JP3265701B2 - Pattern recognition device using multi-determiner - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition apparatus for automatically recognizing patterns of characters, images, sounds, and the like in a computer.

Conventionally, various devices for recognizing patterns such as characters, images, and sounds have been known. However, at present, there is no pattern recognition device having a recognition rate of 100%.

[0003] However, each pattern recognition device has its own unique features, and has areas of specialty and weakness. Therefore, using these various pattern recognition devices, the respective recognition results at the time of learning are recorded as a history, and when an unlearned pattern is recognized, comprehensive judgment is made using the above learning results. This makes it possible to perform pattern recognition utilizing the features of the respective pattern recognition devices, and is expected to improve the recognition rate.

In particular, there is a history of pattern recognition,
By recording what kind of pattern and what kind of recognition is performed, it is possible to partially improve the pattern recognition apparatus.

As a method for comprehensively judging a plurality of pattern recognition results, for example, there is a recognition method using a neural network as disclosed in Japanese Patent Application Laid-Open No. 4-31955 "Character recognition device". In the network recognition method, each learning pattern is
I do not know what effect it has on the recognition result. For example, a pattern that is not very effective in improving the recognition rate is also learned, and there is a problem that a logical explanation of which sample is good remains, and this is not appropriate as a component of the pattern recognition apparatus.

[0006]

2. Description of the Related Art FIG. 5 is a diagram for explaining a conventional pattern recognition device. In the past pattern recognition, various recognition methods and recognition algorithms have been proposed, but each has a different characteristic, and there are target patterns that are good at and not good at.

Therefore, a pattern recognition device (for example, OC
R), it is necessary to select an optimal recognition method (algorithm) according to the recognition target. However, in reality, the input pattern to be recognized undergoes various deformations due to the influence of the outside world. On the other hand, once configured, the recognition device cannot change its recognition method, so that the recognition accuracy is reduced.

In order to improve this, as shown in FIG. 5, a plurality of recognition devices (determiners) 21, 22,... Having different recognition methods are conventionally configured in parallel, and the recognition results are comprehensively determined. Then, a configuration that outputs one determination has been proposed.

In this case, as a method of comprehensively determining the recognition results, 1) the recognition results of the determiners 21, 22,... 2) Use the most recognition result (majority decision). 3) A neural network is used as each of the determiners 21, 22,... To make a comprehensive determination of the determination results in each neural network. Etc. have been proposed.

[0010]

In the above prior art,
There are the following problems. For example, in the case of the configuration that adopts the one with the highest certainty factor in the above item 1), since the other judgment results are ignored and only the specific judgment method is fully trusted, the final result Reliability decreases.

Further, in the case of the configuration adopting the majority decision method of the item 2),
When almost half the result is obtained, the majority decision is forcibly performed, and the reliability of the final decision is reduced. In addition, as described in the item 3), a method using a neural network as the judgment unit 2 is disclosed in Japanese Patent Laid-Open No. Hei 4-31955.
In Japanese Patent Laid-Open Publication “Character Recognition Apparatus”, it is possible to comprehensively judge individual judgment results and proceed with learning so as to derive a final result. However, as described above, In addition, since it is not possible to logically explain whether or not it is affecting the final result, if the recognition rate is poor, it is necessary to add a new learning pattern and learn again, which is time-consuming There is a problem.

As described above, in the conventional pattern recognition apparatus, the characteristics of each recognition method are ignored, and comprehensive judgment by simple logic, for example, selection of the maximum certainty of the recognition result or majority logic is performed. Therefore, although many recognition methods are parallelized, the reliability of the final determination result is not so improved.

In view of the above-mentioned conventional disadvantages, the present invention is an improvement over the above-described conventional method for comprehensive judgment, which not only improves the accuracy of recognizing the final judgment result than before, but also provides a system configuration for various input patterns. It is an object of the present invention to provide a pattern recognition device capable of learning flexibly without changing.

