JP2001160145A - Method and device for classifying pattern - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the storage capacity and error influence in the extraction of features and further to enable classification even when the extraction features of from an input pattern fails. SOLUTION: The feature quantity is extracted from a sample pattern 1 and stored in a feature quantity database 3. A coded data generating part 11 generates coded data from the data of the database 3, an encoding part 13 encodes the data from the database 3, and a classification data generating part 14 generates classification data. A feature quantity extracting part 22 extracts feature quantity from the input pattern 21. An encoding part 23 refers to the generated coded data, and the part 22 encodes the extracted feature quantity. A class dividing part 24 refers to the generated classification data and divides the coding results of the encoding part 23 into classes.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern classification used in the field of character recognition, pattern recognition such as personal identification based on an iris (iris) pattern, or the field of selection of a recognition dictionary. The present invention relates to a method and a pattern classifying apparatus, and particularly has a function of temporarily encoding a feature amount extracted from an input pattern and classifying the feature amount into an appropriate classification class.

[0002]

2. Description of the Related Art Conventionally, a pattern classifying method or a pattern classifying apparatus for classifying a pattern is as follows.
For example, various techniques have been proposed as described in the following documents and the like. Reference: Nagao, "Pattern Information Processing" (1983) Corona, P.113-120

The above-mentioned literature discloses several algorithms for class classification (clustering) for a given set of sample patterns. Among them, a typical method (k-means method) is to first set a representative value of a feature value arbitrarily for a plurality of classification classes, and then classify a classification class of each sample pattern with a minimum distance from the representative value. And the representative value of the classification class is changed according to the average value of the feature amounts of the sample patterns classified into each class. By repeating such classification and the process of changing the representative value, the representative value of the final classification class is determined, and the classification of the input pattern is performed by classifying the input pattern into the classification class having the representative value with the shortest distance. there were.

[0004]

However, in a pattern classification method or a pattern classification apparatus as described in the above-mentioned literature, it is necessary to store all the representative values of the feature amounts of the classification classes. The more types of features used, the more storage capacity is required. Further, since the distance between patterns is directly calculated from the feature amount, if the extraction accuracy of the feature amount is low, an error is included in the result of the distance calculation, which affects the class classification.
Furthermore, for features that failed to be extracted from the input pattern, the values are undefined and distance calculation cannot be performed.
There was a problem that it could not be classified. An object of the present invention is to solve the problems of the prior art and provide a pattern classification method and a pattern classification device capable of temporarily encoding a feature amount extracted from an input pattern and classifying the characteristic amount into an appropriate classification class. I do.

[0005]

According to a first aspect of the present invention, there is provided a pattern classification method for classifying an input pattern into one of a plurality of classification classes in accordance with the characteristic amount. Generating an input code by coding the input pattern according to the feature amount, expressing the classification class with a representative code having the same form as the input code, and inputting a classification class for classifying the input pattern to the input pattern. It is determined by calculating the distance between the code and the representative code. Claim 2
According to the invention of the first aspect, in the pattern classification method according to the first aspect, the coding of the input pattern is performed by a statistic obtained from a feature amount of a sample pattern collected in advance.

According to a third aspect of the present invention, in the pattern classification method of the second aspect, the coding of the input pattern is performed by using an average and a standard deviation of the feature amounts of the sample pattern. According to a fourth aspect of the present invention, in the pattern classification method according to any one of the first to third aspects, the representative code is obtained by coding a sample pattern collected in advance from a sample code obtained by coding the sample pattern. It is selected by calculating the distance between them.

According to a fifth aspect of the present invention, in the pattern classification method according to the fourth aspect, the representative code is selected from the sample codes in order of a minimum value of a distance from another sample code. . According to a sixth aspect of the present invention, in the pattern classification method according to the fourth or fifth aspect, the distance between the sample codes and the distance between the input code and the sample code are statistical values obtained from feature amounts of the sample pattern. Defined by

According to a seventh aspect of the present invention, in the pattern classification method according to any one of the first to sixth aspects, the classification class for classifying the input pattern is a class of the representative code having the shortest distance from the input code. Class. As described above, according to the first to seventh aspects of the present invention, the statistic related to the feature is measured in advance, and the feature is coded using the result. Next, the distance between the codes is calculated based on the statistics,
The code used for the classification is selected based on the distance. This code is a representative code of the classification class.
The input pattern is similarly coded from the result of feature extraction by using a statistic, and is classified into the class of the closest representative code.

By adopting such a configuration, the representative code may be stored instead of the representative value of the feature amount of the classification class, so that the storage capacity can be reduced. Moreover,
For example, by defining the distance value to be 0 between coding of the feature amount and codes whose feature amounts are not so different, the influence of an error in feature extraction can be reduced. If the feature extraction from the input pattern fails, the feature amount is regarded as the average value of the sample pattern and coded, so that normal distance calculation can be performed and classification can be performed.

