JPH02129782A - Character recognizing method - Google Patents

Abstract

PURPOSE:To complete the learning of a neural net in a short time by sharing the recognition of characters in the plural neural nets even when the number of the characters which go to be the subject of the recognition is increased. CONSTITUTION:A character recognizing device has a scanner part 1, a pattern candidate housing part 2, a pattern matching part 3, a deciding part 4 and a data housing part 5 and the deciding part 4 is equipped with the plural neural nets as respective deciding devices so that the decision of a recognized results to each character pattern can be respectively shared. The respective neural nets are corrected with directivity so that the correct character recognized result can be obtained in next decision by using a learning algorithm such as a back propagation algorithm, etc., with the input character pattern as the subject when the correct decision of the character recognition is not executed. Thus, even when the number of the characters to be the subject of the recognition is increased, the learning of the neural net can be completed in a short time.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a character recognition method used in, for example, an optical character reading device.

(Prior Art) In recent years, improvements in character recognition technology such as pattern matching have made it possible for optical character reading devices to read not only printed characters but also handwritten characters.

In such devices, the degree of similarity between an input character pattern and multiple character pattern candidates is calculated by pattern matching, and the character pattern candidate with the highest degree of similarity is determined as the final character recognition result. Character recognition is performed by outputting the following characters.

However, in such character recognition methods, due to changes in character pattern characteristics, the degree of similarity between the input character pattern and the correct character pattern candidate is often lower than the degree of similarity between the input character pattern and the correct character pattern candidate. This poses a problem in that erroneous character recognition may occur.

For this reason, conventional character recognition devices ensure a minimum recognition rate by, for example, making modifications or additions to the matching criteria for input character patterns and character pattern candidates in accordance with changes in character pattern characteristics. However, it is not necessarily a desirable method because it requires a great deal of effort and time to modify the matching criteria itself in this way.

Recently, a method has been proposed in which character recognition results are determined using a neural network using a learning algorithm.

In this method, even if the characteristics of the character pattern change, the learning algorithm corrects the neural network.
Stable character recognition can be performed without changing the matching standard itself during pattern matching.

However, with such character recognition methods, there is a concern that after the neural network is modified, character patterns that were previously correctly determined may no longer be correctly determined.

As a countermeasure in such a case, for example, register all the character patterns that were correctly determined before modification in a predetermined database, and then modify the neural network after monitoring these character patterns so that they are not affected. However, in any case, there is a limit to obtaining correct data for all character patterns using a single neural network, and as the number of characters to be recognized increases, the neural network There was a problem in that a large amount of time was spent on corrections.

(Problems to be Solved by the Invention) As mentioned above, in the conventional character recognition method using a neural network, there is a limit to the number of characters that can be successfully recognized. There was a problem in that it took a lot of time to modify the network.

The present invention is intended to solve the above-mentioned problems.Even if the number of characters to be recognized increases, the recognition of these characters can be divided among multiple neural networks, thereby reducing the time required for recognition. The aim is to provide a character recognition method that can complete the training of a neural network.

[Structure of the Invention] (Means for Solving the Problem) In order to achieve the above-mentioned object, the first invention provides a recognition result of an input character pattern based on each degree of similarity between the input character pattern and a plurality of character pattern candidates. This method uses multiple neural networks to determine the character pattern candidate with the highest degree of certainty, and the recognition results for each input character pattern are divided among the multiple neural networks. be.

In addition, in order to achieve the above-mentioned object, the second invention uses a plurality of neural networks to select a character pattern candidate with the highest degree of certainty as a recognition result of the input character pattern from each degree of similarity between the input character pattern and a plurality of character pattern candidates. In this method, a neural network that is directly used for determining the recognition result is determined according to the content of each degree of similarity.

Furthermore, in order to achieve the above-mentioned object, the third invention selects a plurality of character pattern candidates with the highest degree of certainty as a recognition result of the input character pattern from each degree of similarity between the input character pattern and the plurality of character pattern candidates. This is a method of making a judgment using a neural network, in which a plurality of neural networks divide the judgment of the recognition result for each character to be judged.

