JP4861730B2 - Character recognition device, character recognition method, character recognition program, and integrated circuit - Google Patents

Description

  The present invention relates to a character recognition technique, and more particularly to a character recognition apparatus and method for recognizing characters with a large variation in character shape, such as documents of various typefaces and handwritten forms.

In the conventional character recognition device, feature quantities extracted from a large number of learning character images are clustered, a classification dictionary storing average feature quantities for each cluster as reference feature quantities, and a character image cut out from the read image. Classification feature extraction means for extracting several features, classification means for collating the extracted feature quantity with the reference feature quantity in the classification dictionary, and selecting candidate clusters, and among candidate clusters obtained by the classification means Detailed identification means for identifying which cluster is selectively using only the feature amount that is likely to be different between both clusters from the feature amount used in the classification for a specific group that is likely to be erroneously recognized. There is a verification unit that verifies the validity of the recognition result by using the characteristic value range characteristic for the cluster selected as the first place by the detailed identification unit (see Patent Document 1).
JP-A-8-305804

  However, in the conventional configuration, since the same feature value is used for the classification process, the detailed identification process, and the verification process, the feature value extracted from the cut-out character image is particularly classified in the case of a handwritten character with a significant variation in character shape. If the similarity is low with respect to any of the reference features in the dictionary, the classification process, detailed identification process, and verification process do not function well, and erroneous recognition and rejection of the recognition result It has the problem of increasing.

  The present invention solves the above-mentioned conventional problems. In reading a document or a form, it correctly recognizes even a handwritten character whose character shape is fluctuating, and recognizes a character whose recognition result is wrong or difficult to read. An object of the present invention is to provide a character recognition device that rejects a result with high accuracy.

  In order to solve the above-described problems, the present invention is a character recognition device that receives input of a document image and recognizes each character included in the document image, and includes a classification dictionary in which a reference feature amount of each character is registered, and the document image A character image extracting unit that extracts a character image from the image, a feature amount extracting unit that extracts a feature amount from the extracted character image, a similarity between the extracted feature amount and the reference feature amount is calculated, A candidate character selecting unit that selects a character having a high value as a candidate character, a similarity calculating unit that calculates a similarity between character images in the same document using the feature amount extracted by the feature amount extracting unit, and a candidate A recognition verification means for verifying the certainty of recognition as a candidate character based on the similarity as a character and the similarity between the character images is provided.

With the above-described configuration, even in the case of handwritten characters with a significant variation in character shape, the recognition result is verified by reflecting the similarity between character images in the same document, so that the document image can be recognized accurately. Can do.
The recognition verification unit includes a calculation unit that calculates a product of the similarity r with the verification target character calculated by the similarity calculation unit and the similarity R calculated by the candidate character selection unit; A recognition result changing unit that changes the recognition result of the verification target character to the other candidate character when there is another candidate character having a product larger than the product of the same candidate character selected by the candidate character selecting means. To have.

With such a configuration, it is possible to change the selected candidate character to another candidate character in the same document by utilizing the fact that character images in the same document have similar characteristics.
In addition, an individual verification dictionary that describes, as rules, an affirmative condition indicating a morphological feature quantity to be provided as a candidate character for each character and a negative condition indicating a morphological feature quantity that should not be provided as a candidate character, and a form from the character image An individual character verification means for extracting a feature amount, determining whether the selected candidate character satisfies an affirmative condition or a negative condition, and rejecting the selection of the candidate character when it is determined that the candidate character is satisfied Is further provided.

With such a configuration, the selected candidate characters are individually checked against the rules, and the candidate characters that cannot exist are rejected, so that erroneous character recognition can be prevented.
Also, a similar candidate character identification dictionary that describes two character feature quantities with high similarity and rules for identifying two characters, and a candidate character selected from one character image is a similar candidate character identification dictionary And a candidate character determination means for extracting a morphological feature amount from a character image, collating with the rule, and determining which of the two characters is a candidate character. .

With such a configuration, it is verified whether it is more suitable as a candidate character to determine which of the two characters has high similarity, so that the accuracy of character recognition is improved.
The form feature amount includes at least one of a loop position or number, an average linear density in a predetermined range, a convex end point position or number, and a concave end point position or number.
With such a configuration, it is possible to improve the accuracy of character recognition and reject at least a misrecognized candidate character by verifying a candidate character selected using a unique characteristic morphological feature.

