JP2000076382A - Learning method of hierarchical dictionary to use for character recognition and preparation of the dictionary - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the processing speed of character recognition by allowing a classifying dictionary being a high-order hierarchy to reduce the number of sorting candidates belonging to each. SOLUTION: Processing of inputting a learning database in a system, continually executing detailed sorting to a pattern and writing the number of a template for detailed sorting being a correct answer in a tag is repeated sample number of times (S101 to S104). By this processing, the number of the template for detailed sorting which each leaning pattern should correctly hit is stored. Next, processing is shifted to learning of a sorting template. First, the learning database is inputted to the system, the sorting template of an input pattern is continually evaluated and reading processing of a template number for detailed sorting written in the tag of the input pattern is repeated the times of the number of inputted samples (S113 to S115). Then, processing of obtaining the sorting template being a correcting object for learning is executed finally (S116).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of learning a hierarchical dictionary used for character recognition which analyzes an image and recognizes it as a character string, and more particularly to a dictionary for a case where a hierarchical dictionary is used for character recognition. And a learning method of a hierarchical dictionary suitable for improving the accuracy of character recognition by performing the learning.

[0002]

2. Description of the Related Art Hitherto, a method of using a hierarchical dictionary for character recognition has been proposed. For example, a method of writing a large classification of handwritten kanji based on clustering by Fumitaka Kimura (IEICE) has been proposed. There is a journal, Vol. J70-D No.1). In this method, the recognition process has a two-stage hierarchical structure of classification and detailed classification, and the classification is performed to narrow down the number of objects to be handled in the detailed classification, thereby reducing the processing amount and improving the processing speed. Here, the longest distance method, one of the clustering methods, is applied to the creation of a hierarchical dictionary, and the created upper-level dictionary (classification dictionary) is used. It is reported to decrease from about 1/3 to about 1/4 in comparison. Here, there is no function to automatically feed back the errors generated by the classification and to learn the hierarchical dictionary, and the method of creating a hierarchical dictionary for character recognition found in other conventional technologies is also hierarchical. I don't see the idea of learning a new dictionary.

[0003]

The method of using a dictionary having a hierarchical structure for character recognition described in the above prior art is an effective method for reducing the amount of calculation for character recognition. However, there is a problem that the recognition accuracy will not be improved unless the number of classification candidates to be sent from the classification to the detailed classification is increased. This is because if the number of detailed classifications belonging to one class is small, erroneous recognition may increase. For this reason, it is customary to take a relatively large number of classification candidates to be sent from the classification to the detailed classification. (In other words, this means that the number of classifications cannot be increased.) For example, in the case of recognizing handwritten kanji, one class requires several hundred classification candidates. . Therefore, it is necessary to perform detailed classification on hundreds of candidates belonging to one class,
The amount of calculation is not necessarily small. Therefore, it is necessary to reduce the amount of recognition at the stage of detailed classification, to improve the processing speed, and at the same time, to reduce errors caused by the classification. Thus, high-speed recognition processing can be realized without lowering the accuracy due to the classification.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problem, and an object of the present invention is to make it possible to learn a dictionary of a classification which is an upper hierarchy when a hierarchical dictionary is used for character recognition. The present invention provides a method of learning a hierarchical dictionary that reduces the number of classification candidates belonging to each one, improves the processing speed of character recognition, has high classification accuracy, and does not cause erroneous recognition. It is in.

[0005]

In order to achieve the above object, the present invention relates to a method for learning a hierarchical dictionary used for character recognition according to the present invention, comprising a character string for recognizing a read image as a character string. In the learning method of the hierarchical dictionary used for recognition, a dictionary composed of templates for matching as an image pattern for character recognition is configured as a hierarchical dictionary composed of a hierarchy of a lower detailed classification dictionary and an upper classifying dictionary. The lower-level detailed dictionary is constructed using a learning database, and the upper-level classification dictionary is a dictionary including a template classified based on the characteristics of the template of the lower-level detailed dictionary. The learning method of the hierarchical dictionary used for recognition is
(1) For the sample pattern in the learning database,
A step of writing a template number for identifying a template of the dictionary for detailed classification in a tag added to the pattern; (2) determining a similarity between the input sample pattern and a template in a hierarchy of a classification dictionary; (3) a procedure for obtaining a template of a classification dictionary of a higher class to which the template number belongs from the template number described in the procedure of the above (1); (4) a procedure for obtaining the template in the above (3) Other than the obtained template, a procedure for obtaining a template of a classification dictionary determined to have a high degree of similarity to the sample pattern in the procedure of (2); (5) a classification dictionary of the classification dictionary obtained in the procedure of (3); The template is set as a correct template, and the template of the classification dictionary obtained in the procedure (4) is used as an obstructing template. As a rate, the input sample pattern has a procedure of correcting a correct template so as to increase the similarity, and correcting the disturbing template so that the similarity decreases. After the execution of the procedure, the above (2) is performed for the determined number of times of learning and for the number of input sample patterns.
The steps (5) to (5) are performed in this order to learn the template of the classification dictionary.

