JPH11154231A - Pattern recognition dictionary learning method, pattern recognition dictionary creation method, pattern recognition dictionary learning device, pattern recognition dictionary creation device, pattern recognition method, and pattern recognition device - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To provide a pattern recognition learning method that minimizes the adverse effects of learning and makes use of the advantages of each recognition method in dictionary learning in pattern recognition processing using a plurality of recognition methods. thing. In a method of learning a recognition dictionary in a pattern recognition process using a plurality of recognition methods, a pattern recognition process is performed on an input pattern, and at least a recognition result of the pattern recognition process and a recognition result of each recognition method are used. A learning criterion for the recognition dictionary is determined, and the recognition dictionary is learned according to a recognition method and a learning method for learning the recognition dictionary determined by the learning criterion.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a pattern recognition learning method and a pattern recognition creating method for learning a dictionary in pattern recognition, a pattern pattern recognition dictionary learning device for creating a dictionary in pattern recognition, and a pattern recognition dictionary creation. The present invention relates to a device, a pattern recognition method and a pattern recognition device using the dictionary.

[0002]

2. Description of the Related Art There are known various recognition methods in pattern recognition and methods for learning or creating a dictionary thereof. For example, the document “The Self-Org”
anizing Map ", Proceedings
of the IEEE (1990) includes LVQ (L
learning Vector Quantizatio
n) is introduced. This assumes that the input vector is x,
For each category, m is defined as a reference vector, and the difference is defined as s, and s = {x−m} is defined. The category having the highest degree of difference among the categories is output as the recognition result of x.

At this time, the following procedure is used for learning m.
Category obtained as the recognition results when recognizing the training pattern x c, a reference vector as a _{m c, m c '= m} c + α (x-m c) (1) m c' = m c -α ( x- _mc ) Update the dictionary as in (2).

There are LVQ2 and LVQ3 as methods other than the above, and based on the difference obtained as a recognition result, equation (1)
And equation (2). In addition, GPD (Gene
rarized Probabilistic Dec
A method of applying a concept called “ent) to the Euclidean distance is also known. R (x) is a recognition result obtained when x is recognized, and α (R (x)) is a coefficient determined by R (x). In general, _mc ′ = _mc + α (R ( x)) (x- _mc ) as the update rule.

[0005] The document "Pattern Recognition and Subspace Method", Sangyo Tosho (1986), introduces a learning subspace method (LSM). This defines s = x ^T Px, where x is an input vector, P is projection for each category, and similarity is s. In addition, x ^T is the transpose matrix of x. The category having the highest similarity in each category is output as the result of x recognition.

[0006] When the orthonormal vectors constituting the P was set to [psi _i, shall be expressed as P = GG ^T to a G [psi _i as the r arranged matrix. Let R (x) be the recognition result obtained when x is recognized, and α (R (x)) be a coefficient determined by R (x), and ψ _i ′ = ψ _i + α (R (x)) (x, ψ) x, and learning is performed by orthogonally transforming た と え ば_i ′ by, for example, the Schmidt orthogonalization method to obtain a new ψ _i ′.

[0007] format with a mean vector ^{m, s = ‖x-m‖ 2} - (x-m) T P (x-m) is also known as projection distance, the [psi _i as above is M ′ = m + 2α (R (x)) ((x−m) −P (x−
m)) is also known.

The average subspace method (ALSM) and JP-A-56-137483 introduced in the same document are as follows. Using K as a positive symmetric matrix, P is constructed using a part of the eigenvectors of K, and the similarity is expressed as s = x ^T
Defined as Px.

In the learning, when the recognition is incorrect, K is updated by K ′ = K + αxx ^T (for the correct category) K ′ = K−βxx ^T (for the incorrect category) Done by

[0010] By the way, in a pattern recognition device that has been already put into practical use, such as a character recognition device or a voice recognition device, it has been required to improve the efficiency of the operation from the input of the pattern to the correction of the erroneous recognition result after the recognition process. In order to respond to this, it is necessary to increase the speed and accuracy of the recognition process. However, it is difficult to realize this by a recognition process using one recognition method.

Therefore, various pattern recognition methods using a plurality of recognition methods integrated have been considered. That is,
Since each recognition method has its advantages and disadvantages, it is expected that excellent recognition processing utilizing the advantages of each recognition method can be realized by properly using the recognition methods according to the integrated method.

When dictionary learning is performed in the pattern recognition method using a plurality of recognition methods, conventionally, an incorrect answer pattern or a reject pattern is learned for all the recognition method dictionaries.

However, some of the plurality of recognition methods perform the dictionary learning as exemplified above, for example, for a pattern having a high degree of ambiguity between other categories, so that the dictionary is not recognized. The balance may be lost, and the correct pattern may become incorrect. Because of the adverse effects of such learning, it is difficult to realize dictionary learning that can fully utilize the advantages of each recognition method, and to fully exploit the recognition performance of the entire pattern recognition method using multiple recognition methods. Could not.

[0014]

As described above, conventionally,
As a dictionary learning method for a pattern recognition method using a plurality of recognition methods, there is no learning method capable of learning a dictionary so as to make use of the characteristics of the plurality of recognition methods.

The present invention has been made in consideration of the above circumstances, and minimizes the adverse effects of learning in dictionary learning in pattern recognition processing using a plurality of recognition methods.
An object of the present invention is to provide a pattern recognition learning method, a pattern recognition creation method, a pattern recognition dictionary learning device, a pattern recognition dictionary creation device, a pattern recognition method, and a pattern recognition device that make use of the advantages of each recognition method.

[0016]

Means for Solving the Problems First, terms are defined. The learning criterion is a parameter that can be deduced by a certain algorithm from a recognition result such as the similarity of the first candidate, the category of the first candidate, or the difference between the similarities of the first and second candidates. Say. The learning guideline is an algorithm that determines a dictionary to be learned and a learning method according to a condition with respect to a learning criterion.

The present invention (claim 1) is a pattern recognition dictionary learning method for learning a recognition dictionary in a pattern recognition process using a plurality of recognition methods, wherein the method is determined based at least on the recognition result of each recognition method for an input pattern. Learning the recognition dictionary according to the obtained learning criterion.