According to an eighth aspect of the present invention, in the pattern classification method according to any one of the first to seventh aspects, the ranking of the plurality of classification classes is calculated by calculating a distance between the input code and the representative code. By doing so. According to a ninth aspect of the present invention, in the pattern classification method according to the eighth aspect, the plurality of classification classes are ranked in ascending order of the distance between the input code and the representative code. By adopting such a configuration, the classification classes of the input pattern can be determined in order from the highest matching degree as the classification destination.

According to a tenth aspect of the present invention, in the pattern classification method according to any one of the first to ninth aspects, a frequency of a classification class into which the input pattern is classified is stored for each of the input patterns. . By adopting such a configuration, by using data that can specify an input pattern such as a result of pattern recognition, and recording information on which class the input pattern is classified into, the statistics of the class into which the pattern is classified can be obtained. Information can be provided.

According to the eleventh aspect of the present invention, in a pattern classification device for classifying an input pattern into one of a plurality of classification classes according to its characteristic amount, an input code is generated by coding the input pattern according to the characteristic amount. Coding means, a classification data generating means for generating classification data expressing the classification class with a representative code having the same form as the input code, and a classification class for classifying the input pattern with reference to the classification data, Classifying means for calculating and determining a distance between the input code and the representative code. According to a twelfth aspect of the present invention, in the pattern classification device of the eleventh aspect, the coding means performs coding by using a statistic obtained from a feature amount of a sample pattern collected in advance.

According to a thirteenth aspect of the present invention, in the pattern classification device of the twelfth aspect, the coding means performs coding by using an average and a standard deviation of the feature amounts of the sample pattern. According to a fourteenth aspect of the present invention, in the pattern classification device according to any one of the eleventh to thirteenth aspects, the classification data generating unit converts the representative code from a sample code obtained by coding a sample pattern collected in advance. The distance between these sample codes is calculated and selected.

According to a fifteenth aspect of the present invention, in the pattern classification apparatus according to the fourteenth aspect, the classification data generating means determines the representative code from a sample code having a larger minimum value of a distance from another sample code. It is configured to select in order. According to the sixteenth aspect, in the fourteenth aspect,
In the pattern classification device, the distance between the sample codes and the distance between the input code and the sample code are defined by statistics obtained from feature amounts of the sample patterns. According to a seventeenth aspect of the present invention, in the pattern classification device according to any one of the eleventh to sixteenth aspects, the class classification means is configured to: It is configured to determine the code classification class. According to the inventions of claims 11 to 17, substantially the same operations as those of claims 1 to 7 are performed.

According to the eighteenth aspect of the present invention, the eleventh to seventeenth aspects
The pattern classification device according to any one of the above, further comprising a class order determination unit that ranks the plurality of classification classes by calculating a distance between the input code and the representative code. According to a nineteenth aspect of the present invention, in the pattern classification device according to the eighteenth aspect, the class order determining means performs ranking of the plurality of classification classes in ascending order of the distance between the input code and the representative code. I have to. According to the eighteenth and nineteenth aspects, the same operation as the eighth and ninth aspects is performed. According to a twentieth aspect of the present invention, in the pattern classification device according to any one of the eleventh to nineteenth aspects, there is provided a pattern class recording unit for storing a frequency of a classification class into which the input pattern is classified for each of the input patterns. are doing. This allows
The operation is substantially the same as that of the tenth aspect.

[0016]

DESCRIPTION OF THE PREFERRED EMBODIMENTS (First Embodiment) (Configuration) FIGS. 1A and 1B are configuration diagrams of a pattern classification device according to a first embodiment of the present invention. () Is a configuration diagram of a classification data generation portion, and (b) is a configuration diagram of a class classification portion. The classification data generation portion of the pattern classification device includes a feature amount extraction unit 2 to which a reference sample pattern 1 for generating classification data is given. The feature amount extraction unit 2 extracts a feature amount from the sample pattern 1 and stores the extracted feature amount in a feature amount database (DB) 3. It is assumed that the sample pattern 1, the feature amount extraction unit 2, and the feature amount database 3 are separately prepared when using the pattern classification device of the present embodiment.

The terminal 4a of the changeover switch 4 is connected to the output side of the database 3. The changeover switch 4 is
The data of the feature amount database 3 is input from the terminal 4a,
This is a switch that outputs this data by switching it to terminals 4b and 4c. A coded data generation unit (for example, a coded data generation unit) 11 is connected to the terminal 4 b of the changeover switch 4. The coded data generation unit 11 generates coded data, for example, a counter i,
The coded data storage unit 12 is connected to an output side of the coded data storage unit 12. The coded data storage unit 12 stores the coded data generated by the coded data generation unit 11.

