JPH08202827A - Character recognition method - Google Patents

Abstract

PURPOSE: To reduce recognition errors owing to the separation or synthesis of characters and to reduce processing quantity for correcting the evaluation value of a recognized result. CONSTITUTION: In a step 204a, a final result is selected in accordance with rules on the combination of the separated characters and the synthesized characters, which is mentioned in a rule table. When it is not selected here, a correction value α for improving the priority of the synthesized characters is added in a step 204b. In a step 204c, a correction value β for improving the priority of the synthesized characters is added when a character element size is half and the character type of the separated characters is not appropriate for en. In a step 204d, a correction value γ for improving the priority of the separated characters is added when the separated characters are alphanumerics. In a step 204e, the recognized result of either the separated characters or the synthesized characters is selected as a final result by large or small of the evaluation value average value of the corrected separated characters and the evaluation value of the synthesized characters.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to character recognition, and in particular,
Should two adjacent phonemes be recognized as separate characters,
Alternatively, the present invention relates to a technique for determining whether to recognize as one integrated character.

[0002]

2. Description of the Related Art In an optical character reader (OCR), a pattern of one character is separated and cut into two partial patterns (character elements), and each character element is erroneously recognized as a separate character. is there. For example, the kanji for "ka"
Is recognized as two characters, "power" and "mouth". On the contrary, a combination of different characters may be recognized as one character. In order to avoid such an error, it is necessary to devise whether to recognize two adjacent character elements as separate characters that are separated or as one integrated character. The following two methods are known as examples of the technique.

One is the method described in Japanese Patent Laid-Open No. 3-291779. In this method, after recognizing a pattern (separated character) consisting of individual character elements and an integrated pattern (integrated character) of character elements as characters respectively, one of them is regarded as a character by evaluating the separated character or the integrated character. Although the selection is made, the blank widths before and after the noted character element are compared, and the evaluation value and the selection method used for the selection are switched depending on which blank width is wider.

The other is the method described in JP-A-4-37970. In this method, two separated characters and their integrated characters are individually recognized, and the similarity between the separated character recognition result and the dictionary is multiplied by a correction value of less than 1 to correct the similarity. The average value (representative value) of the similarity of two separated characters is compared with the similarity of the dictionary of integrated character recognition results, and when the former value is larger, the separated character is selected, and when it is not, the integrated character is selected. Select.

In the above-mentioned two methods, when a character is separated into a bias and a whit and is misrecognized (in which "ka" is separated into "a" and "hi", "add" is "force" and "mouth"). It is intended to correctly handle cases such as "separating into") and cases in which two characters are combined into a single character and misrecognized (such as the case where "15" is combined in order). In short, the similarity (or the evaluation value of the certainty) when each character element (a block of black pixels detected by taking a vertical projection of a line image or tracking the connection of black pixels) is individually recognized as a character. , Similarity (or evaluation value of certainty) when a combination of two character elements is recognized as a character, and after performing some processing on them, similarity (or evaluation value)
The higher character (separated character or integrated character) is selected as the correct one. However, each of the above methods has the following problems to be solved.

In the method disclosed in Japanese Patent Laid-Open No. 3-291779, the selection method is switched according to the result of comparison of the blank width before and after the character element of interest. Therefore, even if the image of the problematic character is exactly the same, The selection method is switched depending on the relationship with the character to be used. However, the character spacing is not always constant in a sentence, and in some documents, the character spacing is adjusted line by line, and in this case, the relationship between the space and the adjacent character changes line by line. Therefore, the same character that appears multiple times in a sentence is separated or integrated depending on the position where it appears. This instability is one of the problems to be solved by this conventional method.

On the other hand, in the method disclosed in Japanese Patent Laid-Open No. 4-37970, the similarity is corrected by multiplying the similarity of the separated characters by a correction value of less than 1. To realize this correction processing on a computer. You have to do decimal multiplication,
There is a problem that the calculation amount is heavy and the processing takes time.

Another problem is that the width of correction of the similarity is proportional to the value of the original similarity. Therefore, when the value of the similarity becomes smaller, the correction becomes insufficient and the discrimination ability deteriorates. In addition, there is a tendency that when the value of the degree of similarity is large, the number of cases of overcorrection increases.