More specifically, in the learning method of the hierarchical dictionary used for character recognition, the dictionary composed of templates for matching as an image pattern for character recognition is configured as N stages (N ≧ 2 integers). ),
The upper hierarchical dictionary is configured so as to include the lower hierarchical dictionary as a tree relation, and when learning the N-th dictionary, using the (N−1) -th dictionary lower than that,
Applying the detailed classification dictionary described above to the lower-order N-1 dictionary, applying the classification dictionary described above to the N-th dictionary, and as described above,
The steps (1) and (2) to (5) are executed in this order.

[0007] To achieve the above object, the present invention relates to a method for creating a hierarchical dictionary used for character recognition according to the present invention.
A method for creating a hierarchical dictionary used for character recognition as described above, wherein a similarity of the template is evaluated by a distance, and a template having a shorter distance has a higher similarity. And (10) a procedure for evaluating whether or not the number of the detailed classification templates belonging to a certain class is equal to or more than a predetermined number, and (11) a new procedure. A procedure for generating a template in the vicinity of the template of the class, and after executing the procedure of (L), the procedure of (10) is executed so that the number of the detailed classification templates belonging to a certain class is a fixed number. If the class is evaluated as above, (1
After the procedure of 1) is performed, the procedure of (L) is performed, and the above procedure is repeated until the number of the detailed classification templates belonging to all the classes becomes equal to or less than a certain number. It is.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments according to the present invention will be described below with reference to FIGS. [Outline of Pattern Recognition Process] First, an outline of a process of recognizing a pattern as a character will be described with reference to FIG. FIG. 2 is a flowchart showing an outline of processing for recognizing a pattern as a character.

First, a pattern as an image is input from an input device such as a scanner (S201).

Next, preprocessing such as noise removal and normalization of the size of the input pattern is performed (S202). here,
The noise removal is processing for removing dirt and blurring other than the character to be recognized, and the size normalization is processing for making the size of the character to be recognized uniform and constant (S
203).

After the pre-processing is completed, the features used in the matching processing (for example, vertical / horizontal / right diagonal /
A feature extraction is performed to extract a left oblique direction feature from the normalized pattern (S203).

Then, first, a classification is performed to determine what class this pattern belongs to (S20).
4) Then, detailed classification is performed (S205). When the detailed classification is completed, the character recognition processing is completed.

Here, the procedure of class classification and detailed classification is as follows.
This is a process of matching a template stored therein (features extracted and processed in advance from a learning pattern) with features extracted from an input pattern to find a template closest to the input pattern. For this purpose, a dictionary is prepared for each procedure.

Both processes are still pattern matching, but the reason for providing a hierarchy in the pattern matching process in this way is that if detailed classification is performed suddenly, the number of matching processes to be processed increases, which is practical. This is because it is difficult to process in a short time. Therefore, prior to performing the detailed classification, the classification is performed to narrow down the template, reduce the number of matching processes performed in the detailed classification, and reduce the processing time as a whole.

Next, FIGS. 3 and 4 show the concept of the present invention.
The concept of the present invention will be described with reference to FIG. FIG. 3 is a schematic diagram showing the relationship between the classification template and the detailed classification template. FIG. 4 is a schematic diagram for explaining the principle of the learning vector quantization method.

First, a classification process will be described as an introduction.