A coding means (for example, a coding unit) 13 is connected to the terminal 4c of the changeover switch 4. The coding unit 13 refers to the coded data stored in the coded data storage unit 12 and codes the feature amount of the sample pattern stored in the feature amount database 3. Means and determination means, and a classification data generating means (for example,
A classification data generator 14 is connected. The classification data generation unit 14 generates classification data from the feature amounts coded by the coding unit 13, and is composed of, for example, counters i and j, calculation means, and determination means. The storage unit 15 is connected. The classification data storage unit 15 stores the classification data generated by the classification data generation unit 14.

The class classification portion of the pattern classification device has a characteristic amount extracting means (for example, a characteristic amount extracting unit) 22 for extracting a characteristic amount from the input pattern 21. The feature amount extraction unit 22 performs the same processing as the feature amount extraction unit 2 of the classification data generation part, and a coding unit (for example, a coding unit) 23 is connected to the output side. The coding unit 23 performs the same processing as the coding unit 13 of the classification data generation unit,
The feature amount extraction unit 2 refers to the coded data stored in the
2 has a function of coding the feature quantity extracted in Step 2. A class classification unit (for example, a class classification unit) 24 is connected to the output side of the coding unit 23. The class classification unit 24 classifies the input pattern coded by the coding unit 23 with reference to the classification data stored in the classification data storage unit 15, and includes, for example, a counter i, a calculation unit, and a determination unit. Means.

In this embodiment, the feature amount extraction units 2 and 2
2, coded data generation unit 11, coding units 13, 2
3, the classification data generation unit 14 and the class classification unit 24
For example, it is configured by a central processing unit (hereinafter, referred to as “CPU”). The feature amount database 3, the coded data storage unit 12, and the classification data storage unit 15 include, for example,
It is composed of a storage device.

(Classification Method) In this embodiment, FIG.
By using the sample pattern 1, data necessary for class classification is extracted using the sample pattern 1 (processing (1)), and FIG.
With this configuration, the class classification (process (2)) of the input pattern 21 is performed. Prior to extracting the data necessary for the class classification, the feature amount database 3 already stores the M extracted from the total number S of the sample patterns 1.
It is assumed that individual feature amounts are stored.

(1) Processing of Classification Data Generation Portion of FIG. 1A FIG. 2 is a flowchart showing processing of the coded data generation section 11 of FIG. 1, and FIG. 3 shows processing of the coding section 13 of FIG. FIG. 4 is a flowchart showing the processing of the classification data generator 14 shown in FIG. Hereinafter, the processing contents of the classification data generation portion of FIG. 1A will be described with reference to FIGS. In order to generate the classification data, first, the terminal 4a of the switch 4 is connected to the terminal 4b side, and the feature amount of the sample pattern stored in the feature amount database 3 is supplied to the coded data generation unit 11. The coded data generation unit 11 generates data necessary for coding using the feature amount of the sample pattern, and stores the coded data in the coded data storage unit 12.

The processing of the coded data generation unit 11 is
This will be described with reference to the flowchart of FIG. First, in step S1, the value of the counter i is initialized to 0, and in step S2, the average μi and the standard deviation σi of the i-th feature amount in all the sample patterns are calculated. Next, in step S3, the value of the counter j is initialized to 0, and in step S4, the group number Gj of the i-th feature amount of the j-th sample pattern is determined. The value of the group number Gj is determined as follows.

[0024]

(Equation 1) Here, xi is the value of the i-th feature amount, and if xi is undefined due to a failure in feature extraction or the like, a group number 2 is given.
The calculation of the group number is performed for all sample patterns by adding +1 to the counter j in step S5 and determining whether or not the total number of data is j in step S6. Then, in step S7, the average value of the i-th feature amount of the data having the same group number is calculated for each group number.

The above process is repeated for all the feature amounts by adding +1 to the counter i in step S8 and determining whether or not the number of feature types is i in step S9, and each feature amount is coded data. Mean μi, standard deviation σi, and mean μi (0), μi for each group number
(1), μi (2) and μi (3) are obtained. These values are stored in the coded data storage unit 12 of FIG. 1 as coded data. After generating the coded data, the terminal 4 a of the changeover switch 4 in FIG. 1 is switched to the terminal 4 c side, and the feature of the sample pattern stored in the feature database 3 is supplied to the coding unit 13. The coding unit 13 refers to the coded data stored in the coded data storage unit 12, codes each sample pattern based on the feature amount, and supplies the coding result to the classification data generation unit 14.