Further, in this method, the similarity of the separated characters is multiplied by a correction value of less than 1 to uniformly select the separated characters from the integrated characters. Recognition rate will improve. That is, in the case of an integrated character that can be separated into two characters on the left and right, such as "add", "ka", "ki" (correct is correct), or "caro", "ihi", "kotomi" (correct is correct) In this way, both the recognition result as a separated character and the similarity of the recognition result as an integrated character have a high value when the characters are satisfied even if they are integrated or separated. Considering the frequency of occurrence of characters in, the ratio of cases where the integrated character is correct is high in such cases (for example, "Karo", "Ihi", "Kotomi", "add", "ka", "ki" are correct answers. There are many cases that become). Therefore, the priority of the integrated character improves the recognition rate as a whole. However, even if it is a rare case, it is easy to cause an error in combining original separation characters such as "Karo" and "Ihi", and especially when the correction of similarity is excessive, the separation characters are always combined. , There is a problem.

[0010]

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to improve the accuracy of distinguishing an original separated character from an original integrated character in an optical character recognition device and the like, and
It is to reduce the processing amount for the identification and speed up the processing.

[0011]

According to the character recognition method of the invention described in each of claims 1 to 5, character recognition as two separated characters for two adjacent character elements extracted from an image, and Performing character recognition as one integrated character, obtaining a recognition result of the separated character and an evaluation value representing the certainty thereof, and a recognition result of the integrated character and an evaluation value representing the certainty, and the separating A step of performing correction processing on the evaluation value of the character and the evaluation value of the integrated character, and according to the magnitude relationship between the average value of the evaluation values of the separated characters after the correction processing and the evaluation value of the integrated character, Selecting the separated character recognition result or the integrated character recognition result as a final recognition result.

In the character recognition method of the present invention, the correction value is added to the evaluation value of the separated character or its average value or the evaluation value of the integrated character in the correction processing step. .

In the character recognition method according to the second aspect of the present invention, in the correction processing step, the size of the character element is equal to or smaller than the half-width character equivalent size, and the recognition result of the separated character is unsuitable as the half-width character. When the character belongs to a specific character type, a correction value is added to the evaluation value of the separated character or its average value or the evaluation value of the integrated character in order to increase the selected priority of the recognition result of the integrated character. To be done.

In the character recognition method according to the third aspect of the present invention, in the correction processing step, the size of the character element is equal to or smaller than the half-size character equivalent size, and the recognition result of the separated character is inappropriate as the half-size character. Belongs to a specific character type,
Further, when the recognition result of the separated character is not a previously designated character, the evaluation value of the separated character or its average value or the integrated value is increased in order to increase the priority of selection of the recognition result of the integrated character. The correction value is added to the evaluation value of the selected character.

In the character recognition method of the present invention as defined in claim 4, in the correction processing step, when the recognition results of the two separated characters are both letters or numbers,
A correction value is added to the evaluation value of the integrated character or the evaluation value of the separated character or its average value in order to increase the selected priority of the recognition result of the separated character.

According to the character recognition method of the present invention, in the correction processing step, the recognition results of the two separated characters are both alphabetic characters or numerals, and
When the integrated character recognition result is not a Roman numeral,
A correction value is added to the evaluation value of the integrated character or the evaluation value of the separated character or its average value in order to increase the selected priority of the recognition result of the separated character.

The feature of the invention described in claim 6 is that in the character recognition method of the invention described in each of claims 1 to 5,
Prior to the correction processing step, it is checked whether the combination of the recognition result as the separated character and the recognition result as the integrated character matches with one of a plurality of predetermined specific character combinations. A step of determining a final recognition result according to a predetermined rule regarding the matched specific character combination, and the step and the step are executed only when the final recognition result is not determined by the step. is there.

[0018]

When the average value of the respective evaluation values of the two separated characters (two adjacent character elements) is set on the horizontal axis X and the evaluation value of the integrated character is set on the vertical axis Y, its distribution is roughly plotted.
For example, a distribution chart as shown in FIG. 1 can be obtained. The evaluation value is a value obtained by dividing the distance from the dictionary by the number of contours. By dividing by the number of contours in this way, a lower evaluation value is obtained as the degree of difference with the dictionary is smaller, and a higher evaluation value is obtained as the degree of difference with the dictionary is larger, regardless of the size of the character.