(Function) In other words, in the first invention, the determination of recognition results for each input character pattern is shared among multiple neural networks, so even if the number of characters to be recognized increases. , it is possible to complete learning of a neural network in a short time.

Further, in the second invention, since the neural network to be directly used for determining the recognition result is determined according to the contents of each degree of similarity, the learning of the neural network can also be completed in a short time.

Furthermore, in the third aspect of the invention, since the recognition results are determined by a plurality of neural networks for each character to be determined, short-term learning of the neural networks is possible.

(Example) Embodiments of the present invention will be described below based on the drawings.

FIG. 1 is a block diagram for explaining the configuration of a character recognition device that employs a character recognition method according to an embodiment of the present invention.

In the figure, 1 is a scanner unit that optically reads characters handwritten on a form to obtain a character pattern, and 2 is a registered character pattern candidate group to be matched with the character pattern read by the scanner unit 1. 3 is a pattern matching unit that performs predetermined pattern matching between the input character pattern from the scanner unit 1 and the character pattern candidates stored in the pattern candidate storage unit 2 to calculate the similarity of each. , 4 is a determination unit that determines the most probable character pattern candidate as a character recognition result from the combination of similarities for each character pattern candidate calculated in the pattern matching unit 3, and 5 is a character recognition determined in the determination unit 4. This is a data storage unit that stores results.

By the way, the above-mentioned determining section 4 is configured to include a plurality of neural networks as individual determining devices so that the recognition results for each character pattern can be divided and determined.

Then, each neural network uses a learning algorithm such as a pack propagation algorithm to target the input character pattern when a correct character recognition judgment is not made, so that correct character recognition results can be obtained in the next judgment. It has been revised with a certain direction in mind.

FIG. 2 shows the configuration of the neural network described above.

As shown in the figure, the network configuration of this neural network consists of an input layer consisting of a plurality of input nodes A into which similarity degrees for each pattern candidate calculated by pattern matching are input, and each input node A.
and an intermediate layer consisting of a plurality of intermediate nodes C connected in a network via arc B, and each intermediate node C and arc D connected in a network, and each final correct answer is determined based on the similarity for each pattern candidate. and an output layer consisting of an output node E that outputs a value indicating the confidence level.

Further, each arc BSD described above has a load value 1 indicating the strength of the connection between each node, and these load values are modified by the next network management unit.

The configuration of this network management section will be explained using FIG. 3.

In the same figure, 11 is an input node A and an intermediate node C
12 is a load value table that holds the load values set for each arc B and D, and 13 is an output from each input node A and intermediate node C. An integrator 14 performs m calculation by multiplying each data by a corresponding load value on the load value table 12, and an adder 14 adds up the integration results that are terminated at each intermediate node C or output node E. 15 is the output node E that outputs the maximum similarity in normal character recognition.
The determiner 16 determines the character pattern candidate corresponding to the character pattern candidate as a character recognition result, and the final integrated value for each character pattern candidate output from each output node E during learning should have the maximum degree of similarity. An evaluator 17 takes in and compares each data with the correct answer and evaluates whether or not they match.When the evaluator 16 determines that the data do not match, the evaluator 17 calculates the next time for the input character pattern. A weight value corrector 18 corrects each weight value on the weight value table 12 in a direction that will obtain the correct answer in character recognition processing; A learning database 19 stores the corresponding correct answer data as learning data, and a learning manager 19 manages the registered contents of the learning database 18.

FIG. 4 shows details of the entire determination section 4, which is configured with the plurality of neural networks described above.

In the figure, numeral 21 indicates a neural network constituting each individual judger, 22 a comprehensive judger that judges the final character recognition result from the judgment result of each neural net 21, and 23 a recognition target. 24 is a learning data division for dividing and registering the learning database registered in the whole learning database 23 into the learning database 18 in each neural network 21. A device 25 is a comprehensive evaluation device that compares the judgment results of the comprehensive judgment device 22 with correct data among the learning data extracted from the overall learning database 23, and evaluates whether or not they match.

Here, it is assumed that how the learning data registered in the overall learning database 23 is divided into each neural network 21 can be freely set in advance.