The templates used in the classification are formed by using similar clustering techniques such as the longest distance method, collecting similar detailed classification templates as the same class, and creating one representative from the detailed classification templates belonging to the same class. You. Therefore, the classification template 7
FIG. 3 and the detailed classification template 730 are shown in FIG.
It becomes as shown in. That is, for example, similar detailed classification templates 713 to 717 are collected as one and the same class, and these detailed classification templates 713 are used.
From 717 to 717, one classifying template 703 is created as a representative, and the tree relationship is as shown in the figure.

In the classification process, all the classification templates 701 are matched with features obtained from the input patterns, and the top N candidates are output for the next step of detailed classification. Therefore, if the number of candidates to be sent to the detailed classification procedure as the output of the classification procedure is small, the number of templates to be matched in the detailed classification is reduced, and the processing time can be reduced. Conversely, if the number of candidates is reduced, there is a possibility that the classification accuracy will cause the detailed classification template to be the correct answer to be absent in the candidates and the recognition accuracy to be reduced. That is, it is necessary to learn a hierarchical dictionary used for classification so that the detailed classification template to be the correct answer is included in a small number of candidates.

In order to realize this, it is sufficient to construct a classification having high classification accuracy. For this reason, in the present invention, a "learning vector quantization method" (T. Kohonen, Self-Organization an), which is a dictionary learning method that is known to achieve high recognition accuracy in detailed classification.
d Associative Memory, Springer-Verlag, 1989) or "General learning vector quantization method" including it (Atsushi Sato, Character Recognition by General Learning Vector Quantization, IEICE, Technical Report, PRU95-219, 1996) -03)) can be applied to learning of hierarchical dictionaries used in classification, and the purpose is to construct classification with high classification accuracy.

Therefore, next, a conventionally known “Learning Vector Quantizatio
n) ”Let me explain the algorithm of LVQ 2.1.

Here, the feature extracted from the input pattern can be represented by a vector, which is represented by x, and the character type to which the input pattern belongs is represented by C (x). The templates w _i , w _j and the character types to which they belong are C (w _i ) and C (wj). In the learning vector quantization method, whether a pattern is similar or not similar is represented by a numerical value as a vector distance. That is, when the patterns are similar, the distance between the vectors is determined to be small.

Here, suppose that the template closest to the input pattern x is w _i , and the template w
_Suppose that it becomes _j . Here, the template closest to the input pattern x is called a first candidate template, and the template closest to the input pattern x is called a second candidate template. If it is close to the input pattern and not the correct character type, the template is said to be "interfering".

Since the character type of the correct answer is C (x), C (x) ≠ C
When (w _i ) and C (x) = C (w _j ), the template w _j
Is a template of the second candidate in reality, though it belongs to the correct character type and therefore needs to be short in distance. In other words, the closest template w _i has become a disturbance. Therefore, at this time,
Modify template w _i and template w _j so that template w _j is closest to input pattern x. This means that the template is made to learn.

More specifically, as shown in FIG.
Replace template w _i 802 with □ modified template
The template w _j 803 is changed to the modified template w _j 806 of □ in w _i 805. Thereby, the identification boundary 8
04 becomes 807. Thus, the template w _j 806 in which the template closest to the input pattern x is modified
, And the character types match and can be recognized correctly.

If these algorithms are expressed by equations, the following (Equation 1) is established,

[0026]

[Formula 1] C (x) ≠ C (w _i ) and C (x) = C (w _j ) (Equation 1) and within an area set near the identification boundary between template w _i and template w _j When x exists, the templates w _i and w _j are corrected according to the following (Equation 2) and (Equation 3).

[0027]

W _i ← w _i + k (x−w _i ) (Equation 2) w _j ← w _j + k (x−w _j ) (Equation 3) However, 0 <k <1.

According to the present invention, based on the learning vector quantization method described above, a template used for classifying is
It is intended to be applied as a learning target.

[Embodiment 1] Next, the processing procedure of the first embodiment according to the present invention will be described with reference to FIG. 1, FIG. 3, FIG. 5, and FIG. FIG. 1 is a flowchart showing the procedure of the hierarchical dictionary learning method according to the present invention. FIG. 5 is a flowchart showing in detail a process of selecting a classifying template to be corrected. FIG. 6 is a diagram showing a table showing the template numbers for details and the corresponding classification numbers.

