression learning data set by using a Gabor filter.

37. The facial expression learning method as claimed in claim 36, wherein the expression learning step includes learning a support vector for identifying a face image representing the specific expression on the basis of the face feature extracted from the expression learning data set by using the Gabor filter.

38. The facial expression learning method as claimed in claim 36, wherein the expression learning step includes:

a weak hypothesis generation step of repeating processing to generate a weak hypothesis for estimating whether a provided face image is of the specific expression or not on the basis of the result of filtering by one Gabor filter selected from plural Gabor filters;

a reliability calculation step of calculating reliability of the weak hypothesis generated at the weak hypothesis generation step on the basis of estimation performance of the weak hypothesis with respect to the expression learning data set; and

a data weighting update step of updating data weighting set for the expression learning data set on the basis of the reliability, and

wherein the weak hypothesis generation step includes generating the weak hypothesis while selecting one Gabor filter having the highest estimation performance with respect to the expression learning data set every time the data weighting is updated.

39. The facial expression learning method as claimed in claim 36, wherein the expression learning step includes:

a weak hypothesis generation step of repeating processing to generate a weak hypothesis for estimating whether a provided face image is of the specific expression or not on the basis of the result of filtering by one Gabor filter selected from plural Gabor filters;

a reliability calculation step of calculating reliability of the weak hypothesis generated at the weak hypothesis generation step on the basis of estimation performance of the weak hypothesis with respect to the expression learning data set;

a data weighting update step of updating data weighting set for the expression learning data set on the basis of the reliability; and

a support vector learning step of learning a support vector for identifying a face image representing the specific expression on the basis of the face feature extracted form the expression learning data set by a predetermined Gabor filter, and

wherein the weak hypothesis generation step includes generating the weak hypothesis while selecting one Gabor filter having the highest estimation performance with respect to the expression learning data set every time the data weighting is updated, and

the support vector learning step includes extracting the face feature by using the Gabor filter selected by the weak hypothesis generated by the weak hypothesis generation unit, and thus learning the support vector.

40. A facial expression recognition apparatus comprising:

a face feature extraction unit for filtering a provided face image by using a Gabor filter and thus extracting a face feature; and

an expression recognition unit for recognizing an expression of the provided face image on the basis of the face feature.

41. The facial expression recognition apparatus as claimed in claim 40, wherein the expression recognition unit has an expression identifier for identifying a face image showing a specific expression as a recognition target from the provided face image, and

the expression identifier is a support vector machine (SVM) for identifying whether the face feature extracted by the face feature extraction unit is the specific expression or not, using a support vector that has been learned in advance.

42. The facial expression recognition apparatus as claimed in claim 40, wherein the expression recognition unit has an expression identifier for identifying a face image showing a specific expression as a recognition target from the provided face image, and

the expression identifier comprises:

an estimation result output unit including plural weak hypotheses; and

a discrimination unit for discriminating whether the provided face image shows the specific expression on the basis of the result of estimation outputted from the estimation result output unit;

wherein the estimation result output unit estimates and outputs whether the provided face image shows the specific expression or not on the basis of a feature quantity that has been learned in advance, and

the discrimination unit adds the product of the result of estimation of a weak hypothesis and reliability that has been learned on the basis of estimation performance of the weak hypothesis, every time one hypothesis outputs the result of estimation, and discriminating whether the face image shows the specific expression or not on the basis of the result of addition.

43. The facial expression recognition apparatus as claimed in claim 42, wherein the expression recognition unit has plural expression identifiers for identifying faces images showing specific expressions, and an expression decision unit for deciding the expression of the provided face image, and

the expression decision unit decides the expression of the provided face image on the basis of the result of addition of each expression identifier.