Next, the operation of this embodiment device will be explained.

First, several samples of character patterns to be recognized are collected, and these pattern images are sequentially read by the scanner section 1. Thereafter, the pattern matching unit 3 performs pattern matching between the input character pattern and each character pattern candidate, thereby outputting the degree of similarity for each character pattern candidate to the determination unit 4.

In normal character recognition, the determining unit 4 then determines the character pattern candidate that has obtained the maximum similarity from all the input similarity data, and uses this as the final character recognition result in the data storage unit 4. However, when learning is performed, the processing is performed as follows. This will be explained using the flowchart shown in FIG.

First, the combination of similarities for each character pattern candidate found in the pattern matching section 3 is registered in the learning database 18 as learning data along with correct data that is preferably finally obtained for the combination of similarities. be done.

Thereafter, all the learning data registered in the learning database 18 are divided by the learning database divider 24 (step a), and each divided learning data is assigned to the learning database 18 of each neural network 21 (step b). .

Execution of learning involves extracting one set of similarity combinations and the corresponding correct answer data from each learning database 18 (step c), and applying these to the input node output value table 11 of each neural network 21 and the evaluator. 16 (step d).
.

Now, when the similarity is set in the input node output value table 11, each similarity is set for arc B and arc D from input node A to output node E via intermediate node C. The weight values are integrated and further added (step e), and as a result, each value finally delivered to each output node E becomes a value (certainty) indicating the probability for each character pattern candidate.

Thereafter, the values for each character pattern candidate output from each output node E are output to the evaluator 16. Evaluator 16
compares and evaluates the data output from each output node E with the correct data (step f), and if the two do not match, activates the load value corrector 17 and changes each of the data in the load value table 12. Modify the load value (step g).

At the same time, the learning manager 19 stores the input similarity combination and correct answer data in the current learning in the learning database 18 again as learning data.

When the first learning for all character pattern samples is completed, the learning manager 19 again extracts each similarity combination and correct answer data one by one from the learning database 18 in order to start the second learning. , these are set in each neural net 21, respectively.

The above learning operation is repeated endlessly, and all learning is terminated when correct data can be obtained for all character pattern samples, or when the load value has been corrected a predetermined number of times (step h).

Thus, according to the character recognition device of this embodiment, the judgment for each input character pattern is performed using a plurality of neural networks 2.
1, the correction of the neural network 21 can be completed in a short time even if the number of character patterns to be recognized increases.

Next, other embodiments of the present invention will be described.

FIG. 6 shows the configuration of the determining section of a character recognition device according to yet another embodiment of the present invention, and the other parts are the same as those in the previously described embodiment.

As shown in the figure, if the determination unit in this character recognition device obtains at least one similarity of a certain value or more (for example, 0.9 or more) by pattern matching between the input character pattern and each character pattern candidate, A first neural network 3 that performs learning in a direction that allows correct data to be obtained in character recognition from the next time on only the character patterns of
1 and, conversely, a second neural network 32 that performs learning only on character patterns when no similarity greater than a certain value is obtained by pattern matching.

Further, 33 is a learning database, and this learning database 33 contains learning data for all character patterns to be recognized, that is, the respective degrees of similarity between the input character pattern and each character pattern candidate, and the corresponding correct answers. A pair with the data of is registered.

A rule 34 determines to which neural network 31 or 32 the learning data (similarity) registered in the learning database 33 is assigned.

Furthermore, 35 is an evaluator, and this evaluator 35 determines whether the judgment result output from either the first or second neural network 31 or 32 matches the correct answer data registered in the learning database 33. This is to evaluate whether or not it is true. That is, if the evaluator 35 evaluates that the two do not match, the learning proceeds by correcting the weight values in the neural networks 31 and 32 used in the determination.

Next, the operation of this embodiment device will be explained.

First, an example will be explained in which handwritten characters from "0" to "9" are recognized.