According to a second aspect of the present invention, there is provided a pattern recognition dictionary learning method for learning a recognition dictionary in a pattern recognition process using a plurality of recognition methods, wherein the pattern recognition process is performed on an input pattern. And a learning criterion for the recognition dictionary is determined based at least on the recognition result of each recognition method and the recognition result of each recognition method, and the recognition dictionary is learned according to the determined learning criterion.

[0019] Preferably, a learning guideline depending on an integrated system of the plurality of recognition systems in the pattern recognition process may be determined. Preferably, the learning criterion defines a learning guideline, and learning may be performed according to the learning guideline.

Preferably, the learning criterion includes a difference in similarity between the recognition candidate and the correct candidate in the recognition method, a similarity of the correct candidate, a similarity of the recognition candidate, a recognition candidate category, a correct candidate category, and a candidate rank of the correct candidate. At least one of
May be included.

[0021] Preferably, the learning criterion may determine a recognition method and a learning method for learning a recognition dictionary. The present invention (Claim 6) is a pattern recognition dictionary creating method for creating a recognition dictionary in a pattern recognition process using a plurality of recognition methods by learning an initial recognition dictionary. And learning the recognition dictionary in accordance with a learning criterion determined based at least on the recognition result.

The present invention (claim 7) is a pattern recognition dictionary creating method for creating a recognition dictionary in a pattern recognition process using a plurality of recognition methods by learning an initial recognition dictionary. Executing a pattern recognition process, determining a learning criterion of the recognition dictionary based on at least the recognition result of the pattern recognition process and the recognition result of each recognition method, and learning the recognition dictionary according to the determined learning criterion. Features.

[0023] Preferably, a learning guideline depending on an integrated system of the plurality of recognition systems in the pattern recognition processing may be determined. Preferably, the learning criterion defines a learning guideline, and learning may be performed according to the learning guideline.

Preferably, the learning criterion is a similarity difference between a recognition candidate and a correct answer candidate in a recognition method, a similarity degree of a correct answer candidate, a similarity degree of a recognition candidate, a recognition candidate category, a correct answer candidate category, and a candidate rank of a correct answer candidate. At least one of
May be included.

Preferably, the learning criterion may determine a recognition method and a learning method for learning a recognition dictionary. A pattern recognition dictionary learning apparatus according to the present invention (claim 11) includes: a pattern recognition unit that performs a pattern recognition process using a plurality of recognition methods; and a dictionary storage unit that stores a recognition dictionary for each of the plurality of recognition methods. And learning means for learning the recognition dictionary in accordance with a learning criterion determined based at least on the recognition result of each recognition method for the input pattern.

[0026] A pattern recognition dictionary creating apparatus according to the present invention (claim 11) includes a pattern recognition means for performing a pattern recognition process using a plurality of recognition methods, and a storage unit for storing the recognition dictionaries of the plurality of recognition methods. A dictionary storage unit, an initial dictionary creation unit that creates an initial recognition dictionary stored in the dictionary storage unit, a learning unit that learns the recognition dictionary, and at least based on a recognition result of each recognition method for an input pattern. Learning means for learning the recognition dictionary in accordance with a predetermined learning criterion.

The present invention (claim 13) causes a computer to execute a recognition process for an input pattern by a predetermined pattern recognition process using a plurality of recognition systems, and recognizes each recognition system executed in the pattern recognition process. A learning criterion for the recognition dictionary is determined based on at least the result,
A computer-readable recording medium that records a program for learning the recognition dictionary according to the determined learning criterion is summarized.

The present invention (claim 14) causes a computer to execute a recognition process for an input pattern by a predetermined pattern recognition process using a plurality of recognition systems, and recognizes each recognition system executed in the pattern recognition process. A learning criterion for the recognition dictionary is determined based on at least the result,
The gist of the present invention is a computer-readable recording medium storing a program for learning the recognition dictionary according to a recognition method and a learning method for learning the recognition dictionary determined by the learning criterion.

The present invention (claim 15) relates to a pattern recognition method using a plurality of recognition methods.
The input pattern is recognized using at least a recognition dictionary learned by the pattern recognition dictionary learning method according to any one of the above.

The present invention (claim 16) relates to a pattern recognition method using a plurality of recognition methods.
0, the input pattern is recognized using at least a recognition dictionary created by the pattern recognition dictionary creating method described in any one of the above items.

The present invention (claim 17) relates to a pattern recognition apparatus using a plurality of recognition methods.
A dictionary storage unit that stores a recognition dictionary learned by the pattern recognition dictionary learning method according to any one of the above, and a pattern recognition unit that recognizes an input pattern using at least the recognition dictionary. It is characterized by.

The present invention (claim 18) relates to a pattern recognition apparatus using a plurality of recognition methods.
0, a dictionary storage unit storing a recognition dictionary created by the pattern recognition dictionary creation method according to any one of 0
Pattern recognition means for recognizing the input pattern using at least the recognition dictionary.

According to the present invention, in the recognition dictionary learning in the pattern recognition processing in which a plurality of recognition methods are integrated and one recognition result is obtained, adverse effects relating to the learning are minimized, and appropriate advantages utilizing the advantages of each recognition method are obtained. Each recognition dictionary can be learned in various directions.

Further, with the recognition dictionary learned / created in this way, a highly accurate pattern recognition process utilizing the advantages of each recognition method can be realized. Note that the present invention relating to the apparatus is also realized as an invention relating to a method, and the present invention relating to a method is also realized as an invention relating to an apparatus.

The present invention relating to an apparatus or a method is provided for causing a computer to execute a procedure corresponding to the present invention (or for causing a computer to function as means corresponding to the present invention, or a computer corresponding to the present invention). The present invention is also realized as a computer-readable recording medium in which a program for realizing the function of performing the above is recorded.

[0036]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a basic configuration of a pattern recognition dictionary creating apparatus according to an embodiment of the present invention.

The present pattern recognition dictionary creation device includes a pattern input unit 2, an initial dictionary creation unit 4, a dictionary change unit 10, a recognition processing unit 8, and a dictionary data storage unit 6. The pattern input unit 2 is for inputting learning pattern data.