The processing of the coding unit 13 will be described with reference to the flowchart of FIG. First, step S1
In 1, initialize the generated code with an empty string,
In step S12, the counter i is initialized to 0, and the process proceeds to step S13. In step S13, a group number Gi relating to the i-th feature amount of the data input to the coding unit 13 is obtained. The determination of the group number Gi is performed in the same manner as in the method of the coded data generation unit 11 described with reference to FIG. Next, in step S14, a code expression of the group number Gi is added to the end of the codes created up to now. The code representation of the group number Gi is an integer of 0 to 3 or the like representing the value of Gi. The above processing is performed in step S
By adding +1 to the counter i at 15 and determining whether or not the number of feature types is i at step S16, the process is repeatedly performed for all feature amounts to obtain a code including a code representation of a group number of all feature amounts.

The classification data generator 14 shown in FIG. 1 calculates the distance between each sample pattern coded by the coding unit 13 and generates classification data necessary for class classification. The generated classification data is stored in the classification data storage unit 15. The processing of the classification data generation unit 14 will be described with reference to the flowchart of FIG. First, step S2
At 1, the value of the counter i is initialized to 0. This means that processing relating to the code of the 0th sample pattern (this is referred to as “code 0”) is performed. After step S21, the value of the counter j is initialized to 0 in step S22. Next, step S23
, The distance D (i, j) between the code i and the code j is calculated. Distance D (i, j) between code i and code j
Is defined by the following equation:

[0028]

(Equation 2) Here, Gk (i) and Gk (j) are codes i,
This is the group number of the k-th feature quantity of code j, and Dk (G
k (i), Gk (j)) is the distance between groups of the k-th feature amount. This value is defined as shown in Table 1 below, and is stored in advance in a storage unit such as a table. For example, group Gk (i) = 1 and group Gk (j) = 0
Is the group distance Dk (Gk (i), Gk (j)) of
μk (1) −μk (0).

[0029]

[Table 1] In Table 1, the inter-group distance is defined as the difference between the average values of the groups. However, the distance between the group 0 and the group 2 is defined as treating the groups 1 and 2 in the same manner. The distance between group 1 and group 3 is equal to the distance between group 1 and group 3. The distance between codes is the sum of the distance between groups for each feature value. Next, in step S24, the distance D between the code i and the code j
Of (i, j) and the minimum value Di of the inter-code distance obtained so far, the minimum value is set as Di again.

By repeating the above processing by adding +1 to the counter j in step S25 and determining whether or not the total number of codes is j in step S26, the distance between the code i and the other codes is minimized. Find the value Di.
The above processing is repeated for all the codes by adding +1 to the counter i in step S27 and determining whether or not the total number of codes is i in step S28. Then, in step S29, the upper N Select codes. The selected code is stored in the classification data storage unit 15 in FIG. 1 as a representative code of the classification class.

(2) Processing of Class Classification Portion of FIG. 1B FIG. 5 is a flowchart showing the processing of the class classification section 24 of FIG. In the class classification part of FIG.
The input pattern 21 is input to a feature amount extraction unit 22 to extract a feature amount. The coding unit 23 codes the feature amount with reference to the coded data stored in the coded data storage unit 12, similarly to the coding unit 13 of FIG. Supplied to The class classification unit 24 refers to the classification data stored in the classification data storage unit 15 and performs class classification.

The processing of the classifying section 24 will be described with reference to the flowchart of FIG. Classification unit 24
Classifies the code C obtained as a result of encoding the feature amount extracted from the input pattern 21 by the feature amount extraction unit 22 in the encoding unit 23 as follows. First, in step S31, the value of the counter i is initialized to 0, and then in step S32, the code C
D between the i-th classification class and the representative code Ci
(C, Ci) is calculated.

The above processing is performed by adding +1 to the counter i in step S33, and when the number of classes is i in step S34.
The distance between the code C and the representative codes of all the classification classes is calculated. Then, in step S35, a classification class of the representative code having the shortest distance is obtained. This class is output as the classification class of the input pattern 21.

(Effects) (a) According to the present embodiment, the classification is performed by the class classification unit 24 with reference to the representative code (that is, the classification data) obtained by encoding the feature amount. Not directly affected by distribution. In addition, the coding unit 23 performs coding using the average and the standard deviation of the feature amounts, and the class classification unit 24 classifies the codes based on the defined distance. The influence of the measurement error on the class classification can be reduced. When the feature extraction fails, the feature amount is coded by giving a group number corresponding to the average value to classify the feature amount.

(B) In this embodiment, as shown in the equation (1), the value of the group number Gj is obtained by the calculation based on the threshold judgment, so that the calculation speed can be made higher than in the conventional case. Further, since the group number Gj can be digitally represented as 0, 1, 2, and 3, the number of bits is reduced and the handling is improved.