First, in order to carry out the hierarchical dictionary learning method of the present invention, it is necessary to create a detailed classification dictionary and a classification dictionary which is a layer above it. Since these creating methods are well-known techniques, the description of these creating methods is omitted, and it is assumed that these dictionaries have already been created.

FIG. 1A shows a procedure for creating data required for learning a classification template according to the present invention, and FIG. 1B shows a procedure for learning the classification template. The procedure is shown.

In the above-described learning vector quantization method LVQ2.1, the character type to which the input pattern belongs is determined, and the template corresponding to the character type is modified accordingly. It should be noted that in the present invention, the class to which the input pattern belongs is the class, and the object to be corrected is to learn a classification template.

Hereinafter, description will be made based on a specific example shown in FIG.

As shown in FIG. 1A, first, a learning database is input to the system (S101).

Then, the next process is repeated several times for the samples (S102). In the repetition processing, first, detailed classification is performed on the pattern (S103), and the number of the correct detailed classification template is described in the tag (S104).

For example, when a pattern similar to “circle” is input, a template of the detailed classification is searched, and the template number of the “circle” template 711 is described in the tag. Here, the tag is a storage area provided corresponding to each input pattern.

In the above processing, the number of the detailed classification template to which each learning pattern should hit correctly is stored. That is, the number of the template described in this tag is the template number for the detailed classification that is regarded as the correct answer.

Next, learning of the classification template is executed according to the procedure shown in FIG.

First, a learning database is input to the system (S111).

Next, the number of times of learning the following processing (S112) and the number of input samples (S11
3) Will repeat. Here, the practical number of times for performing the learning is determined from the viewpoint of the effect of the learning and the improvement of the accuracy of the classification.

The following is an application of the learning vector quantization method.

First, the classification template of the input pattern is evaluated (S114). That is, the similarity between the vector of the input pattern and the vector of the class template is evaluated. This is a procedure for evaluating the degree of similarity with each template in terms of distance as seen as a vector, as described in the algorithm of the learning vector quantization method. Here, it is assumed that a “circle” pattern is input and the classification template 704 is evaluated to be the closest.

Next, the template number for the detailed classification described in the tag of the input pattern is read (S115). This was the template number of the "circle" template 711.

Next, there is a process of obtaining a classifying template to be corrected for learning (S11).
6).

This is a two-step process as shown in FIG.

First, from the correct detailed classification number of the read input pattern, the classification number to which it belongs (this is
L), and this template is used as a correct classification template (S301).

The class including each of the detailed classification templates can be obtained by using a table shown in FIG. 6 created in advance. In this table, the classification number 402 can be subtracted from the detailed classification template number 401. This is shown in FIG.
Since the classification template and the detailed classification template have a tree-like relationship, the classification template 701 to which each of the detailed classification templates 730 belongs.
Can be known and can be easily created.

In this example, the template 702 of the class to which the “circle” template 711 belongs is a template for classifying which is the correct answer.

Next, the class number of the closest distance to the input pattern excluding CL is obtained, and this template is used as a disturbance classification template. In this example, the template of the classification template 704 corresponds to this.

Next, the classification template is modified. This is done by modifying the obstructive classification template 704 so as to move it away from the input pattern and bring the correct classification template 702 closer to the input pattern, as already described in the learning vector quantization method. I do.

According to the description symbols of the learning vector quantization method described above, the correct template w _j is used as the classification template 702 and the template
w _i corresponds to the classification template 704. If this is applied to the above algorithm, the classification template is corrected in a desired direction, and learning is performed on the classification template.

[Second Embodiment] Next, a processing procedure of a second embodiment according to the present invention will be described with reference to FIG. FIG. 7 is a flowchart illustrating a processing procedure of a learning method when a dictionary has multiple stages.

In the first embodiment, a dictionary composed of two-stage templates is classified into a detailed classification and a higher-level classification, and a learning vector quantization method is used. Was reflected in the learning of the higher hierarchy.

This procedure is not limited to two steps, but can be applied to learning a dictionary composed of more layers. That is, when learning the dictionary of the N-th level,
The learning data on the dictionary of the (N-1) -th level one step lower is used.

Hereinafter, description will be made in order of FIG.

First, a learning database is input to the system (S501).

Then, the following processing is repeated as many times as the number of levels of the hierarchical processing (S502). Here, it is assumed that the number of stages is N, and a variable that is repeatedly performed is i. And i is 0
To N-2.