In the case of character recognition, the pattern input unit 2 optically reads and digitizes characters by using a photoelectric conversion element such as an image sensor, and in the case of voice recognition, converts voice into a microphone or the like. After conversion into an electric signal using an electroacoustic conversion element, the signal is A / D converted and digitized.

The dictionary data storage section 6 is for storing dictionary data. The recognition processing unit 8 recognizes the input pattern using the dictionary. Here, in the present embodiment, the recognition processing unit 8 has a plurality of different recognition methods and outputs a recognition result based on the plurality of recognition methods. Therefore, the dictionary data storage unit 6 stores a dictionary corresponding to each of the plurality of recognition methods.

There are various forms of recognition processing using a plurality of recognition methods, but there are roughly two types of processing, a parallel type and a serial type. The parallel type is to obtain the final recognition result by integrating the individual recognition results of each recognition method. As a simple example, the category in which the most recognition methods are the first candidates is used as the recognition result. Things.

In the serial type, the execution order of a plurality of recognition methods is determined in advance. For example, (1) each recognition method outputs one recognition result and its accept / reject, and performs recognition in accordance with the execution order. If the method rejects,
Move to the next recognition method, and use the first candidate of the first accepted recognition method as the recognition result. (2) The above (1)
In, only some top candidates of a certain recognition method,
There is a method including a recognition method as a recognition candidate of the next recognition method.

It should be noted that the serial type may be such that at least one stage is of the above-mentioned parallel type or serial type. As will be described later in detail, the dictionary change unit 10 learns the dictionary based on the learning pattern and the recognition result of the learning pattern learning, and changes the contents of the dictionary.
In addition, the recognition processing unit 8 performs learning according to the recognition process and the algorithm of each of the plurality of recognition methods.

The initial dictionary creating section 4 uses the initial dictionary creating algorithm to input the input initial dictionary creating pattern
Initial dictionary data (initial dictionary) is created based on the set. An initial dictionary is created for each recognition method. The initial dictionary creation pattern set is, for example, one in which several standard patterns are prepared for each category (for example, characters) registered in the dictionary. Although various methods are known as initial dictionary creation algorithms, the simplest method is to use the average of feature vectors obtained from each standard pattern as initial dictionary data.

If it is assumed that an external initial dictionary or a learned dictionary is stored in the dictionary data storage unit 6 for learning or re-learning, the initial dictionary creating unit 4
Need not be provided.

In the above configuration, when dictionary learning is performed, first, an initial dictionary is created from the pattern set and stored in the dictionary data storage unit 6, or an initial dictionary or a learned dictionary is read from outside. To store it in the dictionary data storage unit 6.

Next, a recognition process for a pattern to be learned is performed using the initial dictionary or the dictionary during learning, and a learning process is performed based on the result and the learning pattern to update the recognition dictionary. This dictionary learning / updating is repeated as needed to learn and improve the dictionary.

The conditions for terminating the repetition are as follows:
For example, a method in which learning for all learning patterns is defined as one round and a predetermined number of rounds are performed, and a predetermined evaluation parameter satisfies a certain criterion with a predetermined number of rounds as an upper limit (for example, correct reading rate> reference value /
Alternatively, various methods are conceivable, such as a method of performing until the rejection rate <reference value and / or misreading rate <reference value).

It goes without saying that the present pattern recognition dictionary creation device also functions as a pattern recognition dictionary learning device. FIG. 2 shows a basic configuration of a pattern recognition device according to one embodiment of the present invention.

The present pattern recognition apparatus includes a pattern input unit 1
2, a recognition processing unit 16, a dictionary data storage unit 14, and a recognition result output unit 18. These pattern input units 12,
The recognition processing unit 16 and the dictionary data storage unit 14
This is the same as the pattern input unit 2, the recognition processing unit 8, and the dictionary data storage unit 6 in FIG.

The recognition result output section 18 outputs a recognition result. In FIG. 2, first, a dictionary created or improved by the pattern recognition dictionary creating apparatus of FIG. 1 is stored in the dictionary data storage unit 14.

Then, the recognition processing section 16 recognizes the input pattern using the dictionary. The recognition result output unit 18 of FIG. 2 may be added to the configuration of FIG. 1 to form a pattern recognition device having a dictionary learning function or a dictionary creation function.

In the following, several embodiments relating to dictionary learning in a pattern recognition dictionary creating apparatus having the basic configuration as shown in FIG. 1 will be described. In each of the following embodiments, an input pattern is input from the pattern input unit 2 in FIG. 1, a recognition process is performed in the recognition processing unit 8 in FIG. 1, and a learning process is performed in the dictionary change unit 10 in FIG. The recognition dictionary of each recognition method is stored in the dictionary data storage unit 6 of FIG.

(First Embodiment) In the present embodiment, in a recognition process in which a plurality of recognition methods are integrated and used to output one recognition result, an input pattern is obtained from the recognition results of each recognition method in each recognition method. An example in which dictionary learning is performed according to a given standard will be described.

Here, as a parallel type recognition process using a plurality of recognition methods, when the first candidates obtained by the plurality of recognition methods are not all the same, that is, when there are a plurality of types, first candidates are selected from among the candidates. 1 by majority vote
It is assumed that a parallel recognition process of selecting one and determining this as a recognition result is used. However, in this case, it is assumed that it is not necessary for the majority recognition method to be the first candidate in order to accept the final recognition result. Further, when two or more candidate categories are the first candidates in the same number of recognition methods, the candidates are further narrowed down using other criteria.

In the present embodiment, when an incorrect answer occurs for a certain learning pattern in such recognition processing, learning is performed based on the following learning criteria and learning guidelines. It is assumed that the total number of recognition methods in this recognition processing is n, and the number of recognition methods for which the first candidate is the correct answer among the n recognition methods is m. In addition, the number of recognition methods in which the candidate category having the overall recognition result (here, an incorrect answer) is the first candidate is k.