Here, when we calculated the degree of similarity with each character pattern candidate for the input character pattern whose correct answer is "4", we found: 0...... 0.1 1... 0.1 2・・・・・・・・・ 0.1 3・・・・・・・・・ 0.2 4・・・・・・・・・ 0.7 5・・・・・・・・・・・・ 0.8 6・・・・・・・・・0.3 7・・・・・・・・・ 0.1 8・・・・・・・・・0.1 9・・・・・・Assume that a result of 0.1 is obtained.

In this case, the maximum similarity is 0.8, and the character pattern candidate that obtained the maximum similarity is not “4” but ′
5'', the second neural network 32 determines that the degree of similarity to the character pattern candidate ``4'' is higher in certainty than the degree of similarity to the character pattern candidate 5''. It is necessary to modify the load value of each arc.

In this embodiment, the following learning is performed in each neural network 31 and 32 using the above-described similarities for each character pattern candidate and the corresponding correct answer data as learning data.

First, several samples of character patterns to be recognized are collected, and these are sequentially input to this character recognition device, and the degree of similarity between the input character pattern and each character pattern candidate is calculated.

Each calculated similarity is stored in the overall learning database 33 together with the corresponding correct answer data.

Thereafter, when similarities are output one by one from the learning database 33, the similarity data group is assigned and input to either the first or second neural network 31, 32 according to the rule 34. That is, if the similarity data group extracted from the learning database 33 includes at least one item with a similarity of 0.9 or more, the similarity data group is input to the first neural network 31.

On the other hand, if the similarity data group does not include a similarity of 0.9 or higher as 1, the similarity data group is input to the second neural network 32.

The result determined by either neural net 31 or 32 is compared with the correct data in the evaluator 35, and if the two do not match, the weight value of the neural network 31 or 32 that made the determination is used next time. Correct the direction in which the correct answer can be obtained in character recognition from .

Then, the first and second neural networks 31 and 32 are modified, and learning is terminated when correct data can be obtained for all samples.

Thus, according to the character recognition device of this embodiment, the similarity is determined depending on whether or not the similarity data group between the input character pattern and each character pattern candidate contains at least one similarity of a certain value or more. Since the neural networks 31 and 32 that actually judge the recognition results from the data group are determined, even if the number of character patterns to be recognized increases, the correction of the neural network can be completed in a short time. Moreover, it is possible to stably obtain correct recognition results for groups of character patterns with large feature changes.

Next, still another embodiment of the present invention will be described.

FIG. 7 shows a determining section in a character recognition device according to still another embodiment of the present invention, and the other components are the same as those in the embodiment shown in FIG.

As shown in the figure, the determination unit in this character recognition device includes a first neural network 41 that performs learning for the purpose of determining a correct character recognition result from the degree of similarity between an input character pattern and each character pattern candidate. , Despite repeated corrections to some extent in this first 221 runnet 41, or when the learning convergence is delayed, the character pattern for which the correct answer was not obtained (or the data with the maximum value is the correct answer) However, the second step is to perform learning for the purpose of determining correct character recognition results for character patterns whose values are not large enough, or character patterns whose difference from the second highest similarity is extremely small, etc. The main components are a neural network 42 and a third neural network that performs learning for the purpose of determining correct data from the determination results of these neural networks 41 and 42.

Further, 44 is a learning database in which learning data for all character patterns to be recognized is registered, and 45 is a third database.
This is an evaluator that compares the results determined by the neural network 43 with the correct data and determines whether they match.

Next, the learning operation of this embodiment device will be explained.

First, similarity data is extracted from the learning database 44 and input into the first neural network 41 one set at a time, and correct data is set in the evaluator 45. Thereby, in the first neural network 41, learning is performed in a direction such that correct data is determined for each degree of similarity.

Next, when this learning has been repeated to a certain extent, or when the convergence of learning has slowed down, learning data regarding character patterns for which no correct data has been obtained is input to the second neural network 42, learning in the neural network 42 is started.

During this time, the third neural network 43 is trained to determine which data is correct among the determination results of the first and second neural networks 41 and 42.

In this way, the first to third neural networks 41, 42 .
The learning in step 43 is continued, and all learning is ended when correct data is finally obtained for all character patterns.