In FIG. 1, 1 is a set in which separated characters are correct answers, and 2 is a set in which integrated characters are correct answers. In the case of such a distribution, the boundary line 3 of Y = X is drawn, and it is checked which of the upper and lower sides of the boundary line 3 the magnitude relationship between the evaluation value of the integrated character and the average value of the evaluation values of the separated characters is. This makes it possible to determine whether the separated character is the correct answer or the integrated character is the correct answer.

However, the actual distribution is such that the set of separated characters that are correct and the set of integrated characters that are correct overlap each other, and the boundary that best separates each set is above or below the boundary of Y = X. There is a possibility that there is a deviation, and in order to correctly select the separated character / integrated character by comparing the average value of the separated character evaluation values with the integrated character evaluation value, it is necessary to correct such a boundary line deviation. is there. According to the first aspect of the invention, the deviation of the boundary line is corrected by adding a predetermined appropriate correction value to the evaluation value of the separated character or its average value or the evaluation value of the integrated character.

For example, in the example of the distribution shown in FIG. 2, the boundary line 4 that best separates the sets 1 and 2 may be obtained by moving the boundary line 3 of Y = X by α in the Y direction. According to the invention of claim 1, prior to the comparison between the average value of the separated character evaluation values and the integrated character evaluation value, a positive correction value α is added to the separated character evaluation value or the average value thereof. As a result Y
The boundary line 4 obtained by moving the boundary line 3 of = X upward by α selects the separated character / integrated character. However, if the correction value is a negative value, the correction value is added to the integrated character evaluation value.

Then, the correction of such a boundary line (the same applies to the correction in the inventions of claims 2 to 5),
Since it is performed by adding the correction values, there is an advantage that the required calculation processing amount is much reduced and the processing can be speeded up as compared with the multiplication of the correction value less than 1 (decimal multiplication). Further, stable correction can be performed with a constant correction width regardless of whether the evaluation value or the average value of the evaluation values is small or large.

With respect to the overlapping portions of the sets 1 and 2 as illustrated in the distribution chart of FIG. 2, even if the boundary line correction as described above is performed, the separation character / integrated character selection accuracy is further improved. Can't expect improvement. The invention described in each of claims 2 to 5 is intended to achieve the improvement of the selection accuracy for the overlapping portion by the correction processing of the evaluation value in consideration of individual characters.

First, in the invention according to claim 2, the character type of the recognition result as a separated character and the character element size are referenced, and the character element size is equal to or less than the half-width character equivalent size and the character type of the recognition result is If the character type is inappropriate as half-width characters (Hiragana, Katakana, Kanji), the correction value is added to give priority to the recognition result of the integrated character (for example, a positive correction value is evaluated as a separated character. Value or its average). By such a correction, it is possible to reduce erroneous recognition of characters such as "addition", "ization", "writing", "contract", etc.

However, according to the invention of claim 2, the number of character types in which the integrated character is prioritized is small,
There is also a character in which the separated character is the correct answer. For example, in hiragana and katakana, the width of the characters is small.
Characters such as "mo""ri""i""i""to""tsu""ri". However, this varies slightly depending on the target font, so it must be adjusted according to the font. In order to prevent the integrated character from being mistakenly selected for such a specific character, in the invention according to claim 3, the size of the character element is equal to or less than a half-width character equivalent size,
And even if the recognition result of the separated character belongs to an unsuitable specific character type as a half-width character, only when the recognition result of the separated character is not a previously designated character, the recognition result of the integrated character is prioritized, The correction value is added.

Since alphanumeric characters have a narrow character width, integrated characters are often selected. On the contrary, it is rare that the integrated character is the correct answer and the two separated characters are both alphanumeric. The invention according to claim 4 is intended to prevent erroneous recognition of alphanumeric characters by utilizing such a property. According to the invention of claim 1, when the correction value is added in advance so as to raise the selection priority of the integrated character as a whole,
The correction according to the invention is effective.

When the integrated character is a Roman numeral, the correction according to the invention of claim 4 may cause erroneous recognition due to the increased selection priority of the separated character. According to the invention of claim 5, when the integrated character is a Roman numeral even if both of the separated characters are alphanumeric, the correction for increasing the priority of the separated character is not performed. Can be increased.