At this time, among the mn (first-order candidates) recognition methods in which an incorrect answer is obtained, km−m + 1 (in the order of the difference in similarity between the first-order candidate and the correct-answer candidate) in ascending order. A recognition method is selected, and recognition dictionary learning is performed for each of the selected (k−m + 1) recognition methods.

In each of the recognition dictionary learning for the (k−m + 1) recognition methods, learning is performed so that the first candidate for the learning pattern is a correct answer. Note that km-m +
1 is the smallest value that can reliably reverse the majority decision under any condition and make the correct answer candidate the final recognition result.

In this way, only the recognition dictionary of a recognition method suitable (or easy to recognize) for recognizing the pattern is trained, and learning is not performed for the other recognition methods. It becomes possible to learn so that the correct answer candidate becomes the first candidate for the pattern. In other words, this makes it possible to avoid the adverse effects of learning and realize dictionary learning utilizing the good integration.

Further, when learning the dictionary of each recognition method, for example, the first result is obtained from the recognition result of the recognition method to be learned.
Processing such as changing the learning weighting coefficient α can be performed based on the difference in the similarity between the position candidate and the correct answer candidate. At this time,
It is desirable that a recognition method having a larger similarity difference learns with a larger learning coefficient. In this case, the recognition method that is difficult to recognize can be forcibly recognized halfway for the recognition method that is difficult to recognize, so that the recognition method that is easy to recognize among the recognition methods that cannot be correctly recognized has no adverse effect slowly. .

FIG. 3 shows an example of a learning algorithm in the present embodiment. First, a certain learning pattern is input (step S1), recognition processing is performed on the input learning pattern by a plurality of recognition methods, and a final recognition result (recognition category ) Is obtained (step S2).

If the recognition category is correct (step S3), the process proceeds to the next learning if there is a next pattern, and the process is completed if all the patterns have been learned (step S4).

On the other hand, if the recognition result obtained in step S2 is incorrect (step S3), the number m of recognition methods in which the first candidate is correct and the final recognition result are the first The number k of recognition methods that are rank candidates is obtained (steps S5 and S6).

Next, among the recognition methods in which the first candidate was incorrect, km-m + 1 pickup methods having a small difference in similarity between the first candidate and the correct candidate are picked up (step S).
7). This km−m + 1 recognition method is represented by S ₁ ,.
_{(km) +1} .

Next, dictionary learning of an input pattern is performed for each of the recognition methods S ₁ ,..., S _{(km) +1} (step S 8). This means that learning for one learning pattern has been performed.

Further, if there is the next pattern, the process proceeds to the next learning, and if the learning is completed for all the patterns, the process ends (step S9). As described above, the learning process of a plurality of rounds may be performed with the procedure of FIG. 3 as one round. Further, the processing may be performed with the number of rounds defined in advance as an upper limit until the correct reading rate or the like as the whole parallel type recognition processing satisfies a certain standard.

By the way, in the above description, among the nm recognition methods (in which the first candidate is the wrong answer), km−m + 1 are sequentially selected from the one having the smallest similarity difference between the first candidate and the correct candidate.
Although the number of recognition methods is selected, a slightly larger number or a slightly smaller number may be used as the number to be selected.

Further, in the case of the parallel recognition processing in which the majority of the recognition methods need to be the first candidate in order to accept the recognition result, the (first candidate) becomes incorrect. [N / 2] + in the order of the difference in similarity between the first candidate and the correct candidate among the nm recognition methods.
You may make it select 1-m recognition methods. Here, [x] represents an integer part of x.

Although the recognition processing based on majority decision has been described above, in the case of parallel recognition processing other than the recognition processing based on majority decision, the recognition method to be learned is determined in accordance with how to integrate a plurality of recognition methods. It is possible to choose.

In the above description, the recognition method to be learned is selected in order from the one with the smallest similarity difference between the first candidate and the correct answer candidate. A method may be selected. In the case where there is a means for obtaining the certainty factor of the correct answer candidate, the recognition method may be selected in order from the one with the highest certainty factor of the correct answer candidate.

(Second Embodiment) In the present embodiment, a case will be described in which the present invention is applied to an integrated system that performs recognition in a predetermined order based on the recognition rates of the respective recognition systems.

FIG. 4 shows an example of the algorithm of the integration system. In the integration method shown in FIG. 4, a process of integrating N recognition methods in series is considered.

When a pattern is input (step S1)
1), performing recognition processing using the first recognition method (step S12,
S13), the degree of certainty of the recognition result is obtained from the recognition result (the first candidate), and it is determined whether or not the obtained degree of certainty is equal to or higher than a predetermined reference (step S14).

If it is determined that the certainty satisfies the above condition (step S14), the recognition result is accepted and output as the whole recognition result (step S15). If it is determined that the certainty does not satisfy the above condition (step S14), a recognition process using the following second recognition method is performed (steps S16, S17, S13), and the certainty of the recognition result is obtained from the recognition result. Then, it is determined whether or not the certainty factor is equal to or higher than a predetermined reference (step S14).

If the above condition is satisfied here (step S14), the result is accepted and output as the overall recognition result (step S15). The output recognition result may be correct or incorrect.

If the above conditions are not satisfied, the same processing is repeated for the third and subsequent recognition methods. If the Nth recognition method as the final recognition method is not recognized (step S14, S16, S17),
The recognition result is rejected (step S18).

Here, there are various methods for obtaining the degree of certainty from the recognition result. For example, there is a method using a prior probability for each recognition category. In this method, the correct reading rate for each category is calculated in advance by simulation, and the correct reading rate is used as a certainty factor. In addition, there is another more reliable calculation method of the certainty factor, which is disclosed in detail in, for example, Japanese Patent Application Laid-Open No. 9-153113.

