JP3266441B2 - Character recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to character recognition.
The present invention relates to detailed identification of candidates obtained by matching a feature of an unknown character image with a dictionary. The present invention can be applied to a character recognition device, a drawing recognition device, and a document processing device, and can be applied in principle to general pattern recognition including voice recognition and image recognition.

[0002]

2. Description of the Related Art In a method of preliminarily defining one or more models for each character of a character type to be read in a dictionary and performing character recognition of an unknown character image by matching these models, a large classification is used. When a detailed classification process dedicated to each model set that can be matched at the same time is defined, and the detailed classification process is defined for the candidate model set obtained by the major classification, the final recognition result is determined by the detailed classification process. Has already been proposed by the present applicant (Japanese Patent Application No. 5-335104). This method
It has advantages such as improvement of recognition performance and expansion of character types to be read.

As for reject processing in character recognition, a reject method disclosed in Japanese Patent Application Laid-Open No. 2-129780 is known. In this method, the first character type code, its threshold value, and at least one second character type code and their respective thresholds are stored in a table in advance, and the first character code as a result of actually recognizing the unknown character is stored. The contents of the table are searched from the first character type code, and a reject threshold value is determined in consideration of the reject strength for each of the searched threshold values. The second character type of the recognition result is stored in the table. The determined predetermined rejection threshold value is selected according to which of the second-order character types corresponds to the second-order character type, and the selected threshold value is set as a second-order character type obtained as a recognition result of the unknown character. Is determined by comparing the similarity of the rejection.

In this reject method, the method of clearly distinguishing the first place and the second place from the top candidates of the recognition result of the unknown character is described in the contents of the publication (table search). Key is specified as the first character type code). So, for example,
The case where the second candidate is “Think” or “Yui”
The cases where the position candidates are “Yui” and “Thinking” are stored in the table independently as different cases. In other words, even for the same set of candidates, if the order of each candidate is different, they are stored and referenced as separate table elements.

[0005]

SUMMARY OF THE INVENTION The present invention is an extension of the character recognition method described in the specification of Japanese Patent Application No. 5-335104 and the accompanying drawings. It is an object of the present invention to provide specific means for achieving improvement, speeding up processing relating to detailed identification, and reducing the required memory amount.

[0006]

In order to achieve the above object, according to the present invention, there is provided a dictionary in which one or more models are defined for each character of a character type to be read, and a predetermined similar model set. A similar model set storage means in which information relating thereto is stored, and a candidate model set having a small degree of difference from the unknown character image is obtained by matching processing between the feature extracted from the unknown character image and the dictionary; A detailed identification process is performed, and in the detailed identification process, the similar model set storage means is examined using the candidate model set as a key, and a similar model set in which the combination of the candidate model set and the model matches is stored in the similar model set storage means. If it is not stored, the candidate model having the smallest difference in the candidate model set is used as the final recognition result, and the similar model set in which the combination of the candidate model set and the model matches the similar model set When stored in the Dell pair storing means it is stored in said similar model pair storing means,
Selecting one candidate model as the final recognition result from the candidate model set by using information on the similar model set in which the candidate model set and the model combination match;
Based on

Specifically, the features of the invention described in claim 1
The information on the similar model set stored in the similar model set storage means includes a coefficient for re-evaluation of the degree of difference. When the model set is stored in the similar model set storage means, the degree of difference between each candidate model in the candidate model set and the unknown character image is stored in the similar model set storage means. Re-evaluation is performed by multiplying by a coefficient for re-evaluation included in the information on the similar model set that matches the candidate model set and the model combination, and the candidate model having the minimum difference after re-evaluation is finalized. The result is a typical recognition result.

A feature of the present invention is that the information on the similar model set stored in the similar model set storage means does not include the final recognition result in the similar model set. Are included in the similar model set storage means when the similar model set matching the combination of the candidate model set and the model is stored in the similar model set storage means. When the reject flag in the information on the similar model set that matches the candidate model set and the model combination stored in the storage means is valid, the reject is to be the final recognition result.

A feature of the invention according to claim 3 is that, in the information on the similar model set stored in the similar model set storage means, the final recognition result should be a character category other than the similar model set. In the detailed identification process, a similar model set that matches the combination of the candidate model set and the model is stored in the similar model set storage means. Stored in the similar model set storage means.
When the replacement flag in the information on the similar model set that matches the candidate model set and the model combination is valid, the character category specified by the corresponding replacement character code part is to be the final recognition result. .