(Second Embodiment) (Configuration) FIG. 6 is a configuration diagram of a classification data generation portion in a pattern classification device according to a second embodiment of the present invention, and FIG. 1 shows the first embodiment. Elements common to those in the middle are denoted by common reference numerals. In the present embodiment, as in the first embodiment, the generation of the classification data does not refer to the codes of all the sample patterns, but refers to the result of investigating the types of the codes. For this reason, the encoding unit 13 and the classification data generation unit 14 shown in FIG.
1, a code type data generation unit (for example, a code type data generation unit) 16 and a code type data storage unit 17 are added, and a classification data generation unit having a configuration different from that of the classification data generation unit 14 in FIG. 14A is provided.

The code type data generation unit 16 investigates what type of code was generated for the sample pattern code obtained by the coding unit 13, and includes, for example, counters i, j, k, It is composed of a CPU having means and determination means. The code type data storage unit 17 stores the result of the investigation performed by the code type data generation unit 16, and is configured by a storage device. The classification data generation unit 14A generates classification data based on the survey result stored in the code type data storage unit 17, and includes counters i and j, a CPU having an arithmetic unit and a determination unit, and the like. . With respect to the configuration of the class classification portion in the pattern classification device of the present embodiment, the same configuration as the configuration of the class classification portion of the first embodiment shown in FIG. 1B can be applied.

(Processing of Classification Data Generation Portion in FIG. 6) FIG. 7
6 is a flowchart showing the processing of the code type data generation unit 16 in FIG. 6, and FIG. 8 is a flowchart showing the processing of the classification data generation unit 14A in FIG. Hereinafter, FIG.
6 will be described with reference to FIG. 8 and FIG. When the code of the sample pattern is obtained in the coding unit 13 as in the first embodiment, the code is sent to the code type data generation unit 16. The code type data generation unit 16 investigates what type of code is generated for the sample pattern code obtained in the encoding unit 13, and stores the result of the investigation in the code type data storage unit 17.

The processing of the code type data generation unit 16 is
This will be described with reference to the flowchart of FIG. First, in step S41, the code type counter k is initialized to 0, and then in step S42, the value of the data number counter i is initialized to 0. Next, in step S43, the same code check counter j is initialized to 0, and the process proceeds to step S44. If k = 0 in step S44, the code Ci of the first sample pattern is registered as the first type of code C'k in step S45, and the value of the counter k is incremented by +1.

On the other hand, if k ≠ 0 in step S44, it is checked in step S46 whether Ci is equal to the j-th type code C'j. If they are not the same, another type of code is checked by increasing the value of the counter j by +1 in step S47 and determining whether or not j has become k in step S48. If all types of codes do not match, Ci
Is a new type of code, in step S45, Ci is registered as the k-th type of code C'k, and the value of the counter k is incremented by +1. If Ci is equal to the j-th type code C'j in step S46, registration is not performed.

By repeating the above processing by adding +1 to the counter i in step S49 and determining whether or not i has become S in step S50,
Investigate the code of all sample patterns,
Find out how much code was generated. The obtained k types of codes C′k are stored in the code type data storage unit 17 in FIG.

The classification data generator 14A of FIG. 6 generates classification data by referring to the code type data stored in the code type data storage 17. This classification data generation unit 1
The processing of 4A will be described with reference to the flowchart of FIG. First, in step S51, the value of the counter i is set to 0, and in step S52, the value of the counter j is set to 0.
Initialize to Next, in step S53, the distance between the i-th type code i and the j-th type code j stored in the code type data storage unit 17 is calculated. The calculation of this distance is the same as the calculation performed in step S23 of FIG. 4 in the first embodiment. Then, in step S54, the minimum value Di of the distance between the i-th type code i and another type code j is updated.

The above processing is performed by adding +1 to the counter j in step S55 and determining whether or not j is the number of code types in step S56.
Is repeated. Further, by adding +1 to the counter i in step S57 and determining whether or not i is the number of code types in step S58, the process is repeatedly performed for all types of codes i. Then, in step S59, the upper N codes having the larger minimum distance value Di are selected. The selected code is stored in the classification data storage unit 15 in FIG. 6 as a representative code of the classification class (that is, classification data).

(Effect) In the present embodiment, the code of all the sample patterns is not examined when the classification data is generated as in the first embodiment. The type of the code is checked in advance and stored as code type data in the code type data storage unit 17, and the classification data is generated by the classification data generation unit 14A using the storage result. Therefore, the type of code is M ⁴ at maximum, does not depend on the number of sample pattern 1, it is possible to suppress an increase in processing time when the prepared sample pattern 1 a huge number.