[0014]

1 and 2 are diagrams for explaining the principle of the present invention. FIG. 1 shows a basic configuration diagram, and FIG. 2 shows a flow of a basic operation. The above problem is solved by a pattern recognition device configured as follows.

(1) A device for automatically recognizing a pattern such as a character, an image, and a voice, and a controller that manages the input pattern and transmits the pattern to a plurality of determiners 21, 22,. 1 and multiple classifiers implementing different recognition methods 2
When a total determines the overall judgment unit 3 a recognition result output from them, the comprehensive determination dictionary that stores learning data relating comprehensive judgment (hereinafter, sometimes simply referred to as a dictionary) 4 made, the input learning pattern In the category corresponding to
At the time of learning, a set of the recognition result output from each of the determiners 21, 22, and the confidence of the recognition result is used as a recognition result pattern.
(Hereinafter sometimes referred to as an output pattern)
And registers the serial comprehensive determination dictionary 4, each decision unit 21 at the end of learning, the comprehensive determination dictionary seeking accuracy rate per ~
Registered in book 4, in the recognition of {processing steps 100 to 103 see FIG. 2 (a)} unlearned pattern, the total determination unit 3
There total determination 21,22, and recognition result of the ~, previously, the learning
Therefore, the recognition result list in the comprehensive judgment dictionary 4 created
Compared with the turn , if there is a matching recognition result pattern , the learning category registered therein is output as the final recognition result (see processing steps 200 to 204 in FIG. 2B).
In the comprehensive judgment unit 3 and the comprehensive judgment dictionary 4,
When comparing with the recognition result pattern , each judgment unit 21,22,
The recognition result is compared for each, and one value corresponding to the comparison result is calculated, for example, a positive number is determined when the values match, and a negative number is calculated when the values do not match, and matching is performed depending on whether the total value is positive or negative. Is configured to determine the degree of

(2) At the time of the above learning, each of the discriminators 21,2
2. From the correct answer rates for each, a determiner with a rate lower than the reference correct answer rate set in advance is determined as an unsuitable determiner for pattern recognition.
It is configured to be excluded as 1, or 22,.

(3) At the time of the learning, the number of appearances of the recognition result pattern is counted for each learning pattern, and a certainty factor proportional to the number is registered in the comprehensive judgment dictionary 4. (4) In the recognition of the unlearned pattern , the comprehensive judgment
When there is no completely matching recognition result pattern in the dictionary 4, the recognition result of the decision unit 21, or 22, having the highest accuracy rate in the comprehensive judgment dictionary 4 is set as the final result. .

(5) In recognizing the unlearned pattern , a recognition result pattern that completely matches in the comprehensive judgment dictionary 4
If there is no turn , in the comprehensive judgment dictionary 4, the judgment units 21, 22, ... having a high correct answer rate are taken to the top L, and it is judged whether or not the recognition result pattern matches within that range.
The category corresponding to the matching recognition result pattern is configured to be the final result.

(6) In the recognition of the unlearned pattern, a certainty factor corresponding to the number of appearances of the matching result pattern in the comprehensive judgment dictionary 4 is output as a certainty factor of the final result. Constitute.

[0020]

FIG. 1 shows the basic configuration of the present invention. Here, the controller 1 has a function of managing input patterns to be recognized, such as images and sounds, and distributing the input patterns to a plurality of determiners 21, 22,. The determiner group 2 includes the plurality of determiners 21, 22,
And executes different recognition methods. Here, it is assumed that 2 includes n decision units REC (1) to REC (n). The comprehensive determiner 3 comprehensively determines the recognition results of REC (1) to REC (n). At the time of learning, the comprehensive judgment unit 3 creates the comprehensive judgment dictionary 4 independently. Comprehensive judgment dictionary
Numeral 4 stores the recognition results of REC (1) to REC (n) for the learning pattern, refers to each recognition result stored in the comprehensive decision unit 3 when recognizing the unlearned pattern, and sets the recognition result based on a predetermined logical condition. And make a comprehensive decision.