Further, a feature of the invention according to claim 4 is that each model in the dictionary has a flag for indicating that a similar model set including the model is stored in the similar model set storage means, If the flags of all candidate models in the candidate model set obtained by the matching process are valid, the detailed recognition process is performed to determine the final recognition result, and one candidate model in the candidate model set is determined. However, if the flag is invalid, the candidate model having the smallest difference from the unknown character image in the candidate model set is not used as the detailed identification process, and is set as the final recognition result.

[0011]

Was subjected to a recognition experiment 3000 type about handwritten Kanji by [action] is character recognition algorithm, the possibility of a combination of upper candidate models that can occur theoretically in match processing with the dictionary, a very large number and ₃₀₀₀ C ₂ In Although,
The combination of the top candidate models that actually occurred is 1.5
%, And the combination of the top candidate models in which misreading can occur is about 0.5% thereof.

Therefore, according to the present invention , misrecognition is likely to occur only by matching information with a dictionary by merely storing information on a model set that may actually be misread in the similar model set storage means. False recognition of characters is reduced, and the recognition rate is improved. Then, since only the information on a limited number of model sets needs to be stored in the similar model set storage means, the amount of memory for this is small. Further, since the candidate model sets are handled without distinguishing the candidate rank of each model, the similar model set storage means is compared with the case where the candidate ranks are distinguished as in the method of Japanese Patent Laid-Open No. 2-129780. Memory is halved.

Further, according to the first aspect of the present invention, in the detailed identification processing of the candidate model set including the same model combination as the similar model set, the degree of difference of each model is multiplied by a coefficient to reevaluate. , The increase in the amount of processing for detailed identification can be suppressed.

According to the second aspect of the present invention, when there is no correct answer in any of the models in the similar model set and there is not one character category that should be the correct answer, it can be correctly rejected.

According to the third aspect of the present invention, a correct recognition result can be obtained even when there is no correct answer in any of the similar model sets, but there is only one character category that should be the correct answer. . In other words, if the quality of the character image is extremely poor due to crushing or blurring, it may match another model that is impossible when the quality is high, but according to the invention of claim 3 , misunderstandings in such a case can be avoided. decrease.

According to the fourth aspect of the present invention, by merely adding a 1-bit flag to each model in the dictionary, unnecessary search processing of the similar model set storage means is reduced and the processing speed is improved.

[0017]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Some embodiments of the present invention will be described below. First, an example of a hardware configuration of a character recognition device according to each embodiment will be described with reference to the block diagram of FIG. In FIG. 1, reference numeral 1 denotes an input interface (I / F) for inputting image data of a document or a character from an input device such as an image scanner or a tablet. 2 is
A CPU that performs various processes and controls each unit of the apparatus according to a program. Reference numeral 3 denotes a ROM, in which a CPU 1
And fixed information such as a program group 3a for performing character recognition, a dictionary 3b, and a similar model set table 3c. One or more models (or standard patterns) are registered in the dictionary 3b for each character of the character type to be recognized. The similar model set table 3c is a similar model set storage unit for storing a predetermined similar model set which is likely to cause misidentification only by matching with a dictionary and information related thereto. Will be described later.

Reference numeral 4 denotes a RAM which is used as a storage area for image data input from the input interface 1 and used by the CPU 2 as a work area for various programs. Reference numeral 5 denotes an output interface (I / F) for outputting recognition result data and the like to an external device such as a host computer.

The contents of the processing and the structure of the similar model set table 3c will be described below for each embodiment.

<Embodiment 1> The flow of processing in this embodiment is as shown in the flowchart of FIG. The processing shown in this flowchart is
This is executed by the CPU 2 according to the program group 3a. Hereinafter, each step of the processing will be described in order.

First, image data is input via the input interface 1 and stored in a specific area of the RAM 4 (step S10). A character area is cut out from the input image in units of one character (step S11).
The following processing is performed on the image (character image) of each extracted character area. However, it is also possible to sequentially perform character extraction and recognition processing one character at a time, such as extracting one character and performing the following processing, and after that, extracting the next one character. A feature is extracted from the image data of one selected character area (step S12), and the extracted feature is matched with each character registered in the dictionary 3b (step S13). As a result of this matching process, up to n pairs of candidate models and dissimilarities are obtained. Usually, there is no problem if n = 2. In other words, there is generally no practical problem if the second candidate model is obtained.

The degree of difference is a value representing the degree of difference between the characteristics of the character image and the dictionary model, and the smaller the degree of difference, the better the match with the model. Therefore, the model with the smallest difference is the first candidate model, and the model with the next smallest difference is the second candidate model. The processing of steps S12 and S13 may be performed by an existing method.