As will be understood from the above description and the description of the examples below, the inventions of the first to fifth aspects can be implemented in an appropriate combination.

Since the inventions described in the above items are intended to roughly correct the tendency of character combinations, there are cases where the results are uncertain in the correction process. In addition, among such cases, there is a case in which the combination of the recognition results of the separated character and the integrated character and the correct answer can be specified in advance. The invention of claim 6 corresponds to the described case by describing a combination of specific recognition results and a correct rule regarding such a case in the form of a table or the like, and referring to this at a stage. In this case, select the correct answer immediately and omit steps and. By doing so, it is possible to more reliably select the separated character / integrated character, and it is possible to achieve the efficiency and speed of the processing.

An example of a method of selecting a separated character / integrated character according to a rule described in advance is disclosed in Japanese Patent Laid-Open No.
See, for example, 113239.

[0031]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the drawings. FIG. 3 shows an example of a block configuration of the character recognition device according to the present invention. FIG. 12 is a flowchart showing an outline of the processing after the character segmentation of this character recognition device.

In FIG. 3, a binary image input unit 101 inputs binary image data of a document from a scanner or the like. This binary image data is stored in the image memory 102. The line cutout unit 103 cuts out a line image from the binary image data in the image memory 102. The character cutout unit 104 cuts out a character element from this line image by a method such as a vertical projection method or a method of tracking the connection of black pixels (the processing of step 201 in FIG. 12). This character element may be a single character or may be a character portion. Character combination generation unit 105
Generates a combination of character elements (including a single character element) that may form a character, and writes information on the combination in the combination information memory 106 (step 202 in FIG. 12).
Processing).

FIG. 4 shows a processing example of the character "add" appearing in a character string. As shown here, for the Chinese character “K”, the character element 1 corresponding to the “K” part on the left side
51 and the character element 152 corresponding to the right portion "b" are extracted. Then, a character including only the character element 151 (separated character left), a character including only the character element 152 (separated character right), and a combined character (integrated character) of the character elements 151 and 152 are generated. It is an integrated character composed of a combination of primes 151 and 152.

The character recognizing unit 107 includes a combination information memory 1
By referring to the content of 06, the separated character and the integrated character are treated as characters, the image is read from the image memory, the respective features are extracted, and the features are collated with the dictionary 108 to recognize the recognition result (character code ) And an evaluation value representing the certainty thereof are stored in the recognition information memory 109 (step 20 in FIG. 12).
Process 3).

In the present embodiment, a value obtained by dividing the distance from the dictionary by the number of contours is used as the evaluation value in order to avoid the influence of the character size. Therefore, the smaller the evaluation value, the smaller the degree of difference from the dictionary. In addition,
It is also possible to use the similarity with the dictionary as the evaluation value, but this is because the relationship between the magnitude of the value and the dissimilarity with the dictionary is opposite to the evaluation value used in the present embodiment, and therefore the evaluation described later It is necessary to appropriately change the method of selecting the recognition result by comparing the magnitudes of the values and the method of determining which side of the separated character / integrated character the correction value described later is added to.

The correction / selection unit 110 includes the recognition information memory 1
By referring to the information in 09 and the contents of the rule table 111, the final selection of the recognition result and the correction process of the evaluation value related thereto are performed (the process of step 204 in FIG. 12). This process can be divided into five steps 204a to 204e as shown in FIG. The contents of this process will be described below.

First, in step 204a, the combination of the recognition results of the separated character and the integrated character and the rule table 11
The character combination registered in 1 is compared. When the same character combination is found, the final recognition result is determined according to the rules described for the character combination. Only when no matching character combination is found, the process proceeds to step 204b.

An example of the contents of the rule table 111 is shown in FIG. For example, the recognition result on the left of the separated character is "i", the recognition result on the right of the separated character is "hi", and the recognition result of the integrated character is "ka".
Assume that This combination matches the "rule 1" of the rule table 111, but the "selection" of the rule 1 is described as "integration". Therefore, "ka" which is the recognition result of the integrated character is selected as the final recognition result. Similarly, when the recognition result of the separated character left is “bow”, the recognition result of the separated character right is “l”, and the integrated character is “pull”, the rule 2 is matched, but the “selection” of the rule 2 is Since "Integration" is described, "Hiki" is selected as the final recognition result.