Now, the essential feature of the integration method as shown in FIG. 4, namely, (a) if the first candidate of the recognition method which is initially accepted is the correct answer, the result of the entire recognition processing becomes the correct answer. If the answer is incorrect, the result of the entire recognition process is incorrect (in other words, the results of all the recognition methods before the recognition method that was initially accepted are rejected (the certainty factor of the first candidate has reached the standard). And no recognition method after the first recognition method is executed), and (b) the result of the entire recognition process is rejected only when all the recognition methods are rejected. In consideration of the above, some basic dictionary learning can be considered as a dictionary learning for an incorrect answer pattern and a reject pattern in the present integrated system. (1) To learn to accept or reject a pattern that has been accepted by an intermediate recognition method and becomes incorrect, using the recognition method. (2) Learning to give a certain level of certainty higher than the standard in a recognition method that has correctly recognized a pattern that has been accepted and incorrectly accepted by an intermediate recognition method. (3) To learn a pattern that is finally rejected (that is, in all the recognition methods) so that the recognition method that has correctly recognized the pattern provides a certainty factor that is higher than a reference.

Various learning algorithms based on such a concept can be considered. FIG. 5 shows an example of a learning algorithm in the present embodiment. This algorithm performs recognition processing on the input pattern using the integrated method of FIG. 4, and if the result is incorrect or rejected, the learning is performed based on learning criteria and learning guidelines as exemplified later. Is what you do.

First, a certain learning pattern is inputted (step S21).
(Step S22). If the recognition result is a correct answer (step S23), if there is a next pattern, the process proceeds to the next learning, and if learning is completed for all the patterns, the process ends (step S24).

If the recognition result obtained in step S22 is a reject (steps S23 and S25), learning of a reject pattern is performed (step S26). If the recognition result is incorrect (steps S23 and S25), learning of an incorrect answer pattern is performed. Is performed (step S27).

This means that one learning for one learning pattern has been performed. Further, if there is the next pattern, the process proceeds to the next learning, and if the learning is completed for all the patterns, the process is terminated (step S28).

As described above, the learning process of a plurality of rounds may be performed with the procedure of FIG. 5 as one round. Further, the processing may be performed with the number of rounds defined in advance as an upper limit until the correct reading rate or the like as the whole serial type recognition processing satisfies a certain standard.

FIG. 6 shows an example of an algorithm of a dictionary learning process of a reject pattern. According to the integrated system of FIG. 4, the reject pattern is a pattern that has been rejected in all N types of recognition systems. Therefore,
In this processing procedure, learning is performed so that the correct answer candidate is accepted by any of the recognition methods.

First, among the N types of recognition methods, if the first candidate has a correct answer (step S3).
1) The dictionary of the recognition method having the highest certainty is learned (step S33). By learning the dictionary of the recognition method with the highest certainty in this way, the adverse effects of dictionary learning are reduced as much as possible, and dictionary learning that is strong against unknown patterns is realized.

Further, for the same purpose, if there is no recognition method in which the first candidate is the correct answer, dictionary learning is performed for the recognition method in which the correct candidate comes first (step S).
31, S32). If there are two or more such recognition methods, the dictionary of the recognition method with the smallest certainty factor of the first candidate is learned.

Of course, if there is no recognition method in which the first candidate is a correct answer, if the certainty of the correct candidate can be calculated, the certainty of the correct candidate is used instead of the above (for example, The recognition method having the highest certainty is selected), and a more stable and universal dictionary learning can be realized than the learning dictionary determination method based on the candidate ranking in FIG.

Next, FIG. 7 shows an example of a dictionary learning algorithm when the recognition result is incorrect. here,
It is assumed that it is accepted by the i-th method and becomes incorrect.

If i = 1, there is no other option, the dictionary of the first recognition method is learned, and the process is terminated (steps S41 and S42). Otherwise, learn two dictionaries. The two dictionaries are the i-th recognition method dictionary in which the accepted first-rank candidate is incorrect, and the most recognizable pattern in the first to (i-1) -th recognition methods. It is a dictionary of a recognition method that can be done. The former is for learning at least to reject (correctly correct) the pattern, and the latter is for learning the pattern correctly.

In order to perform learning so as to reject a certain pattern at least (correctly if it is successful), for example, there is a method in which the learning is performed by reducing the learning weight coefficient α. In this case, the erroneous reading pattern cannot be immediately converted to the correct answer, but the learning can be slowly performed without any adverse effect.

First, the dictionary of the i-th recognition method is learned (step S41). This is to reject or correct incorrect answers. Furthermore, the first to i-th
The dictionary of the recognition method having the highest certainty among the recognition methods for which the first candidate is the correct answer among the first recognition methods is learned (steps S42 to S44).

If there is no recognition method in which the first candidate is correct, the dictionary of the recognition method in which the correct candidate comes first is learned (steps S42, S43, S4).
5). At this time, when there are a plurality of recognition methods in which the correct candidate comes at the top, the dictionary of the recognition method with the smallest certainty of the first candidate is learned. The purpose of this learning is the same as the purpose of learning the reject pattern.

Of course, as described above, if there is no recognition method in which the first candidate is the correct answer, and if the certainty of the correct candidate can be calculated, the certainty of the correct candidate is used instead of the above. (For example, a recognition method having the highest certainty of the correct answer candidate is selected), and more stable and universal dictionary learning can be realized than the learning dictionary determination method based on the candidate order of FIG.

Note that as a result of learning an incorrect answer pattern, rejection may occur in all recognition methods. Even in such a case, for example, the result of the learning shown in FIG. The correct answer may be accepted by such a recognition method. Even if this is not the case, if the learning of FIG. 5 is executed again with the same pattern, the reject pattern can be learned.

Next, FIG. 8 shows another example of the dictionary learning algorithm when the recognition result is incorrect. In the learning algorithm of the incorrect pattern described above, learning of the input pattern is performed in two dictionaries (one dictionary only when i = 1), but learning of the incorrect pattern is performed by a method of learning only one dictionary. You can do it too. By doing so, it is possible to configure a recognition system capable of correctly recognizing a pattern, particularly when learning a highly ambiguous pattern or the like, while minimizing the adverse effect of the learning.

As shown in FIG. 8, first, when the answer is incorrect in the first recognition method, the dictionary of the first recognition method is learned, and the processing is terminated (step S51,
S52).

If not (steps S51, S5
3) Assuming that the incorrectly recognized recognition method is the i-th recognition method, there is a recognition method in which the correct answer is the first candidate among the first to (i-1) -th recognition methods. If it is (step S55), the dictionary of the recognition method having the highest certainty among those recognition methods is learned (step S5).
6). This is because, due to the nature of the integrated system of the present embodiment, if an incorrect answer is obtained in the i-th recognition system, the previous recognition system (the first to (i-1) -th recognition systems) will be rejected. It is due to becoming.