FIG. 2 is a flowchart showing the basic operation of the present invention. FIG. 2 (a) shows that the learning pattern is stored in each of the discriminators {REC (1) to
Shows the operation of storing the result recognized in step (1) in the comprehensive judgment dictionary. In the comprehensive judgment dictionary, each recognition unit 確信 REC (1) to｝ stores in what pattern the recognition result and the certainty factor (%) are output one by one, and each judgment unit ｛REC (1) to Outputs the total correct answer rate (%) in｝.

A determiner whose overall correct rate (%) is lower than a preset standard correct rate is excluded as an unsuitable determiner for pattern recognition. In addition, the number of appearances of the recognition result pattern is counted for each learning pattern, and a certainty factor proportional to the number is registered in the dictionary.

FIG. 2B shows the operation of comprehensive judgment when recognizing an unlearned pattern. The recognition results of each of the determiners ｛REC (1) to ｛are compared with the created comprehensive judgment dictionary in advance by the operation shown in FIG. If there is, the final judgment result registered in the dictionary is output as the final result. If not, it is output as undeterminable. Or,
The recognition result of the determiner having the highest accuracy rate in the dictionary is set as the final result. Alternatively, a determiner having a high correct answer rate is employed even in the upper L-level, and it is determined whether or not a pattern matches within the range, and a category corresponding to the matched pattern is determined as a final result.

At this time, the above registered in the dictionary
The certainty corresponding to the number of appearances of the recognition result pattern is output as the certainty of the final result. With the above configuration and procedure, the present invention can be realized.

Therefore, the recognition result at the time of learning is recorded sequentially, and the judgment can be made based on the recorded result, so that the recognition accuracy can be improved as compared with the conventional majority logic. In addition, even if the input pattern fluctuates significantly, the comprehensive judgment dictionary changes in accordance with the change in the recognition results of the multiple judgment devices.
Can be adapted without changing the overall configuration.

Also, in an environment where a plurality of decision units are not in the same hardware but are distributed to each server connected to a network, the recognition results of the decision units of the respective servers can be transmitted through the network. Can be collected and recognized based on the result registered in the comprehensive judgment dictionary, so that the basic configuration of the present invention does not need to be changed. And the like.

[0027]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. FIG. 1 and FIG. 2 are explanatory views of the principle of the present invention.
3 and 4 are views showing an embodiment of the present invention,
FIG. 3 (a) shows an example of the structure of the comprehensive judgment dictionary, and FIG.
Fig. 4 shows an example in which the output patterns are sorted in descending order of the number of appearances in the comprehensive judgment dictionary. Fig. 4 shows the comparison of the recognition results for each of the determiners when comparing the patterns in the dictionary with the comprehensive determiner. Then, an example is shown in which the values of the match and the mismatch as the comparison results are calculated, and the degree of matching is determined based on the total value.

In the present invention, characters, images, a automatic device for recognizing a pattern such as voice, is input
A controller 1 that manages the pattern and transmits the same to a plurality of determiners 21, 22,..., A plurality of determiner groups 2 that implement different recognition methods, and a comprehensive determination that comprehensively determines the recognition results output from them. a vessel 3 consists overall judgment dictionary 4 stores learning data relating the overall judgment, the inputted learning
At the time of learning the category corresponding to the pattern, the recognition results output from each of the determiners 21, 22, ... and the recognition result
The combination with the reliability is used as the recognition result pattern for the comprehensive judgment.
Registered in the dictionary 4, and at the end of the learning , each of the decision units 21,22,.
The correct answer rate for each is determined and registered in the dictionary 4, and in the recognition of the unlearned pattern, the comprehensive determiner 3
And the recognition result of ~, in advance, said that was created by the learning
The recognition result patterns in the comprehensive judgment dictionary 4 are compared, and if a matching recognition result pattern exists, the learning category registered therein is output as the final recognition result, and the above-described comprehensive judgment device 3 and the pattern in the dictionary 4 are output. When comparing with each other, the recognition results are compared for each of the determiners 20, 21, and so on, and one value according to the comparison result, for example, a positive number when matching, a negative value when not matching Calculate the numbers and determine their respective decision units 21,22, ...
Means for judging the degree of matching based on whether the total value of the collected values is negative or positive is a means necessary for embodying the present invention. Note that the same reference numerals indicate the same object throughout the drawings.