(Third Embodiment) (Configuration) FIG. 9 is a configuration diagram of a class classification portion in a pattern classification device according to a third embodiment of the present invention, and FIG. Elements common to those in b) are denoted by common reference numerals.

In the class classification portion of the present embodiment, when classifying the input pattern 21, only one class is classified.
Rather than determining only one, the classification classes can be determined in order from the highest degree of matching as the classification destination. For this reason, a class order determining unit (for example, a class order determining unit) is used instead of the class classifying unit 24 in FIG.
25 are provided. The class order determination unit 25 refers to the classification data stored in the classification data storage unit 15 and determines the classification classes of the codes coded by the coding unit 23 in order from the highest degree of matching as the classification destination. And comprises, for example, a CPU having a counter i, a calculating means, and a determining means.

The configuration of the classification data generation portion in the pattern classification device of the present embodiment is the same as the configuration of the classification data generation portion of the first embodiment shown in FIG. 1A, or the second configuration shown in FIG. Any of the configurations of the classification data generation part of the embodiment can be applied.

(Processing of Class Classification Part of FIG. 9) FIG.
10 is a flowchart showing the processing of the class order determination unit 25 in FIG. Hereinafter, the processing content of the class classification portion of FIG. 9 will be described with reference to FIG. In the flowchart of FIG. 10, the distance Di between the code C generated from the feature amount of the input pattern 21 and the representative code Ci of the classification class is obtained by the same processing in steps S61 to S64 as the processing in steps S31 to S34 in FIG. 5. For all classification classes. Next, in step S65, the distance Di between the code C and the representative code Ci is arranged in ascending order. Then, in step S66, a permutation of the numbers i of the arranged distances Di is obtained. In this permutation, the classification classes are arranged in descending order of the degree of suitability as the classification destination.

(Effect) In the present embodiment, the class order determination unit 25 is provided, so that the classification classes of the input pattern 21 can be determined in order from the highest matching degree as the classification destination. This determines, for example, in the case where a recognition dictionary such as a large classification for character recognition is divided into a plurality of types, in which order the dictionary for recognizing the input pattern 21 should be applied. And so on.

(Fourth Embodiment) (Configuration) FIG. 11 is a configuration diagram of a class classification portion in a pattern classification device according to a fourth embodiment of the present invention.
Elements common to those in FIG. 9 showing the third embodiment are denoted by the same reference numerals. In the class classification portion of the present embodiment, pattern data that can specify the input pattern 21 such as the result of pattern recognition is used by using the pattern data of the third embodiment shown in FIG. By recording such information, statistical information of the class into which the input pattern 21 is classified can be provided. Therefore, a pattern data 18, a pattern-specific class recording unit (for example, a pattern-specific class recording unit) 26, and a pattern-specific class storage unit 27 are added.

The pattern-specific class recording section 26 uses the pattern data 18 to record information as to which class the input pattern 21 is classified into.
x, Px (Cx), CP having arithmetic means and determination means
U and the like. Class storage unit 27 by pattern
Stores the information of the class recording unit 26 for each pattern, and is constituted by a storage device.

(Processing of Class Classification Part of FIG. 11) FIG.
12 is a flowchart showing the processing of the pattern-specific class recording unit 26 of FIG. Hereinafter, the processing content of the class classification portion of FIG. 11 will be described with reference to FIG. In the class classification part of the present embodiment, the feature amount is extracted from the input pattern 21 by the feature amount extracting unit 22 and the coding unit 23
, The feature is coded. Further, the class order is determined by the class order determination unit 25, and the result of the determination is sent to the pattern-specific class recording unit 26.

The processing of the class recording unit 26 for each pattern will be described with reference to the flowchart of FIG. First, in step S71, it is checked whether or not information on the input pattern 21 has already been stored in the class storage unit 27 for each pattern. This can be executed by searching the pattern-specific class storage unit 27 for pattern data 18 (for example, a result of character recognition) specifying the input pattern 21. If stored, step S
At 72, the value of the counter Px representing the number of appearances of the pattern is increased by +1.
, The counter Px (Cx) for the class Cx determined to be most suitable for classification is increased by +1. The information on the input pattern 21 is stored in the class storage unit 2 for each pattern.
If not stored in step S73, a new information storage location is secured in the pattern-specific class storage unit 27, and the counters Px and Px (Cx) are set to 0 in step S73.
Initialize to

(Effect) In the present embodiment, the pattern data 18 which can specify the input pattern 21 such as the result of pattern recognition is used, and information on which class the input pattern 21 is classified into is recorded by the pattern-specific class recording section 26. By doing so, it is possible to provide statistical information of the class into which the input pattern 21 is classified. This statistical information is, for example, when the recognition dictionary is divided into a plurality of types, such as a large classification in character recognition, the recognition data is easily classified using the statistical information of the pattern. It can be moved to the class dictionary and used to optimize the contents of the recognition dictionary.