If there is a rejected recognition method while the correct answer is the first candidate, letting the recognition method learn the pattern and accepting it as the correct answer reduces the adverse effects of the learning. Considered optimal from a viewpoint.

If there is no recognition method in which the correct answer is the first candidate among the first to i-th recognition methods (step S53), learning is performed in the dictionary of the i-th recognition method. Then, the pattern is learned at least in a direction to reject the pattern (step S54), and the correct answer recognition is the (i + 1) th.
Leave it to the third and subsequent recognition methods.

(Third Embodiment) The present embodiment relates to dictionary learning in an integrated system of N kinds of recognition systems as in the second embodiment.

This embodiment differs from the second embodiment only in the learning algorithm in the case of an incorrect answer, and the other parts are the same. Therefore, only the difference will be described below. By the way, in the learning in the case of an incorrect answer in the second embodiment, the recognition processing is performed only by the first to ith recognition methods by the integrated method, and the (i + 1) th and subsequent recognition methods are considered. However, in the present embodiment, even when an incorrect answer pattern is obtained in the i-th recognition method, the i + 1-th and subsequent recognition methods are also considered.

FIG. 9 shows an example of an algorithm of a dictionary learning process of an incorrect answer pattern in the present embodiment. First,
It is determined whether or not the answer is incorrect in the first recognition method (step S61).

If the answer is incorrect in the first recognition method, the dictionary of the first recognition method is learned unconditionally,
The process ends (step S62). If not,
It is checked whether any of the first to (i-1) -th recognition methods has the correct answer as the first candidate (step S63). If there is such a recognition method, a recognition method having the highest certainty is found among them, and the dictionary is learned (steps S64 and S65).

Here, due to the nature of the integration method, a recognition method in which the correct answer is the first candidate is accepted before the recognition processing of a recognition method which is incorrect (here, the i-th recognition method) is started. If this is the case, the pattern will be the correct answer, so this algorithm is used.

[0104] In practice, the recognition method that is incorrect is that the correct answer is the first.
Since it is not a position candidate, it is considered that it is not suitable for recognizing the pattern. Therefore, an attempt is made to first recognize the most easily recognizable recognition method among the recognition methods in the hierarchy before such a recognition method.

However, even if the correct answer is not the first candidate, there is a case where the certainty of the correct answer is high and the effect of the learning is likely to appear. For such a case, the first
The guideline for selecting a dictionary to be learned from the recognition methods from the i-th to the i-th recognition method may be the one with the highest certainty of the correct answer candidate.

Next, if no correct answer is the first candidate among the first to (i-1) -th recognition methods (step S63), if i = N, the N-th recognition method is used. The system dictionary is learned, and the process is terminated (step S66,
S67).

Although not shown in the flowchart of FIG. 9, the dictionary of the recognition method with the highest certainty of the correct answer candidate among the N-1 types of recognition methods except the N-th recognition method is also used here. May also be learned together.

Next, if i <N in step S66, the recognition processing from the (i + 1) th to the Nth recognition method is performed (step S68), and among the (i + 1) th to the Nth recognition methods, It is checked whether there is a recognition method that has the correct answer as the first candidate, and if there is such a recognition method (step S69), the recognition method with the highest certainty among them (the j-th recognition method) ) Is learned (step S71), and the dictionary of the i-th recognition method and the dictionary of the j-th recognition method are learned (step S72).

[0109] This is because the i-th
Learning the dictionary of the second recognition method, at least rejecting the pattern, and searching for the most easily correct recognition method by the method of judging from the above-mentioned certainty factor so as to be able to obtain the correct answer by the recognition method of the subsequent rank, The obtained dictionary of the j-th recognition method is learned (so that correct candidates are accepted).

However, if the j-th recognition method selected here is a method that already accepts the pattern without learning, it does not extend to learning. Therefore, a check as to whether or not the recognition method accepts the correct answer candidate is performed prior to the learning (for example, in step S69), and if it is determined that the recognition method accepts the correct answer candidate, the i-th determination is made in step 72. The learning may be performed only for the dictionary of the recognition method.

Next, in step S69, if there is no recognition method in which the correct answer is the first candidate among the (i + 1) -th to N-th recognition methods, only the i-th recognition method which is incorrect is used. Learning is performed (step S70). Also in this case, similarly to the above, the dictionary of the recognition method in which the certainty of the correct answer candidate is the maximum among the first to (i-1) -th recognition methods may also be learned.

(Fourth Embodiment) The present embodiment provides a learning method that uses the same integration method as the second and third embodiments and further depends on the integration method. .

This embodiment is different from the second and third embodiments only in the learning algorithm in the case of an incorrect answer, and the other parts are the same. Therefore, only the difference will be described below.

FIG. 10 shows an example of an algorithm of a dictionary learning process for an incorrect answer pattern in the present embodiment. The processing procedure of the present embodiment is the same as the processing procedure of the third embodiment, except that the correct answer is not the first candidate among the first to (i-1) -th recognition methods (see FIG. 9). Only the processing content in the case where the incorrectly recognized recognition method is not the last recognition method in step S63) (that is, i <N) (step S63 in FIG. 9) is different.

That is, in the present embodiment, in the above case, the i-th recognition method (an incorrectly recognized recognition method)
The learning is performed with the policy of rejecting the recognition result in step (i), and the learning is performed with the policy of accepting the correct answer in accordance with the integration algorithm in the (i + 1) th and subsequent recognition processes.

Steps S81 to S87 in FIG.
Are the same as steps S61 to S67, respectively.
Hereinafter, only the case where the determination in step S86 of FIG. 10 is NO, that is, the processing of steps S88 to S97 will be described.

When the answer to the determination of step S86 is NO, the recognition dictionary of the i-th recognition method is first learned (step S88). Subsequently, j = i + 1 is set (step S8).
9) Recognition by the j-th recognition method is performed, and the recognition result is obtained (step S90).