By utilizing such rules, the frequent separation / integration decisions are made in step 204.
It is possible to surely and quickly perform the processing regardless of the processing by the evaluation value after b. It should be noted that there may be a rule that a character that is neither a recognition result of a separated character nor a recognition result of an integrated character is the final recognition. This is different from selecting either the separated character recognition result or the integrated character recognition result as the final recognition result in step 204e.

Steps 204b to 204d are steps of correction processing for the evaluation value, and Step 2
Reference numeral 04e is a part for comparing the average value of the evaluation values of the separated characters after the correction processing with the evaluation value of the integrated character, and selecting the recognition result on the side having the smaller value as the final recognition result. Step 2
In connection with the processing of 04, for example, the recognition information memory 109
One data structure as shown in FIG. 5 is created for each combination of two adjacent character elements. The elements of this data structure are "character code", "evaluation value", "evaluation value correction" and "judgment" for each separated character and integrated character, and "evaluation value average value" and "average value correction" for separated characters. The recognition result and the evaluation value are stored as they are in the “character code” and the “evaluation value”. The “evaluation value correction” stores the correction value or the cumulative value to be added to the evaluation value in steps 204c and 204d. In "average value correction", the correction value is stored in step 204b. In "determination", a flag is written on the correct answer side in step 204a or step 204e. If the correct answer determined in step 204a is different from the recognition result, the "character code" is rewritten with the character code of the correct answer.

In step 204b, a process of adding a predetermined plus correction value α (for example, 10) to the average value of the evaluation values of the separated characters is performed (in the direction of increasing the selection priority of integrated characters). Will be corrected). Practically, this addition operation is executed in step 204e, and in the step 204b, "average value correction" is performed on the separated character left and the separated character right of the data structure shown in FIG.
The correction value α is written in.

As an example, when recognizing the "ka" of a kanji appearing in a certain character string, as shown in FIG. 7, the separated character "i" on the left and the separated character "hi" on the right. ,
It is assumed that the integrated character “” is obtained, and the evaluation values of 65, 75, and 70 are obtained. Step 20
4b to 204d, if the correction is not added, the average value of the evaluation values on the left and right of the separated character and the evaluation value of the integrated character are both 7
It becomes 0.

In the case of a character that can be a separated character, the evaluation value of both the integrated character and the separated character is small, so such a case often occurs. So, in step 204b,
If the correction value α = 10 is added to the average value of the separated characters, FIG.
In the example shown in, the average value of the evaluation values of the separated characters is increased to 80, and the evaluation value of the integrated character is relatively decreased (the recognition result of the integrated character has a higher selection priority). As a result, the integrated characters are integrated in step 204e. Characters will be selected correctly.

In step 204c, if the size of the separated character (character element) is equal to or smaller than the size corresponding to the one-byte character and the recognition result of the separated character is an unsuitable character type (hiragana, katakana, kanji), In order to increase the selection priority of integrated characters, a positive correction value β (for example, 1
0) is added to the evaluation value of the separated character. Practically, FIG.
The correction value β is added to the values of the “evaluation value correction” (initial value is 0) on the left and right of the separated characters in the data structure shown in FIG. However, even if the character element size is equal to or less than the half-width character and the recognition result of the separated character is an unsuitable half-width character type, the recognition result is a specific character with a narrow character width (for example, "ku").
There is an exceptional process in which the correction value β is not added when “to”, “mo”, “ri”, “i”, “i”, “to”, “tsu”, “ri”, etc.).

The purpose of the correction in step 204c will be described with reference to a concrete example. In order to facilitate understanding, immediately before step 204b and immediately after step 2
The description will be made assuming that the correction of 04d is not performed.

As an example, consider the case of processing the Chinese character "ki" appearing in a certain character string. As shown in FIG. 8, the recognition result of the separated character left is “Koto”, the evaluation value thereof is 60, the recognition result of the separated character right is “O”, the evaluation value thereof is 70, and the recognition result of the integrated character is “ Note that the evaluation value is 70. Further, it is assumed that the size of a character element (character width) is about half of a full-width character and is determined to be “half-width size”. However, the word "Koto" in the recognition result to the left of the separated character is Kanji and belongs to a character type with an incorrect half-size character. Therefore, in order to prevent it from being selected, that is, to increase the selection priority of integrated characters, The correction value β = 10 is added to the “evaluation value correction” on the left. Since the same applies to the right of the separated character, the correction value β = 10 is added to the “evaluation value correction”.