(Usage Form) The pattern classification method and pattern classification apparatus of the present invention can be widely used in the field of general pattern recognition such as character recognition and personal identification, and are used for character recognition and personal identification, for example. When the dictionary is divided, the dictionary can be used for determining a dictionary in which the recognition data of the input pattern 21 is to be registered, selecting a dictionary to be used when recognizing the input pattern 21, or optimizing the division of the registered dictionary. FIG. 13 is a configuration diagram illustrating a personal identification device that identifies an individual by the characteristics of an iris (iris) using, for example, the class classification portion of FIG. 9.

In the personal identification device using the feature of the iris, a face image is photographed using a camera, and the photographed image is converted into digital data to obtain an input pattern 21. The feature amount extraction unit 22 extracts the feature amount of the face such as the distance between eyes, the degree of opening of the eyes, the iris outer edge radius, and the like, and the extracted feature amount is encoded by the encoding unit 23. In this personal identification device, as a dictionary 29 for identifying an iris in the input pattern 21, a dictionary 29-1 relating to a first class defined by a feature amount of a face, a dictionary 29-2 relating to a second class, A plurality of dictionaries such as the dictionary 29-3 for the third class are prepared. Then, the class order determination unit 25 orders the codes of the feature amounts coded by the coding unit 23 into classes, and determines the order of the dictionary used at the time of matching, for example, 29-1, 29-2, and 29-3. I do.

The collating unit 28 first collates the contents of the dictionary 29-1 relating to the first class defined by the facial features with the code coded by the coding unit 23 to identify an individual. Do. If the individual cannot be identified, the dictionary is checked against the dictionary 29-2 to identify the individual.
If an individual cannot be identified with this dictionary, the dictionary is further collated with the dictionary 29-3 to identify the individual. As a result, the number of objects to be compared at the time of comparing the iris features in the matching unit 28 can be reduced while the number is initially small, and can be gradually increased. The result of the identification can be obtained.

[0058]

As described above in detail, according to claims 1 to 7 and 11 to 17 of the present invention, the following effects (1) to (4) can be obtained. is there. (1) Since a representative code may be stored instead of the representative value of the feature amount of the classification class, the storage capacity can be reduced. (2) Classification is performed with reference to a representative code obtained by encoding feature values, so that the distribution is not directly affected by the sample pattern distribution. In addition, the coding is performed using the average and standard deviation of each feature, and the classification is performed based on the distance defined for the obtained code. Can be reduced.

(3) If the feature extraction fails, a class number can be obtained by assigning a group number corresponding to the average value to the feature amount and coding the feature amount. (4) When generating classification data, instead of examining the codes of all sample patterns, the types of sample pattern codes are checked in advance and stored as code type data, and the classification data is used using the result. If it is generated, since the type of code does not depend on the number of sample patterns, it is possible to suppress an increase in processing time when an enormous number of sample patterns are prepared.

According to the eighth, ninth and eighteenth aspects of the present invention, a plurality of classification classes are ordered, so that the classification classes of the input pattern are ordered in descending order of the degree of suitability as the classification destination. Can be determined. According to the tenth and twentieth aspects of the present invention, information that identifies an input pattern is recorded by using data that can specify an input pattern such as a result of pattern recognition. Can provide statistical information for the class that was created.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a classification data generation part and a class classification part in a pattern classification device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a process of a coded data generation unit 11 of FIG.

FIG. 3 is a flowchart showing a process of a coding unit 13 in FIG. 1;

FIG. 4 is a flowchart showing a process of a classification data generation unit 14 of FIG.

FIG. 5 is a flowchart illustrating a process of a class classification unit 24 of FIG. 1;

FIG. 6 is a configuration diagram of a classification data generation portion in a pattern classification device according to a second embodiment of the present invention.

FIG. 7 is a flowchart illustrating a process of a code type data generation unit 16 of FIG. 6;

FIG. 8 is a flowchart illustrating a process of a classification data generation unit 14A of FIG. 6;

FIG. 9 is a configuration diagram of a class classification portion in a pattern classification device according to a third embodiment of the present invention.

FIG. 10 is a flowchart illustrating a process of a class order determination unit 25 of FIG. 9;

FIG. 11 is a configuration diagram of a class classification portion in a pattern classification device according to a fourth embodiment of the present invention.