Hereinafter, referring to FIGS. 1 and 2, FIG.
The configuration and operation of the pattern recognition device of the present invention will be described with reference to FIG. 1. The categories to be prerequisite learning targets are C1 to Cm (m categories). Ni (i = i =
1 to n). For now, m = 4 and Ni = 5. At this time, the total number of learning patterns is N1 + N2 +.
+ Nm. The j-th learning pattern of category k is represented as Pkj.

Each of the decision units is REC (1) 21 to REC (n) 2n. As a specific example, n = 4 is set. For example, taking character recognition as an example, REC (1) 21 is a “pattern matching method”, REC (2) 22 is a “fold line approximation method”,
C (3) 23 realizes a known relaxation method such as the “relaxation method”, and REC (4) 24 realizes a known recognition method such as the “composite similarity method”. It is assumed that the necessary dictionaries are also included in each of the determiners 21, 22,.

Each of the determiners 21, 22,... Outputs the recognized category name (ci) and the certainty factor (ri) at that time (i is the number of the determiner). 2. Learning 2.1 Basic operation of learning The operation at the time of learning will be described. When the learning pattern Pkj is input, the controller 1 outputs the data to each decision unit R
It is transmitted to EC (1) 21 to REC (4) 24. Each judgment unit 21 ~
24 performs pattern recognition using a built-in dictionary (fixed) in advance, and outputs the result as (ci, ri). In the comprehensive judgment device 3, each (ci, ri) is
Create the table shown in (a). This is repeated for all learning patterns. Although the certainty factor ri is not shown in FIG. 3A, the certainty factor ri of each determiner i is added to each determination result ci and output. {Refer to processing steps 100 and 101 in FIG. 2} When the learning is completed, the category Ci as shown in FIG.
A table of 5 rows and 4 columns is completed for each (“A”, “B”, “C”, “D”). Here, if there is a discriminator that outputs the correct answer (Ci) for all the learning patterns (5 patterns), this discriminator {REC (2) 22 in the example shown in FIG.
% Correct answer ability. Similarly, the total correct answer rate is obtained for the other determiners, and is entered in the table. ｛Figure 2
2.2 Sorting of learning result {Example corresponding to claim 3} Next, the same REC (1) 21 to REC (4) 24 recognition results are totaled for each output learning pattern. In order to check what kind of output pattern is large, sorting is performed on the number of appearances.

In FIG. 3A, within the same category ("A" in the illustrated example), the same output pattern (c1, c2, c3,
Those having c4) are collected and arranged in descending order of the accumulated value. FIG. 3B shows the rearranged table. This ends the learning.