  A character recognition method of a character recognition device having a classification dictionary in which reference feature amounts of characters are registered, receiving input of a document image, and recognizing each character included in the document image. A character image extraction step for extracting the feature amount, a feature amount extraction step for extracting the feature amount from the extracted character image, a similarity between the extracted feature amount and the reference feature amount is calculated, and a character with high similarity A candidate character selection step for selecting a character as a candidate character, a similarity calculation step for calculating a similarity between character images in the same document using the feature amount extracted in the feature amount extraction step, and a candidate character as Based on the similarity and the similarity between the character images, a recognition verification step for verifying the certainty of recognition as a candidate character is included.

By using such a method, even in the case of handwritten characters with significant variations in character shape, the similarity between character images in the same document is reflected to verify whether the recognition result is correct, so that the document image can be recognized accurately. Can do.
In addition, a character image extraction unit that has a classification dictionary in which reference feature amounts of each character are registered, extracts a character image from a document image, and receives a document image and recognizes each character included in the document image. A feature amount extraction step for extracting the feature amount from the extracted character image, a similarity between the extracted feature amount and the reference feature amount is calculated, and a character having a high similarity is selected as a candidate character A candidate character selection step, a similarity calculation step of calculating a similarity between character images in the same document using the feature amount extracted in the feature amount extraction step, a similarity as a candidate character, and the character A character recognition program for executing a recognition verification step of verifying the certainty of recognition as a candidate character based on the similarity between images.

By applying such a character recognition program to a character recognition device, it is verified whether the recognition result is correct by reflecting the similarity between character images in the same document, even in the case of handwritten characters with significant variations in character shape. Therefore, the document image can be recognized accurately.
An integrated circuit of a character recognition device that receives input of a document image and recognizes each character included in the document image, and extracts a character image from the classification dictionary in which the reference feature amount of each character is registered. A character image extracting unit, a feature amount extracting unit for extracting the feature amount from the extracted character image, a similarity between the extracted feature amount and the reference feature amount is calculated, and a character having a high similarity is selected as a candidate character Candidate character selection means for selecting as, a similarity calculation means for calculating the similarity between character images in the same document using the feature amount extracted by the feature amount extraction means, and the similarity as a candidate character And a recognition verification means for verifying the certainty of recognition as a candidate character based on the similarity between the character images.

  By using such an integrated circuit for a character recognition device, even in the case of handwritten characters whose character shape fluctuates significantly, it is verified whether the recognition result is correct by reflecting the similarity between character images in the same document. Document images can be accurately recognized.

Hereinafter, embodiments of a character recognition device according to the present invention will be described with reference to the drawings.
(One embodiment)
FIG. 1 is a block diagram of an embodiment of a character recognition device according to the present invention.
This character recognition device includes a classification dictionary 101, a similar candidate character identification dictionary 102, an individual verification dictionary 103, a document image input reception unit 104, a character image extraction unit 105, a feature amount extraction unit 106, and a candidate character selection. Unit 107, morphological feature amount extraction unit 108, candidate character determination unit 109, individual character verification unit 110, similarity calculation unit 111, recognition verification unit 112, and recognition result output unit 113.

  The classification dictionary 101 is composed of a magnetic disk or the like and has a data structure as shown in FIG. As shown in FIG. 3 as an example, a large number of dictionary data 306 including a cluster number 301, an identifier 302, a character code 303, an attribute flag 304, and an average feature 305 are registered. FIG. 3 shows the character “middle” as indicated by the identifier 302. The classification dictionary 101 is normally a multi-template dictionary that includes a plurality of reference feature amounts having different character shapes for one character. In the case of a multi-template dictionary, candidate characters may include a plurality of identical characters with different clusters.

The similar candidate character identification dictionary 102 includes a magnetic disk or the like. FIG. 4 shows an example of the similar candidate character identification dictionary. In the similar candidate character identification dictionary 102, a similar character combination 402, a feature amount 403, and a rule 404 are described. In the similar character combination 402, two characters having high similarity are described.
The feature quantity 403 describes the form feature quantity for identifying these two characters. The form feature amount includes a loop position, the number of loops, an average linear density in a predetermined range, a convex end point position, a convex end point number, a concave end point position, a concave end point number, and the like.