Thus, according to the character recognition device of this embodiment, for a character pattern for which it is difficult to obtain a correct answer only by repeating learning in the first neural network 41, the second neural network 42 can obtain a correct answer. Since the learning is performed in a directional manner, the modification of the neural network can be completed in a short time as in the previously described embodiment.

In this embodiment, a third neural network 43 is further used to make a final character determination based on the determination results of the first and second neural networks 41 and 42. Rules may be used instead of the neural network 43.

Next, still another embodiment of the present invention will be described.

FIG. 8 is a block diagram showing a determining section in a character recognition device according to still another embodiment of the present invention.

As shown in the figure, this determination section mainly includes a plurality of neural networks 51 that perform determinations for each character pattern to be recognized. In other words, if there are 10 recognition targets from "0" to 9, then 1
0 neural nets 51 are required.

Learning in these neural nets 51 is carried out by first reading learning data registered in the overall learning database 52 via the data selector 53, and passing the respective learning data to the learning database of each neural net 51 (not shown).
After registering one set at a time, the load values are repeatedly corrected in a direction that allows correct data to be obtained for each neural network 51.

The learning is then terminated when correct data can be obtained from each neural network 51.

When determining characters, if even one of the determination results of each neural network 51 is determined to be correct, the determination device 5
4, the correct data is determined and output as the final recognition result.

Furthermore, if there are two or more character patterns that are determined to be correct, that is, if two or more neural networks 51 determine that the corresponding character patterns are correct, then the one that yields the largest value as the degree of correctness is selected. Judgment is made as a recognition result.

Furthermore, if all the neural networks are determined to be incorrect, it is determined that the determination is impossible.

Thus, according to the device of this embodiment, each neural network 51 divides the judgment of the recognition result for each input character pattern for each input character pattern, so that the character pattern to be recognized is Even if the number of characters increases, it is possible to complete the correction of the neural network in a short time, and furthermore, with this embodiment device, even if the characteristics of each character pattern change significantly, it is possible to perform stable and correct character recognition. can.

[Effects of the Invention] As explained above, according to the character recognition method of the present invention, even if the number of characters to be recognized increases, the recognition of these characters can be shared among multiple neural networks. By doing this, you can complete the study of 221 Urunet in a short time.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining the configuration of a character recognition device that employs a character recognition method according to an embodiment of the present invention, FIG. 2 is a block diagram for explaining the configuration of a neural network in FIG. 1, and FIG. 3 is a block diagram for explaining the configuration of the network management section that manages the neural network in FIG. 2, FIG. 4 is a block diagram showing the configuration of the determination section in FIG. 1, and FIG. A flowchart showing the flow of the learning operation in the judgment section, FIG. 6 is a block diagram showing the configuration of the judgment section in a character recognition device adopting a character recognition method according to another embodiment of the present invention, and FIG. FIG. 8 is a block diagram showing a determination section in a character recognition device according to another embodiment of the present invention. FIG. 3... Pattern matching section, 4... Judgment section, 2
1... Neural network, 22... Comprehensive judger, 2
3... Whole learning database, 24... Learning data divider, 25... Comprehensive evaluator.

Claims (3)

[Claims] (1) A method of determining a character pattern candidate with the highest confidence as a recognition result of the input character pattern from each degree of similarity between an input character pattern and a plurality of character pattern candidates, using a plurality of neural networks, the method comprising: A character recognition method characterized in that the plurality of neural networks share the determination of recognition results for each input character pattern. (2) A method for determining a character pattern candidate with the highest confidence as a recognition result of the input character pattern from each degree of similarity between the input character pattern and a plurality of character pattern candidates, using a plurality of neural networks, the method comprising: A character recognition method characterized in that the neural network to be directly used for determining the recognition result is determined according to the contents of each degree of similarity. (3) A method for determining a character pattern candidate with the highest degree of certainty as a recognition result of the input character pattern from each degree of similarity between the input character pattern and a plurality of character pattern candidates, using a plurality of neural networks, the method comprising: A character recognition method characterized in that the plurality of neural networks divide the judgment of the recognition result for each character to be judged.