2.3 Exclusion of Inappropriate Judgment Unit by Learning {Embodiment Corresponding to Claim 2} Here, in the table of FIG. 3A, a judgment unit whose overall correct answer rate is lower than a preset reference value If there is a {decision unit REC (4) 24 when, for example, the reference value is set to 30%}, it is excluded as a decision unit inappropriate for pattern recognition. Therefore, among REC (1) 21 to REC (4) 24, REC (1)
(1) Only 21 to REC (3) 23 are used as practical judgment units.
{See processing step 103 in FIG. 2} 3. Recognition of Unlearned Pattern 3.1 Basic Operation of Recognition {Example Corresponding to Claim 1} Next, recognition of an unlearned pattern will be described. The input pattern (unknown) is sent from the controller 1 to each of the determiners 21, 22,
The recognition result is output from each of the decision units 21, 22, and. The comprehensive decision unit 3 searches the comprehensive decision dictionary 4 stored in the form shown in FIGS. 3A and 3B and outputs the same output pattern (c1, c).
2, c3, c4) are checked.で は In this example,
Assume that REC (4) 24 mentioned above is not excluded.あ れ ば If a completely matched pattern exists in the comprehensive judgment dictionary 4, the corresponding category is output as the final recognition result. {See processing steps 200, 201, 202, and 204 in FIG. 2} If there is no matching pattern, the result is output as unrecognizable. Alternatively, countermeasures such as 3.3 and 3.4 described later are performed. << Refer to processing step 203 in FIG. 2 >> 3.2 Matching Method for Table Search << Embodiment Corresponding to Claim 1 >> In the comparison between the comprehensive judgment dictionary 4 and the output pattern, there is a case where they do not completely match as described above. .

In such a case, weighting is performed as described below, and the determination is made based on the total value. FIG. 4 schematically shows the operation at this time. First, a certain pattern in the comprehensive judgment dictionary 4 is (T1, T2, T3, T4), the certainty factor of each judgment unit at that time is (R1, R2, R3, R4), and the recognition result for the unlearned pattern "X" To (x1, x2, x
3, x4), and the confidence at the time of recognition is (r1, r2, r3, r
4). If the recognition result of the decision unit REC (i) 2i,
If the results in the comprehensive judgment dictionary 4 match (xi = Ti), Qi ≡ Min (ri, Ri) If they do not match, a Qi such that Qi≡−Min (ri, Ri) is obtained. Q = Q1 + Q2 + Q3 as total value
Find + Q4. If this Q is a negative value, no matching is performed. Among the Qs having positive values, Pkj with the maximum total value is searched over all the learning patterns, and the category Ck is output as the final result.

3.3 Case of No Matching (1)
<< Embodiment Corresponding to Claim 4 >> In 3.1, the recognition result pattern matching with the comprehensive judgment dictionary 4 is described.
If there is no turn , do the following: Total in dictionary
Find the decision unit REC (i) with the highest correct answer rate , and find that decision unit REC
The result xi of (i) is output as the final result. In addition, the certainty factor for the recognition pattern is maximum (max (r1, r2, r
(3, r4)) The search for the determiner i and the result xi as the final result are the same as described in the prior art.

3.4 When there is no matching (2)
<< Embodiment Corresponding to Claim 5 >> This is a second countermeasure in the case where there is no match. Comprehensive judgment dictionary 4
Are arranged in descending order of the total correct answer rate , and the same L as the above-mentioned item 3.1 is matched. For example, match up to the top three
If there is an output pattern , the category is output as the final result.

3.5 Calculation of Final Certainty Eventually, one category name and one certainty factor must be output as a recognition result. The category name is obtained by the above method. The certainty must be obtained by integrating the results of a plurality of determiners. First, the number of appearances corresponding to the table of FIG. 3B is searched for the matched dictionary pattern. For example, in FIG. 3B, when the output pattern matches the leftmost output pattern , the number of appearances is three. The value 3/5 = 60% normalized by the number of times of learning “5” is output as the certainty factor of the final result. As described above, the present invention can be realized.

[0038]

As described above, according to the pattern recognition apparatus of the present invention, the following effects can be obtained. (1) The recognition accuracy can be improved by using a plurality of determiners. (2) By storing recognition results at the time of learning, recognition accuracy can be improved as compared with the conventional majority logic or the like. (3) Even if the input pattern fluctuates significantly, the overall judgment dictionary also changes in accordance with the change in the recognition result of the multiple determiners, so that it can be applied without changing the overall configuration. (4) Even in an environment where a plurality of determiners are not in the same hardware but are distributed to each server connected to the network, the recognition results and the like from each determiner can be collected via the network. By creating a comprehensive judgment dictionary,
Since the overall judgment can be made according to the present invention, there is no need to change the basic configuration of the present invention.