In the repetition processing, first, a hierarchical dictionary of the (N-i) th stage is created (S503).

Next, learning data is created by using the hierarchical dictionary of the (Ni-1) th stage in the same manner as the processing procedure of the flowchart of FIG. 1A (S504). That is, it is checked whether or not the input pattern belongs to the template of the lower dictionary, and the template number of the (N-i-1) th hierarchical level that is the correct answer is described in the tag.

Then, using the learning vector quantization method,
The (N-i) th hierarchical dictionary is learned by the same method as the processing procedure of FIG. 1B (S505). That is, (N−i)
In the templates belonging to the hierarchical dictionary at the lower level, search for the correct template and the template to be disturbed.
Correction is made so that the distance of the correct template is closer to the input pattern and the template to be disturbed is further away.

Finally, the created dictionary is stored (S50).
6).

[Embodiment 3] Next, the processing procedure of a third embodiment according to the present invention will be described with reference to FIG. FIG.
13 is a flowchart illustrating a processing procedure of a method of creating a hierarchical dictionary using an LBG clustering algorithm.

The first embodiment and the second embodiment relate to a method for learning a hierarchical dictionary, but the present embodiment relates to a method for creating a hierarchical dictionary.

In general, in order to reduce the amount of calculation in the recognition processing, the number of detailed classification templates belonging to a class is averaged, and there is no exceptionally large number of detailed classification templates belonging to a specific class. Is desirable. For this reason, in the present embodiment, when creating a hierarchical dictionary, it is devised not to have a class having a certain number or more of detailed classification templates.

The hierarchical dictionary is created using a clustering algorithm. In the present embodiment, for the above purpose, the LBG clustering algorithm (Y. Linde, A. Buzo, and RM Gray, An Algorithm f
or Vector Quantizer Design, IEEE Trans. on Comm.,
Vol. COM-28, No.1, pp.84-95) is used. This LBG clustering algorithm is known to have high classification performance among the clustering algorithms, and furthermore, it is possible to avoid a class having a large number of detailed classification templates from being created so as to achieve the above-mentioned object. it can.

First, the outline of the LBG clustering algorithm will be briefly described.

When the LBG clustering algorithm is applied, first, an appropriate number of templates are set as initial values. Then, two processes of dividing the clustering sample by Voronoi division, finding the center of gravity of each division, and updating this as a new template are repeated. In this repetitive processing, first, the sum (distortion) of the distance between the template and the clustering sample is obtained in each division, and further, the sum of the distortions in all divisions is obtained. This is called the total distortion, and the LBG clustering algorithm repeats until the total distortion becomes equal to or less than a certain value.

Here, the template described above corresponds to a template for classification, and the clustering sample corresponds to a template for detailed classification. In the present invention, if the result of the LBG clustering algorithm is M or more detailed classification templates belonging to a specific class, a new template is set near an existing template, and the LBG clustering algorithm is performed again. Can divide that particular class into two.

Hereinafter, a method of creating a hierarchical dictionary using the LBG clustering algorithm will be described with reference to FIG.

First, the variable Flag is set to 0 (S60).
1).

Next, the LBG clustering algorithm is executed (S602).

Then, the processes from S604 to S606 are performed several times (i) in the class (S603).

In the repetition processing, first, it is determined whether the number of the detailed classification templates belonging to the i-th class is M or more (S604).

If it is not less than M, a new template is generated near the i-th classification template (S
605), and sets the variable Flag to 1 (S60)
6). After the repetition of S603, it is determined whether the variable Flag is 0 or 1 (S607).

If the variable Flag is 1, a new template for classifying has been created.
01 (S608), and the processing up to this point is repeated again.

When the variable Flag is 0, the number of detailed templates belonging to all classes is smaller than M, and the process is terminated.

[0077]

According to the present invention, when a hierarchical dictionary is used for character recognition, the number of classifying candidates belonging to each class is reduced by learning a dictionary for classifying which is a higher hierarchy. It is possible to provide a method of learning a hierarchical dictionary that can improve the processing speed of character recognition while reducing the number, and that has high classification accuracy and does not cause erroneous recognition.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a procedure of a hierarchical dictionary learning method according to the present invention.