44. The facial expression recognition apparatus as claimed in claim 40, wherein the expression recognition unit has an expression identifier for identifying a face image showing a specific expression as a recognition target from the provided face image,

the expression identifier is a support vector machine (SVM) for identifying whether the face feature extracted by the face feature extraction unit is the specific expression or not, using a support vector that has been learned in advance, and

the support vector has been learned in advance on the basis of a face feature extracted from an expression learning data set including plural face images showing a specific expression as a recognition target and plural images showing expressions different from the specific expression to extract a face feature from the expression learning data set, by using all Gabor filters selected as weak hypotheses when repeatedly generating a weak hypothesis for estimating whether a provided face image shows the specific expression or not on the basis of the result of filtering of one Gabor filter selected form plural Gabor filters.

45. A facial expression recognition method comprising:

a face feature extraction step of filtering a provided face image by using a Gabor filter and thus extracting a face feature; and

an expression recognition step of recognizing an expression of the provided face image on the basis of the face feature.

46. The facial expression recognition method as claimed in claim 45, wherein the expression recognition step includes a specific expression identification step of identifying a face image showing a specific expression as a recognition target from the provided face image, and

at the specific expression identification step, whether the face feature extracted at the face feature extraction step is the specific expression or not is identified by a support vector machine (SVM) using a support vector that has been learned in advance.

47. The facial expression recognition method as claimed in claim 45, wherein the expression recognition step has a specific expression identification step of identifying a face image showing a specific expression as a recognition target from the provided face image, and

the specific expression identification step includes:

an estimation result output step of estimating whether the provided face image shows the specific expression or not by each plural weak hypotheses on the basis of a feature quantity that has been learned in advance; and

a discrimination step of discriminating whether the provided face image shows the specific expression on the basis of the result of estimation outputted from the estimation result output step;

wherein at the estimation result output step, the product of the result of estimation of a weak hypothesis and reliability that has been learned on the basis of estimation performance of the weak hypothesis is added every time one hypothesis outputs the result of estimation, and whether the face image shows the specific expression or not is discriminated on the basis of the result of addition.

48. The facial expression recognition method as claimed in claim 45, wherein the expression recognition step has a specific expression identification step of identifying a face image showing the specific expression from the provided face image,

the specific expression identification step is a step of identifying whether the face feature extracted at the face feature extraction step represents a face image showing the specific expression or not by a support vector machine (SVM) using a support vector that has been learned in advance, and

the support vector has been learned in advance on the basis of a face feature extracted from an expression learning data set including plural face images showing a specific expression as a recognition target and plural images showing expressions different from the specific expression to extract a face feature from the expression learning data set, by using all Gabor filters selected as weak hypotheses when repeatedly generating a weak hypothesis for estimating whether a provided face image shows the specific expression or not on the basis of the result of filtering of one Gabor filter selected form plural Gabor filters.

49. An autonomously acting robot apparatus comprising:

an image pickup unit for picking up an image of its surroundings;

a cut-out unit for cutting out a window image of an arbitrary size from the image picked up by the image pickup unit; and

a detection apparatus for detecting whether the window image is an image representing a detection target or not,

wherein the detection apparatus has

an estimation result output unit including plural weak hypotheses, and

a discrimination unit for discriminating whether the window image is an image representing a detection target or not on the basis of the result of estimation outputted from the estimation result output unit, and

wherein the estimation result output unit estimates and outputs whether the provided data is a detection target or not for each weak hypothesis on the basis of a feature quantity that has been learned in advance, and

the discrimination unit has an abort unit for adding the product of the result of estimation of a weak hypothesis and reliability learned on the basis of estimation performance of the weak hypothesis every time one weak hypothesis outputs the result of estimation, and deciding whether or not to abort processing by the estimation result output unit on the basis of the result of addition.

50. An autonomously acting robot apparatus comprising:

an image pickup unit for picking up an image of its surroundings;

a face image detection apparatus for detecting a predetermined area as a face image from the image picked up by the image pickup unit; and

a facial expression recognition apparatus for recognizing an expression of the face image,

wherein the facial expression recognition apparatus has

a face feature extraction unit for filtering the face image detected by the face image detection apparatus by using a Gabor filter and thus extracting a face feature, and

an expression recognition unit for recognizing an expression of the provided face image on the basis of the face feature.