[Brief description of the drawings]

FIG. 1 illustrates the principle of the present invention (part 1).

FIG. 2 is a diagram illustrating the principle of the present invention (part 2).

FIG. 3 shows an embodiment of the present invention (part 1).

FIG. 4 shows an embodiment of the present invention (part 2).

FIG. 5 is a diagram illustrating a conventional pattern recognition device.

[Explanation of symbols]

1 Controller 2 Judgment group 21,22, ~ 2n Each judge {REC (1) ~ REC (n)}, Recognition device 3 Comprehensive decision unit 4 Comprehensive decision dictionary, dictionary 100 ~ 103,200 ~ 204 Processing steps

Continuation of the front page (56) References JP-A-2-21383 (JP, A) JP-A-2-129782 (JP, A) JP-A-3-220686 (JP, A) JP-A-60-108981 (JP) , A) JP-A-62-219091 (JP, A) JP-A-63-263588 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06T ⁷ /00-7/60 G06K 9/03 G06K 9/62 G06N 3/00 G10L 15/28

Claims (6)

(57) [Claims] 1. An apparatus for automatically recognizing a pattern such as a character, an image, and a voice , comprising:
And management, storage and a controller that transmits a plurality of determination devices, a plurality of determination unit group that implements the different recognition methods, and comprehensive determining overall judgment unit a recognition result output from them, a learning data relating to the total determination It consists of a comprehensive judgment dictionary, which is used when learning the category corresponding to the input learning pattern.
And the recognition result output from each decision unit and the recognition result
The combination with the certainty factor of
And registers the determination dictionary, seeking accuracy rate for each decision unit at the end of the learning register the success rate at the comprehensive determination dictionary in the recognition of the unlearned pattern, the comprehensive determination unit is fully determiner Before the recognition result and the previously created by the learning
Comparing the recognition result pattern in the comprehensive judgment dictionary ,
If there is a matching recognition result pattern , the learning category registered therein is output as the final recognition result, and the comprehensive determinator outputs the recognition result pattern in the comprehensive determination dictionary.
When performing a comparison with a turn, the recognition results are compared for each of the decision units, and a predetermined
A pattern recognition apparatus comprising: calculating a value; and determining a degree of matching based on a total value that is a sum of the predetermined values . 2. A pre-SL during learning, the correct answer rate for each decider, advance, and characterized in that exclude the rate decision unit below the reference accuracy rate set as inappropriate determiner pattern recognition The pattern recognition device according to claim 1. 3. A pre-SL during learning, count the number of occurrences of the recognition result patterns for each learning pattern, and registers the confidence in proportion to the number of times before Symbol comprehensive determination dictionary, claim 1 A pattern recognition device according to claim 1. 4. A recognition before Symbol unlearned pattern, before if <br/> Symbol no recognition result pattern that exactly matches the total determination dictionary, with maximum accuracy rate before Symbol the comprehensive determination dictionary The pattern recognition apparatus according to claim 1, wherein the recognition result of the determiner is used as a final result. In recognition of 5. Before SL unlearned pattern, if the previous <br/> Symbol no recognition result pattern that exactly matches the comprehensive judgment dictionary, before Symbol the comprehensive determination dictionary, the accuracy rate is high determination vessel was taken up to the upper L-position, that the range in the recognition result <br/> pattern it is determined whether match, the category corresponding to the matching recognition result <br/> pattern to the final result The pattern recognition device according to claim 1, wherein 6. The recognition of prior SL unlearned pattern, the pre <br/> Symbol comprehensive determination dictionary matched recognition result pattern certainty factor according to the number of occurrences of, to output as the reliability of the final result The pattern recognition device according to claim 1, wherein