The rule 404 describes a criterion for determining which of the two characters is to be determined using this form feature amount.
FIG. 5 illustrates the form feature quantities “0-6”, “4 (closed type) -9”, and “8 (slanted type) -9” of the similar character combinations described in the similar candidate character identification dictionary 102. It is a figure for doing.
FIG. 5A shows the relationship between the morphological feature amounts of the handwritten characters “0” 501 and “6” 502 and the determination criteria. The number of lines crossed by the horizontal lines 511 and 512 drawn above the handwritten characters “0” 501 and “6” 502, that is, the line density, is “2” and “1”, respectively. As a reference, when “1.5” or more, a score is added to “0”, and when it is less than “6”, a score is added. Both the handwritten characters “0” 501 and “6” 502 form a loop. For the center position of this loop (indicated by x in the figure), the height 513, 514 from the bottom of the circumscribed rectangle of the character is used as the feature value, and in the case of the lower side, “6” is added, In this case, “0” is added. Although not described as “0-6” as a feature value, a recessed point 515 is formed in the handwritten character “6” 502.

FIG. 5B shows the relationship between the morphological feature amounts of the handwritten characters “4” 503 and “9” 504 and the determination criteria. A right-pointing convex end point 521 is formed at the center on the right side of the handwritten character “4” 503. If the convex end point 521 is present as the form feature amount, a criterion for adding “4” is provided.
FIG. 5C shows the relationship between the morphological feature amounts of the handwritten characters “8” 505 and “9” 506 and the determination criteria. A criterion for adding points to “9” is added when the lower average line widths 531 and 532 are equal to or less than the line width of the entire character, and “8” is added otherwise. In addition, if there is a center-left opening concave point 533 as a form feature amount, a determination criterion for adding “9” is provided.

The feature amount 403 and the rules 404 described with reference to FIGS. 5A to 5C are described in the similar candidate character identification dictionary 102 for a large number of similar character combinations 402.
The individual verification dictionary 103 is composed of a magnetic disk or the like. FIG. 6 shows an example of the individual verification dictionary 103. The individual verification dictionary 103 describes positive conditions 603, 605,..., Negative conditions 604, 606,.

FIG. 7 is a diagram for explaining the contents of the individual verification dictionary 103. FIG. 7A shows handwritten characters “0” 701 and 702. The loop center position of the handwritten character “0” 701 is at the center of the character circumscribed rectangle. The number of loops is “1”, the number of convex end points is “0”, and the number of concave end points is “0”. The average linear density of the central region 711 is “2”.
The average linear density of the center region 712 of the handwritten character “0” 702 is “3”, the number of convex end points is “2”, and the positions of the convex end points are in the upper right portion 713 and the lower left portion 714.

  In the individual verification dictionary 103, such morphological features are described as (1) and (2) in the rule 607 which is the determination condition of the affirmative condition 603. The negative condition 604 of the individual verification dictionary 103 describes morphological feature quantities that are impossible for the handwritten characters “0” 701 and 702. For example, in the negative condition 604, the number of loops is “3” or more, the average linear density of the central region is “4” or more, and the like.

In FIG. 7B, handwritten characters “4” 703 and 704 are shown. The handwritten character “4” 703 has one loop 720. The convex end point positions are four convex end points at the upper right part 722, the left central part 723, the right central part 724, and the lower part 725. The average linear density of the upper region 721 is “2”. The average linear density of the lower region 731 is “1”.
According to the rule 607 (1) of the affirmative condition 605 for the character “4” in the individual verification dictionary 103, the number of loops is 1, and the convex end point is determined if 3 or 4 convex end points in the 4 convex end points are confirmed. The average linear density of the upper region is greater than 1.5 and the average linear density of the lower region is described as being less than 1.5.

The handwritten character “4” 704 has no loop. Convex end point positions are five convex end points at the upper left portion 726, the upper right portion 729, the left central portion 727, the right central portion 728, and the lower portion 730. The average linear density of the upper region 721 is “2”. The average linear density of the lower region 731 is “1”.
According to the rule 607 (2) of the affirmative condition 605 for the character “4” in the individual verification dictionary 103, if the number of loops is 0 and the convex end point is confirmed as 4 or 5 convex end points among the five convex end points, the determination condition The average linear density of the upper region is greater than 1.5 and the average linear density of the lower region is described as being less than 1.5.

In the rule 607 of the negative condition 606 of the individual verification dictionary 103, morphological feature quantities that are impossible for the handwritten characters “4” 703 and 704 are described. It is described that the number of convex end points is 6 or more, the average linear density of the upper region is 3 or more, and the average linear density of the lower region is 2 or more.
The document image input receiving unit 104 includes an optical reading device such as a scanner. The document image input receiving unit 104 receives a document image input from a user, converts the document image into binary image data, cuts out a character line from the document image, and outputs the character line to the character image extraction unit 105. .