Next, the obtained candidate model set (in this embodiment, a set of the first and second candidate models) is used as a key.
The similar model set table 3c is searched (step S1).
4) It is checked whether or not a similar model set matching the candidate model set and the model combination is registered in the similar model set table 3c (step S15).

In this search, the candidate models are handled without distinguishing their ranking (first and second). That is,
Regardless of the case where the first place is "I think" and the second place is "Yui", and the case where the first place is "Yui" and the second place is "Yui",
The similar model set table 3c is configured to refer to the same table element. With such a configuration, the size of the similar model set table 3c is reduced by half as compared with the case where the candidate rank of each model in the model set is distinguished.

The candidate model set is a similar model set table 3c
If not, the first candidate model is selected, its character category is determined (step S17), and it is output as the final recognition result (step S1).
8). If the candidate model set is registered, the final recognition result is selected by performing the identification process described later (step S16), and is output (step S18).

From step S12 to step S
The processing up to 18 is repeatedly executed in units of one character up to the last character area. When it is determined that the processing has been completed up to the last character area (step S19), the processing is completed.

Now, as can be inferred from the description of the processing after step S14, simply comparing the degree of difference with the character image of each candidate model obtained by the matching processing in step S13 (that is, the degree of difference If the smallest first-place candidate model is used as the final result immediately), there is a risk of erroneous recognition.
is there. In the present embodiment, a set of two models having such a possibility and information for determining a final recognition result relating to the two models are registered in the similar model set table 3c. Next, three specific examples of the similar model set table 3c will be described in relation to the identification processing in step S16. Here, the number of models in the similar model set is assumed to be 2 on the premise that the second candidate model is obtained.

Example 1: FIG. 4 shows a first example of the similar model set table 3c.
Ai is a table element of the i-th similar model set, which is composed of Mi1, Mi2, ki1, and ki2. Here, Mi1 and Mi2 are the names of the two models that constitute this similar model set. ki1 and ki2 are coefficients for re-evaluating the degree of difference regarding the models Mi1 and Mi2. The two candidate models obtained by the matching process in step S13 are Mi1, Mi2
(Regardless of which of Mi1 and Mi2 is the first and second place), the table element Ai is referred to. Then, in step S16, the character image to be recognized and each candidate model (here, Mi1, Mi2)
And the corresponding re-evaluation coefficients ki 1 and k 2
Multiplied by i2 and re-evaluated. If ki1 × Di1 ≦ ki2 × Di2, the character category of model Mi1 is used as the final result, and ki1 ×
If Di1> ki2 × Di2, the final category is the character category of model Mi2.

Example 2: FIG. 5 shows a second example of the similar model set table 3c.
Ai is a table element of the i-th similar model set, which is composed of Mi1, Mi2, a difference re-evaluation coefficient ki1, ki2, and a reject flag FREJi. Here, Mi1, Mi2
, Ki1 and ki2 are the names of the two models constituting the similar model set and the respective models Mi as in the example shown in FIG.
This is a coefficient for re-evaluating the degree of difference regarding 1, Mi2. The reject flag FREJi is a flag that is set to “valid” when a model of a character category as a final recognition result is not included in the similar model set.

When two candidate models obtained by the matching process in step S13 are Mi1, Mi2 (Mi1, Mi2
Mi2 is the first or second place),
The table element Ai will be referred to. In step S16, the reject flag FREJi is checked. If the flag is "invalid", the character image to be recognized and each of the candidate models (here, Mi1 and Mi2) are recognized in the same manner as in the first example. The differences Di1, Di2 are re-evaluated by multiplying them by the corresponding re-evaluation coefficients ki1, ki2, and ki1 × Di1 ≦
If ki2 × Di2, the final category is the character category of model Mi1, and if ki1 × Di1> ki2 × Di2, the final category is the character category of model Mi2. However, when the reject flag FREJi is "valid", the final result is rejected and a reject code is output.

Example 3: FIG. 6 shows a third example of the similar model set table 3c.
Ai is the table element of the i-th similar model set, which is Mi1, Mi2, the difference re-evaluation coefficients ki1, ki2, the replacement flag FCNGi, and the replacement character code COD.
It consists of Ei part. Here, Mi1, Mi2 and ki1, ki2 are:
As in the example shown in FIG. 4, the names of the two models constituting the similar model set and the coefficients for re-evaluating the degree of difference regarding each of the models Mi1 and Mi2. Replacement flag FC
NGi is a flag that is set to “valid” when the final recognition result should be a character category other than the similar model set. When this flag is “valid”, the content of the replacement character code part CODEi is finally set. To provide a character code as a recognition result.