Therefore, the average value of the values after adding the value of "evaluation value correction" to the value of "evaluation value" of the separated character is 75, which is larger than the value 70 of "evaluation value" of the integrated character,
As a result, the correct recognition result "" of the integrated character is selected as the final recognition result.

On the other hand, if the correction value β is not added, the average value of the evaluation values of the separated characters becomes 65, which is smaller than the table value 70 of the integrated characters, so that the separated characters are erroneously selected. Such an error is less likely to occur by the correction in step 204c.

Next, the exception processing will be specifically described. As an example, consider the case of processing a character string "ito" appearing in a character string. Each character in this string "ito" is a single character, but since it is a character with a narrow width, when printed at proportional pitch, the spaces between the characters may be clogged, and these integrated characters may be generated. is there. In addition, since each character is vertically long, it may be determined to be a half-size character.

When the recognition result and the evaluation value as shown in FIG. 9 are obtained for this character string "ito", the correction value β is added to the evaluation value of the separated characters "a" and "to" as in the example of FIG. = 10
Is added, the average value of the evaluation values of the separated characters becomes larger than the evaluation value of the integrated character, so that the integrated character “Ito” is selected. However, since the separated characters “a” and “to” correspond to the specific character that is the target of the exception processing, the correction value β = 10 is not added to each evaluation value. Therefore, the average value of the evaluation values of the separated characters is 70, which is smaller than the evaluation value 75 of the integrated characters. As a result, "i" and "to", which are the recognition results of the separated characters, are correctly selected instead of the recognition results of the integrated characters.

In the next step 204d, when the recognition results of the two separated characters are both letters or numbers, a positive value is added to the evaluation value of the integrated character in order to increase the selection priority of the recognition result of the separated characters. Correction value γ (for example, 1
0) is added. Practically, the value of "evaluation value correction" for the integrated character of the data structure shown in FIG. 5 (initial value is 0)
The correction value γ = 10 is added to. However, there is an exception process in which the correction value γ is not added when the recognition result of the integrated character is a Roman numeral even if both recognition results of the separated characters are alphabetic characters or numbers.

The purpose of the correction in step 204d will be described with reference to a concrete example. In addition, in order to facilitate understanding, it is assumed that the corrections in steps 204b and 204c are not performed.

As an example, "2 appearing in a character string
Consider the case of processing the character string "5". When the characters "2" and "5" are printed in half-size characters, the integrated character "25" is generated because the character width is narrow.

As shown in FIG. 10, when the recognition results of the separated characters are "2" and "5", the evaluation values thereof are 45 and 65, the recognition result of the integrated characters is "phantom", and the evaluation value thereof is 50. To do.
Therefore, when the correction is not performed, the average value 55 of the separated character evaluation values is larger than the integrated character evaluation value 50.
The integrated character “25” is selected, which is not the correct answer. However, since the recognition results of the separated characters are both numbers, the correction value γ = 10 is added to the evaluation value of the integrated character, and as a result, the evaluation value of the integrated character after this addition is increased to 60 and the evaluation of the separated character is performed. Since the average value of the values is larger than 55, the recognition results “2” and “5” of the separated character are correctly selected.

Next, the exception processing will be specifically described. As an example, consider the case of processing a Roman numeral “IV” that appears in a character string. This Roman numeral is one character, but it is separated into two character elements "I" and "V" by character cutting and each is generated as a character, and the integrated character "IV" is also generated. .
When the recognition result and the evaluation value of each separated character and integrated character are as shown in FIG. 11, it is assumed that both separated characters are alphabetic characters as described above, and the correction value γ is added to the integrated character evaluation value.
== 10 is added, the evaluation value after the addition is 6
As a result of becoming 0 and becoming larger than the average value 55 of the separated character evaluation values, the separated character recognition result is erroneously selected. However, since the recognition result of the integrated character is Roman numeral “IV” and corresponds to the above exception processing, the correction value γ is not added to the evaluation value of the integrated character, and as a result, the average value of the evaluation values of the separated characters is obtained. Since the integrated character evaluation value 50 is smaller than 55, the integrated character recognition result is correctly selected.