The character image extraction unit 105 extracts a character image from the character line input from the document image input reception unit 104, and outputs the character image to the feature amount extraction unit 106, the morphological feature amount extraction unit 108, and the authentication verification unit 112.
FIG. 8 shows a state in which the character images 803 to 806 of the character string 802 on the first line in the document image 801 received by the document image input receiving unit 104 are extracted by the character image extracting unit 105.

The feature amount extraction unit 106 extracts feature amounts from the character images 803 to 806, and notifies the candidate character selection unit 107 and the similarity calculation unit 111 of the feature amounts.
The candidate character selection unit 107 calculates the similarity between the notified feature amount and the reference feature amount of the classification dictionary 101, and selects N candidate characters in descending order of similarity. Here, for example, the urbanization distance or the Euclidean distance is used for the calculation of the similarity. This selection of candidate characters is described in detail in the above-mentioned Patent Document 1.

  FIG. 9 is a diagram illustrating an example of candidate characters selected by the candidate character selection unit 107. Here, N is set to 10. Here, as described above, since the cluster number 901 uses a multi-template dictionary, a plurality of clusters exist for the same character. The similarity 902 is indicated by an integer up to the second decimal place. Character images 803 to 806 in the character line 802 are selected as the first candidate characters “middle”, “inside”, “sato”, and “ho”, respectively.

The candidate character selection unit 107 notifies the selected candidate character to the candidate character determination unit 109 and the recognition verification unit 112.
The morphological feature amount extraction unit 108 stores the character image input from the character image extraction unit 105 in the storage area, and extracts the morphological feature amount from the character image in accordance with the instruction of the candidate character determination unit 109, and the candidate character The determination unit 109 is notified. Similarly, when the notification of the rule of the character affirmative condition and the negative condition is received from the individual character verification unit 110, the form feature amount corresponding to the rule is extracted from the corresponding character image and notified to the individual character verification unit 110.

  Candidate character determination unit 109 combines the first to Nth candidate characters notified from candidate character selection unit 107. It is determined whether the combination of each candidate character is the same character (whether the cluster is different but the same character). If it is not the same character, it is determined whether it is described in the similar character combination 402 of the similar candidate character identification dictionary 102. When it is described in the similar character combination 402, the dictionary data is read out. At the same time, the feature quantity 403 of the similar candidate character identification dictionary 102 is notified to the morphological feature quantity extraction unit 108, and an instruction is given to extract the morphological feature quantity from the character image of the candidate character. The morphological feature amount notified from the morphological feature amount extraction unit 108 is compared with the rule 404 of the similar candidate character identification dictionary 102 to determine which of the similar characters is superior.

  For example, when the first candidate character of the handwritten character “0” 501 is “6” and the second candidate character is “0” in FIG. Since it is “2” and the center position of the maximum loop is almost the center, “0” is added according to the determination criterion, which is advantageous. Accordingly, the candidate character determination unit 109 copies the first candidate character to “0” and the second candidate character to “6”. Candidate character determination unit 109 notifies individual character verification unit 110 of the determination result.

  When the individual character verification unit 110 receives the notification of the candidate character determined from the candidate character determination unit 109, the individual character verification unit 110 reads the rule 607 of the positive condition and the negative condition described for each corresponding character 602 in the individual verification dictionary 103, The morphological feature quantity extraction unit 108 is notified. When the individual character verification unit 110 receives the notification of the morphological feature amount from the morphological feature amount extraction unit 108, the individual character verification unit 110 compares the rule of the positive condition and the negative condition with the morphological feature amount, and whether the positive condition and the negative condition are satisfied. Determine whether or not. The individual character verification unit 110 “dismisses” the first candidate character when the negative condition is satisfied with the first candidate character, and the first condition when the negative condition is not satisfied and the positive condition is satisfied with the first candidate character. The position candidate character is set to “valid” and the process is terminated. When neither a negative condition nor an affirmative condition is satisfied for the first candidate character, the rule 607 of the affirmative condition and negative condition of the corresponding character in the individual verification dictionary 103 is read for the second candidate character, and morphological feature amount extraction is performed. Notification to the unit 108. When receiving the notification of the morphological feature amount from the morphological feature amount extraction unit 108, the rule of the affirmative condition and the negative condition is compared with the morphological feature amount, and it is determined whether the positive condition and the negative condition are satisfied. When the negative condition is not satisfied and the positive condition is satisfied, the first candidate character is “rejected”. When a negative condition is satisfied, or when neither a negative condition nor an affirmative condition is satisfied, the same processing is performed for the third and lower candidate characters. When neither a negative condition nor an affirmative condition holds for all candidate characters, the first candidate character is verified as valid.