When the two candidate models obtained by the matching process in step S13 are Mi1, Mi2 (Mi1, Mi2
Mi2 is the first or second place),
The table element Ai will be referred to. In step S16, the replacement flag FCNGi is checked. If the replacement flag FCNGi is reset (invalid), the character image to be recognized and each candidate model (here, Mi1, Di2 from Mi2)
1, Di2 are multiplied by the corresponding re-evaluation coefficients ki1 and ki2, and re-evaluated. When ki1 × Di1 ≦ ki2 × Di2, the model M
Let the character category of i1 be the final result, ki1 × Di1> ki2 ×
In the case of Di2, the final category is the character category of the model Mi2. However, when the replacement flag FCNGi is set (valid), the content (character code) of the replacement character code CODEi is output as the final result.

<Embodiment 2> The flow of processing in this embodiment is as shown in the flowchart of FIG. The processing shown in this flowchart is
This is executed by the CPU 2 according to the program group 3a.

Here, differences between the present embodiment and the first embodiment will be described. In the first embodiment, the similar model set table 3c is always searched for the model set obtained as a candidate by the matching process with the dictionary. However, the model set may not be registered in the similar model set table 3c. Not a few. In order to reduce the time consumption for the useless search of the unregistered model set, in the present embodiment, each model in the dictionary is provided with a 1-bit search flag, and a similar model including the model is provided. If the set is registered in the similar model set table 3c, this search flag is set to "valid", and if not, the search flag is set to "invalid". Then, for the model obtained as a candidate by the dictionary matching process, it is determined whether or not to search the similar table set table 3c by checking the search flag.

Hereinafter, each processing step will be described with reference to the flowchart of FIG. Step S20 of inputting image data, Step S21 of cutting out a character region, Step S22 of extracting a feature of an image of a character region, Step S2 of a matching process between the extracted feature and each model of the dictionary 3b.
3 corresponds to steps S10, S11, S1 of the first embodiment.
2, S13. Further, similar to the first embodiment, the similar model set table 3c is stored in FIG. 4, FIG.
The configuration shown in FIG.

Following the matching process in step S23, the search flag of each model obtained as a candidate is checked, and the first
It is determined whether both the search flag for the candidate model and the search flag for the second candidate model are “valid” (steps S24 and S25). If the search flag of any one of the candidate models is “invalid”, the character category of the first candidate model is selected as the final result (step S29), and is output (step S30). That is, the detailed identification process using the similar model set table 3c is not performed.

When the search flags of the first and second candidate models are both "valid", a similar model set composed of the same model combination as the candidate model set is registered in the similar model set table 3c. There is a possibility that the similar model set table 3
The search for c is performed, and it is determined whether or not it has been registered (steps S26 and S27). In this search, the candidate models are handled without distinguishing their ranking (first and second), as in the case of the first embodiment.

When a similar model set including the same model combination as the candidate model set is not registered in the similar model set table 3c, the character category of the first candidate model is selected as the final result (step S29). Is output (step S30). When a similar model set composed of the same model combination as the candidate model set is registered, an identification process related to the contents of the corresponding table element of the similar model set table 3c is performed to select the final recognition result (step S28), and outputs it (step S30). Although the processing content of step S28 depends on the configuration of the similar model set table 3c, it is the same as that of step S16 of the first embodiment, and thus the description thereof will not be repeated here.

From step S22 to step S22
The processing up to 30 is repeatedly executed in units of one character up to the last character area, and when it is determined that the processing has been completed up to the last character area (step S31), the processing is completed.

[0040]

According to the present invention, misrecognition is likely to occur only in the matching process with the dictionary, by merely storing the information on the model set in which misreading may actually occur in the similar model set storage means. It is possible to prevent erroneous recognition of characters and improve the recognition rate. Since it is only necessary to store information on a limited number of model sets in the similar model set storage means, the amount of memory for this is small. Further, since the candidate model sets are handled without distinguishing the candidate rank of each model, the memory capacity of the similar model set storage means can be reduced by half as compared with the case where the candidate ranks are distinguished.

According to the first aspect of the present invention, the detailed identification process for the candidate model set including the same model combination as the similar model set is performed simply by multiplying the difference of each model by a coefficient and re-evaluating. Since only simple processing is performed, it is possible to suppress an increase in the processing amount for detailed identification.