In step 204e, the average value of the evaluation values of the separated characters after the above correction processing is compared with the evaluation value of the integrated character, and the recognition result with the smaller value is selected as the final recognition result. . Specifically, after adding the value of "evaluation value correction" to the value of "evaluation value" on the left and right of the separated characters in the data structure shown in FIG. 5, the average value is calculated, and this average is calculated. The value of "average value correction" is added to the value to obtain the corrected average value. Further, the value of "evaluation value correction" is added to the value of "evaluation value" of the integrated character to obtain a corrected evaluation value. When the corrected average value of the separated characters <the corrected evaluation value of the integrated characters, the recognition result of the separated characters is selected as the final recognition result, and the “judgment” of the separated characters is flagged. If not, the recognition result of the integrated character is selected as the final recognition result, and the "judgment" of the integrated character is flagged.

The final recognition result thus obtained is output from the result output unit 112. Specifically, referring to the data structure shown in FIG. 5 in the recognition information memory 109, if the flag of “judgment” of the integrated character is set, the code of the “character code” of the integrated character is output. If a flag is set in the "judgment" of the separated character, the code of the "character code" of the separated character left and the separated character right is output.

In the above description, the data structure as shown in FIG. 5 is used, but this is only an example, and the specific processing algorithm and data structure may be appropriately changed. It is a thing.

[0059]

According to the invention described in each of claims 1 to 6, it is possible to reduce the recognition error due to the erroneous separation or merging of characters, improve the recognition rate, and correct the evaluation value of the recognition result. Is an addition operation that requires a small amount of processing, so that high-speed processing can be achieved. According to the second aspect of the present invention, by paying attention to the relationship between the character element size and the character type of the recognition result, it is possible to reduce erroneous recognition of a character or a half-width character that is separated into a bias and a drought. According to the invention described in claim 3, in the invention described in claim 2, it is possible to reduce erroneous recognition of a specific character having a narrow width that causes an error. According to the invention described in claim 4, it is possible to reduce erroneous recognition of alphanumeric characters. According to the invention described in claim 5, it is possible to reduce erroneous recognition of Roman numerals. According to the invention as set forth in claim 6, it is possible to reliably prevent erroneous recognition with a small processing amount for a character that cannot be erroneously recognized by the inventions as claimed in each of claims 1 to 5 or a character for which separation or integration can be predicted. can do.

[Brief description of drawings]

FIG. 1 is a diagram showing a distribution example of a set in which a separated character is a correct answer and a set in which an integrated character is a correct answer in association with selection of a separated character / integrated character.

FIG. 2 is a diagram showing a distribution example of a set in which a separated character is a correct answer and a set in which an integrated character is a correct answer, in association with selection of a separated character / integrated character.

FIG. 3 is a block diagram showing a configuration example of a character recognition device according to the present invention.

FIG. 4 is a diagram for explaining extraction of a character element and generation of a character combination.

FIG. 5 is a diagram showing an example of a data structure of recognition information.

FIG. 6 is a diagram showing an example of a rule table.

FIG. 7 is a diagram showing an example of recognition results of separated characters and integrated characters, evaluation values, and the like.

FIG. 8 is a diagram showing an example of recognition results of separated characters and integrated characters, evaluation values, and the like.

FIG. 9 is a diagram showing examples of recognition results of separated characters and integrated characters, evaluation values, and the like.

FIG. 10 is a diagram showing examples of recognition results of separated characters and integrated characters, evaluation values, and the like.

FIG. 11 is a diagram showing an example of recognition results of separated characters and integrated characters, evaluation values, and the like.

FIG. 12 is a flowchart showing an outline of processing from character extraction to determination of a final recognition result.

[Explanation of symbols]

 101 binary image input unit 102 image memory 103 line cutout unit 104 character cutout unit 105 character combination generation unit 106 combination information memory 107 character recognition unit 108 dictionary 109 recognition information memory 110 correction / selection unit 111 rule table 112 result output unit

Claims (6)