From the viewpoint of preventing misrecognition as much as possible, the first candidate character may be “rejected” when a negative condition and an affirmative condition are not satisfied for all candidate characters.
To explain using a specific example, when the first candidate character of the character image 704 in FIG. 7B is “4”, the individual character verification unit 110 determines the positive condition for the character “4” in the individual verification dictionary 103. And the negative condition rule are read out and notified to the morphological feature quantity extraction unit 108. In the morphological feature quantity extraction unit 108, the number of loops “0”, the number of convex end points “5”, the average line density “2” of the upper area, and the average line of the lower area are stored from the character image 704 stored in the storage area. The density “1” is extracted and notified to the individual character verification unit 110. The individual character verification unit 110 matches the rule 607 (2) loop number 0, convex end point number 5, upper region average line density 1.5 or more, lower region average line density 1.5 or less. It is determined that the condition is satisfied, and it is determined that the negative conditions of the number of convex end points of 6 or more, the average linear density of the upper region of 3 or higher, and the average linear density of 2 or lower of the lower region are not satisfied, and the first candidate character “4” Is verified as “valid”.

In rule 607 (2), the number of convex end points is set to 4 or 5 in consideration of the case where all the convex end points cannot be extracted by the morphological feature quantity extraction unit 108. Similarly, in the positive condition (1) for the character “4”, the number of convex end points is set to 3 or 4 in consideration of the case where all the convex end points cannot be extracted.
The individual character verification unit 110 notifies the recognition verification unit 112 of the verification result.

  The similarity calculation unit 111 stores the feature amount of the character image notified from the feature amount extraction unit 106 in the storage area. When a candidate character to be verified is designated from the recognition verification unit 112, the similarity between the corresponding feature amount and the feature amount of the candidate character on the same line is calculated. The calculated similarity is notified to the recognition verification unit 112. The similarity is calculated based on the urbanization distance or the like and normalized to a range of 0 to 1. At this time, the larger the numerical value, the higher the similarity.

The recognition verification unit 112 determines whether or not the first candidate character verified by the individual character verification unit 110 should be adopted as a recognition result for the entire document image or a series of character images such as the same line or the same field of a form. In order to make a determination, inter-character verification is performed based on the position and shape of each character and the similarity between the characters.
The recognition verification unit 112 stores each character image input from the character image extraction unit 105 in a storage area, calculates the character width and character height of each character image, and calculates the average character width and average character height. Then, the first candidate character exceeding the predetermined threshold value from the average value and the first candidate character below the predetermined threshold value from the average value are rejected.

  FIG. 10 shows a character string image 1001 written in one line. Here, for each of the character images C1011 to C1018, it is determined whether or not there is a character image having a character width exceeding the threshold value from the average character width 1003 or a character image having a character height exceeding the threshold value from the average character height 1004. . Character image C1011 is rejected because it has a character width that exceeds the threshold value from average character width 1003 and has a character height that exceeds the threshold value from average character height 1004. Character image C1017 is rejected because it has a character width equal to or smaller than a threshold value from average character width 1003 and has a character height equal to or smaller than a threshold value from average character height 1004.

In addition, the recognition verification unit 112 obtains an average character position 1002 (vertical position for horizontal writing, horizontal position for vertical writing), and rejects a character image having a large deviation from the average character position. For example, the character image C1018 is rejected because the character position shift is large.
Note that even if the character image has a character width less than the threshold value from the average character width 1003 or has a character height less than the threshold value from the average character height 1004, the first candidate character is “.”, “−. In the case of a symbol such as “” or “•”, the character image is not rejected.

In addition, the same character shape character is identified by the character positions such as “.” And “•”, “_” and “−”, and position processing is performed to correct the first candidate character as necessary.
Next, the recognition verification unit 112 performs inter-character verification based on the similarity between the characters.
When the recognition verification unit 112 receives a notification of the recognition result that the first candidate character is “valid” from the individual character verification unit 110, the recognition verification unit 112 notifies the similarity calculation unit 111 of the candidate character to be verified, and the feature amount An instruction is given to calculate the similarity between the character string of one line notified from the extraction unit 106 and other candidate characters.