According to the second aspect of the present invention, when there is no correct answer in any of the models in the similar model set and there is not one character category to be the correct answer, the rejection can be correctly performed.

According to the third aspect of the present invention, when there is no correct answer in any of the models in the similar model set, but there is only one character category to be the correct answer, for example, the character image is crushed or blurred. And the like, the quality is remarkably poor, so that a correct recognition result can be obtained in a case where the model matches another model which is impossible when the quality is high.

According to the fourth aspect of the present invention, unnecessary search processing of the similar model set storage means can be reduced by simply adding a 1-bit flag to each model in the dictionary, and the processing speed can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an example of a hardware configuration of each embodiment of the present invention.

FIG. 2 is a flowchart illustrating a flow of processing according to the first exemplary embodiment of the present invention.

FIG. 3 is a flowchart illustrating a flow of a process according to a second exemplary embodiment of the present invention.

FIG. 4 is a diagram showing a first example of a similar model set table.

FIG. 5 is a diagram showing a second example of a similar model set table.

FIG. 6 is a diagram illustrating a third example of a similar model set table.

[Explanation of symbols]

 1 input interface 2 CPU 3 ROM 3a program group 3b dictionary 3c similar model set table

Claims (4)

(57) [Claims] 1. A dictionary in which one or more models are defined for each character of a character type to be read, and a predetermined similar model set and a coefficient for re-evaluating a difference degree regarding the similar model set are defined as: Means for storing candidate model sets having a small degree of difference from the unknown character image by matching processing between the feature extracted from the unknown character image and the dictionary, and then performing detailed identification processing. In the detailed identification process, the similar model set storage unit is checked using the candidate model set as a key, and a similar model set that matches the combination of the candidate model set and the model is not stored in the similar model set storage unit. In this case, the candidate model with the smallest difference in the candidate model set is used as the final recognition result, and a similar model set that matches the combination of the candidate model set and the model is stored in the similar model set storage means. If it is 憶 is,
The degree of difference between each candidate model in the candidate model set and the unknown character image is stored in the similar model set storage means.
The candidate model set and the model combination are re-evaluated by multiplying by a coefficient for re-evaluation in the similar model set in which the model combination matches, and the candidate model having the minimum difference after the re-evaluation is finally recognized. A character recognition method characterized by being a result. 2. The character recognition method according to claim 1, wherein
The final recognition in the similar model set is stored in the similar model set storage means.
Reject to indicate that no knowledge result is included
The candidate model set and the model combination in the detailed identification process.
Is stored in the similar model set storage means.
If it is remembered, it is stored in the similar model set storage means.
That the candidate model set and the model combination match
When the reject flag in the similar model group is valid
Makes the reject the final recognition result,
When the flag is invalid, each candidate in the candidate model set
The degree of difference between the model and the unknown character image is calculated using the similar model set.
The candidate model set and the model stored in the storage means.
For re-evaluation of the degree of dissimilarity in similar model sets with matching combinations
Re-evaluated by multiplying by the coefficient of
The candidate model with the smallest difference later is used as the final recognition result.
A character recognition method characterized by: 3. The character recognition method according to claim 1, wherein
A similar model pair storing hand stage, the similar mode the final recognition result
To indicate what should be a character category other than Dell
Of the candidate model set and the model in the detailed identification process.
Is stored in the similar model set storage means.
If it is remembered, it is stored in the similar model set storage means.
That the candidate model set and the model combination match
When the replacement flag in the similar model set is valid,
Character specified by the corresponding replacement character code part
The category is the final recognition result, and the replacement flag is
When invalid, each candidate model in the candidate model set is
The similarity model set storage unit stores the degree of difference from the unknown character image.
The combination of the candidate model set and the model stored in
Coefficients for re-evaluating dissimilarity in matching similar model sets
Re-evaluated by multiplication and the difference after the re-evaluation
Make the candidate model with the lowest degree the final recognition result
Character recognition method to be characterized. 4. The character recognition method according to claim 1, wherein each model in said dictionary is for indicating that a similar model set including the model is stored in said similar model set storage means. A flag, and if the flags of all the candidate models in the candidate model set obtained by the matching process are valid, perform the detailed identification process to determine a final recognition result; Even for one candidate model, if the flag is invalid, the candidate model having the smallest difference from the unknown character image in the candidate model set without performing the detailed identification processing is used as the final recognition result. A character recognition method characterized in that: