Signal pattern recognition apparatus comprising parameter training controller for training feature conversion parameters and discriminant functions

• US 5,754,681 A
• Filed: 06/22/1995
• Issued: 05/19/1998
• Est. Priority Date: 10/05/1994
• Status: Expired due to Term

• Current Assignee: Seiko Epson Corporation (Seiko Group)
• Original Assignee: ATR Interpreting Telephony Research Laboratories
• Inventors: Watanabe, Hideyuki, Yamaguchi, Tsuyoshi, Katagiri, Shigeru
• Primary Examiner(s): Couso, Yon
• Assistant Examiner(s): NGUYEN, HA T
• Application Number:
  US08/493,719
• Time in Patent Office: 1,062 Days
• Field of Search: 395/2.4-2.49, 395/23, 395/24, 395/22, 382/181, 382/197, 382/224, 382/155-160, 382/190
• US Class Current: 382/159
• CPC Class Codes: G06F 18/2135 based on approximation criteria, e.g. principal component analysis
• CPC Class Codes: G06F 18/22 Matching criteria, e.g. proximity measures

Abstract

In a signal pattern recognition apparatus, a plurality of feature transformation sections respectively transform an inputted signal pattern into vectors in a plurality of feature spaces corresponding respectively to predetermined classes using a predetermined transformation parameter corresponding to each of the classes so as to emphasize a feature of each of the classes, and a plurality of discriminant function sections respectively calculates a value of a discriminant function using a predetermined discriminant function representing a similarity measure of each of the classes for the transformed vectors in the plurality of feature spaces. Then, a selection section executes a signal pattern recognition process by selecting a class to which the inputted signal pattern belongs based on the calculated values of a plurality of discriminant functions corresponding respectively to the classes, and a training control section trains and sets a plurality of transformation parameters of the feature transformation process and a plurality of discriminant functions so that an error probability of the signal pattern recognition is minimized based on a predetermined training signal pattern.

Claims

1. A signal pattern recognition apparatus for classifying an inputted signal pattern into one of a plurality of predetermined classes so as to recognize the inputted signal pattern, comprising:

a plurality of feature transformation means for respectively transforming the inputted signal pattern into vectors in a plurality of feature spaces corresponding respectively to said classes by executing a feature transformation process by means of a predetermined transformation parameter corresponding to each of said classes so as to emphasize a feature of each of said classes, said feature transformation means being provided respectively for said plurality of classes;

a plurality of discriminant function means for respectively calculating a value of a discriminant function by means of a predetermined discriminant function representing a similarity measure of each of said classes for said vectors in said plurality of feature spaces which are transformed by said plurality of feature transformation means, said discriminant function means being provided respectively for said plurality of classes;

selection means for executing a signal pattern recognition process by selecting a class to which the inputted signal pattern belongs based on the values of said plurality of discriminant functions corresponding respectively to said classes, said discriminant functions being obtained through said calculation executed by said plurality of discriminant function means; and

training control means for training and setting said plurality of transformation parameters of said feature transformation process and said plurality of discriminant functions, so that an error probability of said signal pattern recognition is minimized based on a predetermined training signal pattern.

2. The signal pattern recognition apparatus as claimed in claim 1,

wherein each of said plurality of feature transformation means linearly transforms the inputted signal pattern into vectors in said plurality of feature spaces corresponding respectively to said classes by projecting the inputted signal pattern onto a predetermined basis vector and multiplying a resulting vector by a predetermined real number.

3. The signal pattern recognition apparatus as claimed in claim 1,

wherein each of said plurality of discriminant functions of the discriminant function means is a predetermined quadric discriminant function representing the similarity measure of each of said classes.

4. The signal pattern recognition apparatus as claimed in claim 1,

wherein said training control means performs adaptation of said plurality of transformation parameters of said feature transformation process and said plurality of discriminant functions of the discriminant function means, so that the error probability of said signal pattern recognition is minimized, based on said predetermined training signal pattern, by means of an adaptive minimization method utilizing a probabilistic descent theorem.

5. The signal pattern recognition apparatus as claimed in claims 1, further comprising:

signal conversion means for converting an inputted speech into a speech signal and outputting the speech signal; and

feature extraction means for converting the speech signal outputted from said signal conversion means into a predetermined speech feature parameter, and outputting the obtained feature parameter as a signal pattern to said plurality of feature transformation means and said training control means, thereby recognizing the inputted speech.

6. The signal pattern recognition apparatus as claimed in claim 5,

wherein said feature extraction means transforms the speech signal outputted from said signal conversion means into LPC cepstrum coefficient vectors through linear prediction analysis, and outputting resulting vectors as a signal pattern to said plurality of feature transformation means and said training control means.

7. The signal pattern recognition apparatus as claimed in claim 1, further comprising:

image conversion means for converting a character into dot image data, and outputting the dot image data as a signal pattern to said plurality of feature transformation means and said training control means, thereby recognizing the character.

8. The signal pattern recognition apparatus as claimed in claim 1, further comprising:

further image conversion means for converting an image into dot image data, and outputting the dot image data as a signal pattern to said plurality of feature transformation means and said training control means, thereby recognizing the image.

13. The signal pattern recognition apparatus as claimed in claim 1, further comprising a feature extraction section for receiving a first signal from an input device and for producing said inputted signal pattern, said feature extraction section extracting vectors of features parameters from said first signal, such that data representing said extracted vectors are included in said inputted signal pattern.

14. The signal pattern recognition apparatus as claimed in claim 1, wherein said training control means is connected to receive the inputted signal pattern, so as to train and set said plurality of transformation parameters of said feature transformation process and said plurality of discriminant functions.

15. The signal pattern recognition apparatus as claimed in claim 1, wherein a unique feature metric space is provided for each class.

9. A method for classifying an inputted signal pattern into one of a plurality of predetermined classes so as to recognize the inputted signal pattern, including the following steps of:

transforming the inputted signal pattern into vectors in a plurality of feature spaces corresponding respectively to said classes by executing a feature transformation process by means of a predetermined transformation parameter corresponding to each of said classes so as to emphasize a feature of each of said classes;

calculating a value of a discriminant function by means of a predetermined discriminant function representing a similarity measure of each of said classes for said vectors in said plurality of feature spaces which are obtained through said feature transformation process;

executing a signal pattern recognition process by selecting a class to which the inputted signal pattern belongs based on the calculated values of said plurality of discriminant functions corresponding respectively to said classes; and

training and setting the transformation parameter of said feature transformation process and each of said discriminant functions, so that an error probability of said signal pattern recognition is minimized based on a predetermined training signal pattern.

10. The method as claimed in claim 9, wherein said transforming step includes a step of linearly transforming the inputted signal pattern into vectors in said plurality of feature spaces corresponding respectively to said classes by projecting the inputted signal pattern onto a predetermined basis vector and multiplying resulting vectors by a predetermined real number.

11. The method as claimed in claim 9,

wherein each of said discriminant functions is a predetermined quadric discriminant function representing the similarity measure of each of said classes.

12. The method as claimed in claim 9,

wherein said training step includes a step of performing adaptation of the transformation parameter of said feature transformation process and said discriminant functions, so that the error probability of said signal pattern recognition is minimized, based on the predetermined training signal pattern, by means of an adaptive minimization method utilizing a probabilistic descent theorem.

16. The method as claimed in claim 9, further comprising the step of:

extracting vectors of feature parameters from an initial signal and producing the inputted signal pattern based upon the extracted vectors.