FIG. 12 is a flowchart showing a process of a class recording unit for each pattern in FIG. 11;

13 is a configuration diagram of a personal identification device based on the characteristics of an iris using FIG. 9;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Sample pattern 2, 22 Feature amount extraction part 3 Feature amount database 4 Changeover switch 11 Coded data generation part 12 Coded data storage part 13, 23 Coding part 14, 14A Classification data generation part 15 Classification data storage part 16 Code type Data generation unit 17 Code type data storage unit 18 Pattern data 21 Input pattern 24 Class classification unit 25 Class order determination unit 26 Pattern class storage unit 27 Pattern class storage unit

 ────────────────────────────────────────────────── ─── Continuing from the front page (72) Inventor Takahiro Watanabe 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (72) Inventor Yasuhiro Cho 1-7-112 Toranomon, Minato-ku, Tokyo Inside Electric Industry Co., Ltd. (72) Inventor ▲ Takaji Koji 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. F-term (reference) 4C038 VA07 VB04 VC05 VC20 5L096 BA20 DA02 FA32 FA39 FA66 JA11 JA22 KA13 MA07

Claims (20)

[Claims] 1. A pattern classification method for classifying an input pattern into one of a plurality of classification classes according to a characteristic amount thereof, wherein an input code is generated by coding the input pattern according to the characteristic amount. A pattern classification, wherein the classification class is expressed by a representative code having a similar form, and a classification class for classifying the input pattern is determined by calculating a distance between the input code and the representative code. Method. 2. The pattern classification method according to claim 1, wherein the coding of the input pattern is performed by a statistic obtained from a feature amount of a sample pattern collected in advance. 3. The pattern classification method according to claim 2, wherein the coding of the input pattern is performed using an average and a standard deviation of the feature amounts of the sample pattern. 4. The method according to claim 1, wherein the representative code is selected by calculating a distance between the sample codes from a sample code obtained by coding a sample pattern collected in advance. The pattern classification method according to any one of Items 1 to 3. 5. The pattern classification method according to claim 4, wherein the representative code is selected from the sample codes in order of a minimum value of a distance from another sample code. 6. The method according to claim 4, wherein a distance between the sample codes and a distance between the input code and the sample code are defined by statistics obtained from characteristic amounts of the sample pattern. The described pattern classification method. 7. The classification class according to claim 1, wherein the classification class for classifying the input pattern is a classification class of the representative code having the shortest distance from the input code. Pattern classification method. 8. The method according to claim 1, wherein the ranking of the plurality of classification classes is performed by calculating a distance between the input code and the representative code. Pattern classification method. 9. The pattern classification method according to claim 8, wherein the ranking of the plurality of classification classes is performed in ascending order of the distance between the input code and the representative code. 10. The pattern classification method according to claim 1, wherein a frequency of a classification class into which the input pattern is classified is stored for each of the input patterns. 11. A pattern classifying apparatus for classifying an input pattern into one of a plurality of classification classes according to a feature amount thereof, comprising: coding means for generating an input code by coding the input pattern according to the feature amount; A classification data generating unit that generates classification data expressing the classification class with a representative code having the same form as the input code; a classification class that classifies the input pattern by referring to the classification data; A class classification means for calculating and determining a distance from the representative code. 12. The pattern classification apparatus according to claim 11, wherein said coding means performs coding using a statistic obtained from a feature amount of a sample pattern collected in advance. 13. The pattern classification apparatus according to claim 12, wherein said coding means performs coding using an average and a standard deviation of the feature amounts of said sample pattern. 14. A configuration in which the classification data generating means selects the representative code by calculating a distance between these sample codes from a sample code obtained by coding a sample pattern collected in advance. The pattern classification device according to any one of claims 11 to 13, wherein: 15. The classification data generating means according to claim 14, wherein said representative code is selected from said sample codes in order of a minimum value of a distance from another sample code. A pattern classification device as described. 16. The method according to claim 1, wherein a distance between the sample codes and a distance between the input code and the sample code are defined by statistics obtained from feature amounts of the sample pattern. Item 15. The pattern classification device according to Item 14. 17. The apparatus according to claim 11, wherein said class classification means is configured to determine, as a classification class for classifying said input pattern, a classification class of said representative code having the shortest distance from said input code. The pattern classification device according to any one of Items 1 to 16. 18. The ranking of the plurality of classification classes,
The pattern classification apparatus according to any one of claims 11 to 17, further comprising a class order determining unit that calculates a distance between the input code and the representative code. 19. The apparatus according to claim 18, wherein said class order determining means ranks said plurality of classification classes in ascending order of a distance between said input code and said representative code. Pattern classification device. 20. The apparatus according to claim 11, further comprising pattern-specific class recording means for storing, for each of said input patterns, a frequency of a classification class into which said input pattern is classified.
20. The pattern classification device according to any one of claims 19 to 19.