FIG. 2 is a flowchart showing an outline of processing for recognizing a pattern as a character.

FIG. 3 is a schematic diagram showing a relationship between a classification template and a detailed classification template.

FIG. 4 is a schematic diagram for explaining the principle of a learning vector quantization method.

FIG. 5 is a flowchart showing in detail a process of selecting a classifying template to be modified.

FIG. 6 is a diagram showing a table showing a template number for details and a classification number corresponding thereto.

FIG. 7 is a flowchart showing a processing procedure of a learning method when a dictionary has multiple stages.

FIG. 8 is a flowchart illustrating a processing procedure of a method of creating a hierarchical dictionary using an LBG clustering algorithm.

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Hiroshi Shinjo 1-280 Higashi-Koigabo, Kokubunji-shi, Tokyo Inside the Central Research Laboratory of Hitachi, Ltd. (72) Inventor Kazuki Nakajima 1-280 Higashi-Koigabo, Kokubunji-shi, Tokyo Hitachi, Ltd. Central Research Laboratory (72) Inventor Tateyuki Sugimoto 2880 Kozu, Odawara-shi, Kanagawa Prefecture Hitachi, Ltd.Storage Business Unit (72) Inventor Seiji Ida 2880 Kofu, Odawara-shi, Kanagawa Prefecture Hitachi Storage Device Business F-term ( Reference) 5B064 AB16 DA00 DA05 DA14 DA27 DC07

Claims (3)

[Claims] In a method of learning a hierarchical dictionary used for character recognition for recognizing a read image as a character string, a dictionary including a template for matching as an image pattern for recognizing a character is stored in a lower order. And a hierarchical dictionary including a hierarchy of a higher-level classification dictionary. The lower-level detailed dictionary is constructed using a learning database, and the upper-level classification dictionary includes the lower-level detailed classification. This is a dictionary composed of templates classified into classes based on the features of the dictionary template. The learning method of the hierarchical dictionary used for character recognition is as follows. Steps to write the template number to identify in the tag attached to this pattern,
(2) a procedure of quantifying and evaluating the similarity between the input sample pattern and the template in the hierarchy of the classification dictionary, and (3) a template number based on the template number described in the procedure (1). To find a template for the classifying dictionary of the higher class to which the
(4) A procedure for obtaining a template of a classification dictionary that has a similarity with the sample pattern in the procedure of (2) other than the template obtained in (3), (5) the (3) The template of the classification dictionary obtained in the procedure of (4) is set as a correct template, and the template of the classification dictionary obtained in the procedure (4) is set as an obstructing template. The method has a procedure of modifying the similarity to be large, and modifying the disturbing template so that the similarity is small. After the execution of the procedure (1), a predetermined number of learnings is performed. Also,
Perform steps (2) to (5) in this order for the number of input sample patterns.
A learning method of a hierarchical dictionary used for character recognition, wherein learning of a template of a classification dictionary is performed. 2. The structure of a dictionary composed of templates for matching as image patterns for character recognition is defined as N stages (N ≧ 2 integers). In order to learn the dictionary of the N-th stage, the dictionary for the detailed classification according to the claim 1 is used by using the dictionary of the (N-1) -th stage below the dictionary. Lower N
Applying the classification dictionary described in claim 1 to the N-th dictionary, as described in claim 1, (1) and (2) to (2) A method of learning a hierarchical dictionary used for character recognition, characterized by performing the procedure of (5) in this order. 3. The method for creating a hierarchical dictionary used for character recognition according to claim 1, wherein the similarity of the template is evaluated by a distance, and when the similarity is determined to be higher for a template having a shorter distance. , (L) a procedure for executing the LBG clustering algorithm to classify the detailed classification templates into classes to which they belong, and (10) evaluating whether the number of the detailed classification templates belonging to a certain class is equal to or more than a certain number. Procedure, (11) a procedure for generating a new template in the vicinity of the template of the class. After execution of the procedure of (L), the procedure of (10) is executed to execute detailed classification belonging to a certain class. If the number of templates for use is evaluated to be equal to or more than a certain number, (11)
After executing the procedure of (L), the procedure of (L) is executed, and the above procedure is repeated until the number of the detailed classification templates belonging to all the classes falls below a certain number. A method of creating a hierarchical dictionary used for recognition.