If the result is reject (step S91), it is checked whether the first candidate is the correct answer (step S92). If the answer is correct, the dictionary of the j-th recognition method is learned (step S93). , And the process ends.

If the answer is not correct in step S92,
Since the first candidate in this recognition method is incorrect, j is updated by j = j + 1 without learning (step S9).
6), proceed to the next recognition method.

In step S91, if it is not reject (ie, if it is accept), it is checked whether the first candidate is correct (step S94).
If the answer is correct, it can be recognized by the recognition method without learning, and the process ends without learning the dictionary of the j-th recognition method.

However, in the case where the accepted first candidate is incorrect, the j-th candidate is rejected as a result of learning the dictionary of the i-th recognition method.
Since erroneous recognition occurs in the second recognition method, the dictionary of this recognition method is learned so as to be at least rejected (step S95). Then, j is updated by j = j + 1 (step S96), and the process proceeds to the next recognition method.

The above loop processing is performed in step S9.
After 3 or until the determination is YES in step S94 and the processing ends, or until j> N in step S97, the processing is repeatedly executed while j is incremented by one.

(Fifth Embodiment) In the present embodiment, as an example of serial recognition processing using a plurality of recognition methods, only some upper candidates of the previous recognition method are recognized by the next recognition method. A case in which the present invention is applied to an integrated system that is a target category will be described.

This integration method is effective in the case of recognizing a character set having many types of patterns to be recognized such as Chinese characters. FIG. 11 shows an example of the algorithm of the integration method.

In the integration system shown in FIG. 11, as an example, a case where two recognition systems are integrated in series is shown. When a pattern is input (step S10)
1) First, recognition candidates are narrowed down by a recognition process called a large classification (step S102). Next, recognition processing is performed only with the narrowed down recognition candidates, and final recognition called detailed identification is performed (step S103).

In this way, when recognizing a character set having more than 4000 categories such as Chinese characters, for example, a category which is completely similar to the input pattern is rejected by using a recognition method which does not require a large amount of calculation.
It is possible to narrow down the category to a category similar to the input pattern and perform a full-fledged (computation amount is greater than a large classification) recognition process, thereby greatly reducing the overall computation amount.

In such recognition processing, the cause of the incorrect final recognition result is roughly classified into the following two. (1) The recognition candidates narrowed down by the large classification do not include a correct answer. (2) The recognition candidates narrowed down by the large classification included correct answers, but the recognition in the detailed identification was incorrect.

It is necessary to change the learning method depending on the difference between the causes. FIG. 12 shows an example of a learning algorithm in the present embodiment. Here, the recognition dictionary to be learned is changed depending on whether or not the correct category is included in the candidate category according to the large classification.

First, a certain learning pattern is input (step S111), and a large classification recognition process, which is the first recognition method in FIG. 11, is performed on the input learning pattern (step S112).

Here, it is checked whether or not the correct category is included in the candidate categories obtained by the large classification recognition processing. If the correct category is not included (step S113), the large classification dictionary is learned (step S114).

When the correct category is included in the candidate categories obtained by the large classification recognition process (step S
113) Then, a detailed identification process, which is the second recognition method, is performed (step S116).

Here, it is checked whether the final recognition result obtained by the detailed identification processing is correct, and if it is not correct (step S11).
7), learn the detailed identification dictionary (step S118).
Of course, if the final recognition result obtained by the detailed identification processing is a correct answer (step S117), learning is unnecessary.

This means that the learning processing for one learning pattern has been performed. Further, if there is a next pattern in step S115 or step S119, the process proceeds to the next learning, and if the learning is completed for all the patterns, the process ends.

As described above, the procedure shown in FIG.
As a round, a plurality of rounds of learning processing may be performed. Further, the processing may be performed with the number of rounds defined in advance as an upper limit until the correct reading rate or the like as the whole serial type recognition processing satisfies a certain standard.

In the above, it is determined whether or not the correct category is included in the candidate categories obtained by the large classification recognition processing in step S113. When the candidate category is specified, the fact that the correct answer candidate does not fall within the N / 2 rank may be used as a condition for learning the large classification dictionary. In this way, it is possible to stably include a pattern that has just entered a candidate category into a candidate category, which is effective.

In the above-described integration method, the large classification may be divided into several stages. In such a case, the result is checked for each stage, and the result is omitted from the candidate category ( Alternatively, the dictionary may be learned at the stage when the dictionary has been omitted from the top candidates up to the specified order).

In the case where the detailed identification is obtained by integrating a plurality of recognition methods, learning can be performed by the methods shown in the first to fourth embodiments. The above procedures, means, and functions can be realized as software. Further, the present invention is embodied as a computer-readable recording medium on which a program for causing a computer to execute the above-described procedures (or for causing the computer to function as the above-described units, or for causing the computer to realize each of the above-described functions) is recorded. You can also. The present invention is not limited to the above-described embodiment, and can be implemented with various modifications within the technical scope.

[0138]

According to the present invention, in the recognition dictionary learning in the pattern recognition processing in which a plurality of recognition methods are integrated and one recognition result is obtained, the adverse effects on the learning are minimized, and the advantages of each recognition method are obtained. Each recognition dictionary can be learned in an appropriate direction to utilize. As a result, highly accurate pattern recognition processing utilizing the advantages of each recognition method can be realized.

[Brief description of the drawings]

FIG. 1 is a diagram showing a basic configuration of a pattern recognition dictionary creation device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a basic configuration of a pattern recognition device according to an embodiment of the present invention.

FIG. 3 is a flowchart illustrating an example of a learning algorithm according to the first embodiment of the present invention.

FIG. 4 is a flowchart illustrating an example of a serial recognition processing algorithm according to the second, third, and fourth embodiments of the present invention;

FIG. 5 is a flowchart illustrating an example of a learning algorithm according to the second, third, and fourth embodiments of the present invention;

FIG. 6 is a flowchart illustrating an example of an algorithm of a reject pattern learning process according to the second embodiment of the present invention;

FIG. 7 is an exemplary flowchart illustrating an example of an algorithm of a learning process of an incorrect answer pattern in the embodiment.