[Claims] 1. A character recognition as two separated characters and a character recognition as one integrated character are performed for two adjacent character elements extracted from an image, and the recognition result of the separated characters and the recognition result Obtaining an evaluation value representing the certainty and a recognition result of the integrated character and an evaluation value representing the certainty, and performing a correction process for the evaluation value of the separated character and the integrated character evaluation value, A final recognition result of the recognition result of the separated character or the recognition result of the integrated character according to the magnitude relationship between the average value of the evaluation values of the separated character after the correction process and the evaluation value of the integrated character. And a correction value is added to the evaluation value of the separated character or its average value or the evaluation value of the integrated character in the correction processing step. Character recognition method. 2. A character recognition as two separated characters and a character recognition as one integrated character are performed for two adjacent character elements extracted from an image, and the recognition result of the separated characters and the recognition result thereof. Obtaining an evaluation value representing the certainty and a recognition result of the integrated character and an evaluation value representing the certainty, and performing a correction process for the evaluation value of the separated character and the integrated character evaluation value, A final recognition result of the recognition result of the separated character or the recognition result of the integrated character according to the magnitude relationship between the average value of the evaluation values of the separated character after the correction process and the evaluation value of the integrated character. In the correction step, the size of the character element is equal to or smaller than the half-width character equivalent size, and the recognition result of the separated character is unsuitable as the half-width character. When the character belongs to a fixed character type, a correction value is added to the evaluation value of the separated character or its average value or the evaluation value of the integrated character in order to increase the selected priority of the recognition result of the integrated character. Character recognition method characterized by being performed. 3. A character recognition as two separated characters and a character recognition as one integrated character are performed for two adjacent character elements extracted from an image, and the recognition result of the separated characters and the recognition result thereof. Obtaining an evaluation value representing the certainty and a recognition result of the integrated character and an evaluation value representing the certainty, and performing a correction process for the evaluation value of the separated character and the integrated character evaluation value, A final recognition result of the recognition result of the separated character or the recognition result of the integrated character is obtained according to the magnitude relationship between the average value of the evaluation values of the separated characters after the correction processing and the evaluation value of the integrated character. In the correction step, the size of the character element is equal to or smaller than the half-width character equivalent size, and the recognition result of the separated character is unsuitable as the half-width character. Belongs to a certain character type, and
When the recognition result of the separated characters is not a predetermined character, the evaluation value of the separated characters or the average value thereof or the integrated character is increased in order to increase the priority of selection of the recognition result of the integrated characters. A character recognition method characterized in that a correction value is added to the evaluation value of. 4. The character recognition as two separated characters and the character recognition as one integrated character are performed on two adjacent character elements extracted from an image, and the recognition result of the separated characters and the recognition result Obtaining an evaluation value representing the certainty and a recognition result of the integrated character and an evaluation value representing the certainty, and performing a correction process for the evaluation value of the separated character and the integrated character evaluation value, A final recognition result of the recognition result of the separated character or the recognition result of the integrated character according to the magnitude relationship between the average value of the evaluation values of the separated character after the correction process and the evaluation value of the integrated character. And a step of selecting the recognition result of the separated characters when both of the recognition results of the two separated characters are alphabetic characters or numbers in the correction step. A character recognition method, wherein a correction value is added to an evaluation value of the integrated character or an evaluation value of the separated character or an average value thereof in order to increase the priority of the character recognition. 5. The character recognition as two separated characters and the character recognition as one integrated character are performed on two adjacent character elements extracted from an image, and the recognition result of the separated characters and the recognition result thereof. Obtaining an evaluation value representing the certainty and a recognition result of the integrated character and an evaluation value representing the certainty, and performing a correction process for the evaluation value of the separated character and the integrated character evaluation value, A final recognition result of the recognition result of the separated character or the recognition result of the integrated character according to the magnitude relationship between the average value of the evaluation values of the separated character after the correction process and the evaluation value of the integrated character. In the correction step, the recognition results of the two separated characters are both letters or numbers, and the integrated character recognition result is a Roman character. When it is not a number, a correction value is added to the evaluation value of the integrated character or the evaluation value of the separated character or its average value in order to increase the priority of the recognition result of the separated character selected. Character recognition method characterized by. 6. The character recognition method according to claim 1, 2, 3, 4 or 5, wherein prior to the step of correction processing,
It is determined whether the combination of the recognition result as the separated character and the recognition result as the integrated character matches one of a plurality of predetermined specific character combinations, and if they match, the predetermined specific character combination is predetermined. A character recognition method, comprising: determining a final recognition result according to a rule set, and performing the step and the step only when the final recognition result is not selected by the step.