FIG. 8 is a flowchart showing another example of the algorithm of the incorrect answer pattern learning process in the embodiment.

FIG. 9 is a flowchart illustrating an example of an algorithm of a learning process of an incorrect answer pattern according to the third embodiment of the present invention.

FIG. 10 is a flowchart illustrating an example of an algorithm of a learning process of an incorrect answer pattern according to the fourth embodiment of the present invention.

FIG. 11 is a flowchart illustrating an example of a serial recognition processing algorithm according to a fifth embodiment of the present invention.

FIG. 12 is a flowchart illustrating an example of a learning algorithm according to the embodiment;

[Explanation of symbols]

 2, 12 ... pattern input unit 4 ... initial dictionary creation unit 10 ... dictionary change unit 8, 16 ... recognition processing unit 6, 14 ... dictionary data storage unit 18 ... recognition result output unit

Claims (18)

[Claims] 1. A pattern recognition dictionary learning method for learning a recognition dictionary in a pattern recognition process using a plurality of recognition methods, the method comprising: a learning criterion determined at least based on a recognition result of each recognition method for an input pattern; A pattern recognition dictionary learning method characterized by learning a recognition dictionary. 2. A pattern recognition dictionary learning method for learning a recognition dictionary in a pattern recognition process using a plurality of recognition methods, wherein the pattern recognition process is performed on an input pattern, the recognition result of the pattern recognition process and each recognition method A learning criterion of the recognition dictionary based on at least the recognition result of the above, and learning the recognition dictionary according to the determined learning criterion. 3. The pattern recognition dictionary learning method according to claim 1, wherein the learning criterion defines a learning guideline, and learning is performed according to the learning guideline. 4. The learning criterion includes a difference in similarity between a recognition candidate and a correct answer candidate in a recognition method, a similarity of a correct answer candidate, a similarity of a recognition candidate, a recognition candidate category, a correct answer candidate category, and a candidate rank of a correct answer candidate. 4. The method according to claim 1, wherein at least one of them is included.
The pattern recognition dictionary learning method described in the section. 5. The pattern recognition dictionary learning method according to claim 1, wherein the learning criterion defines a recognition method and a learning method for learning the recognition dictionary. 6. A pattern recognition dictionary creating method for creating a recognition dictionary in a pattern recognition process using a plurality of recognition methods by learning an initial recognition dictionary, wherein a recognition result of each recognition method for an input pattern is obtained. A method for creating a pattern recognition dictionary, characterized by learning the recognition dictionary according to at least a learning criterion determined based on the reference. 7. A pattern recognition dictionary creating method for creating a recognition dictionary in a pattern recognition process using a plurality of recognition methods by learning an initial recognition dictionary, comprising: executing a pattern recognition process for an input pattern; A learning criterion for the recognition dictionary is determined based at least on the recognition result of the pattern recognition process and the recognition result of each recognition method, and the recognition dictionary is learned according to the determined learning criterion. Method. 8. The pattern recognition dictionary creating method according to claim 6, wherein the learning criterion defines a learning guideline, and learning is performed according to the learning guideline. 9. The learning criterion includes a difference in similarity between a recognition candidate and a correct answer candidate in a recognition method, a similarity of a correct answer candidate, a similarity of a recognition candidate, a recognition candidate category, a correct answer candidate category, and a candidate rank of a correct answer candidate. 9. The method according to claim 6, wherein at least one of them is included.
Method for creating a pattern recognition dictionary described in the section. 10. The pattern recognition dictionary creation method according to claim 6, wherein the learning criterion defines a recognition method and a learning method for learning the recognition dictionary. 11. Pattern recognition means for performing pattern recognition processing using a plurality of recognition methods; dictionary storage means for storing recognition dictionaries of the plurality of recognition methods; and recognition results of each recognition method for an input pattern. Learning means for learning the recognition dictionary in accordance with a learning criterion determined at least based on the pattern recognition dictionary learning device. 12. A pattern recognition means for performing pattern recognition processing using a plurality of recognition methods, a dictionary storage means for storing recognition dictionaries of the plurality of recognition methods, and an initial state stored in the dictionary storage means. Initial dictionary creating means for creating a typical recognition dictionary; learning means for learning the recognition dictionary; and learning for learning the recognition dictionary according to a learning criterion determined based at least on the recognition result of each recognition method for an input pattern. Means for creating a pattern recognition dictionary. 13. A computer that executes a recognition process for an input pattern by a predetermined pattern recognition process using a plurality of recognition methods, and performs the recognition based on at least a recognition result of each recognition method executed in the pattern recognition process. A computer-readable recording medium storing a program for causing a dictionary learning standard to be determined and for learning the recognition dictionary according to the determined learning standard. 14. A computer which executes a recognition process for an input pattern by a predetermined pattern recognition process using a plurality of recognition methods, and performs the recognition based at least on the recognition result of each recognition method executed in the pattern recognition process. A computer-readable recording medium storing a dictionary learning standard, and a program for learning the recognition dictionary according to a recognition method and a learning method for learning the recognition dictionary defined by the learning standard. 15. A pattern recognition method using a plurality of recognition methods, wherein a pattern is input using at least a recognition dictionary learned by the pattern recognition dictionary learning method according to any one of claims 1 to 5. A pattern recognition method characterized by recognizing a pattern. 16. A pattern recognition method using a plurality of recognition methods, wherein a pattern is input using at least a recognition dictionary created by the pattern recognition dictionary creation method according to any one of claims 6 to 10. A pattern recognition method characterized by recognizing a pattern. 17. A pattern recognition device using a plurality of recognition methods, a dictionary storage means storing a recognition dictionary learned by the pattern recognition dictionary learning method according to any one of claims 1 to 5, A pattern recognition device comprising: pattern recognition means for recognizing an input pattern using at least the recognition dictionary. 18. A pattern recognition device using a plurality of recognition methods, a dictionary storage means for storing a recognition dictionary created by the pattern recognition dictionary creation method according to any one of claims 6 to 10, A pattern recognition device comprising: pattern recognition means for recognizing an input pattern using at least the recognition dictionary.