When the similarity calculation result is obtained from the feature amount extraction unit 106, the product of the similarity notified from the candidate character selection unit 107 is obtained.
A specific example will be described with reference to FIGS.
FIG. 11 shows character images C1111 to C1114 of the character string 1101 described in one line. The selection and verification results of the candidate character selection unit 107 and candidate character determination unit 109 and the individual character verification unit 110 of these character images C1111-C1114 are “4” as the first candidate character, as shown in FIG. “6”, “5”, and “6”. The similarity R to the reference feature amount of the classification dictionary 101 when selected as the first candidate character by the candidate character selection unit 107 is “0.7”, “0.5”, “0.6”, respectively. ”,“ 0.8 ”. In the verification result in the individual character verification unit 110, all are verified as “valid”.

Now, let the character image C1112 be the character (C) to be verified. The character image C1112 is selected as the verification target character (C) because the character image C1114 of the same character string 1101 is selected as the same candidate character “6” as the character (C).
The similarity calculation unit 111 calculates the similarity r between the character image C1112 and the other character images C1111, C1113, and C1114. As a result, the similarities r of the other character images C1111, C1113, and C1114 are “0.1”, “0.75”, and “0.3”, respectively.

  The authentication verification unit 112 sets the character image C1113 having the maximum similarity r as a character (C1). The character image C1114 selected as the same candidate character as the character (C) is set as the character (C2). If the candidate characters of the character (C1) and the character (C2) are the same, the candidate character of the verification target character (C) is valid. However, the character (C1) is “5” and the character (C2) is “6”.

The r · R values of these characters (C1) and (C2) are calculated by the authentication verification unit 112 as “0.45” and “0.24”, respectively. If the product of r · R is C1> C2, the authentication verification unit 112 changes the recognition result of the verification target character (C) to a candidate character of the character (C1). When C1 <= C2, the verification target character (C) is rejected.
In this case, since C1> C2, the candidate character “6” of the character image C1112 is changed to the candidate character “5” of the character (C1). The verification result is “valid”.

The recognition verification unit 112 notifies the recognition result output unit 113 of the first candidate character determined to be “valid” as the recognition verification result. For the first candidate character determined as “rejected”, an error is notified.
The recognition result output unit 113 displays the first candidate character notified from the recognition verification unit 112 as a recognition result, for example, on a display or outputs it as a character code to an external device. At this time, the fact that the error notification has been made is displayed as being difficult to read.

Next, the operation of the above embodiment will be described using a flowchart.
FIG. 13 is a flowchart for explaining operations up to selection of a candidate character.
First, the document image input receiving unit 104 receives a document image input from the user, converts it into binary image data (S1302), detects a character line from the document image, and outputs it to the character image extracting unit 105 (S1304). ). The character image extraction unit 105 detects a character image from the input character line (S1306), stores it in the storage area of the morphological feature amount extraction unit 108 and the recognition verification unit 112, and notifies the feature amount extraction unit 106 (S1308). ).

The feature amount extraction unit 106 extracts the feature amount from the character image, stores the feature amount in the similarity calculation unit 111, and notifies the candidate character selection unit 107 (S1312).
The candidate character selection unit 107 calculates the similarity between the notified feature amount and the reference feature amount registered in the classification dictionary 101, and stores the calculated similarity and the candidate character in the recognition verification unit 112. (S1316), N candidate characters are selected in descending order of similarity (S1318). The selection result is notified of the candidate character determination unit 1092 (S1320), and it is determined whether or not the processing of all characters in one line is finished (S1322). If not, the process returns to S1306. It is determined whether or not the processing has been completed for all the lines of the image (S1324). If not completed, the process returns to S1304, and if completed, the process ends.

FIG. 14 is a flowchart for explaining the operation of the candidate character determination unit 109.
When the candidate character determination unit 109 receives the notification of the first to Nth candidate characters for one character image from the candidate character selection unit 107, the candidate character determination unit 109 determines whether the combination of the candidate characters is the same character (S1402). . If it is the same character, the process moves to S1414, and if it is not the same character, it is determined whether or not it is in the similar character combination of the similar candidate character identification dictionary 102 (S1404). If not, the process moves to S1414. If there is, the dictionary data is read (S1406).

  The morphological feature amount of the character image is extracted by the morphological feature amount extraction unit 108 (S1408), and the extracted morphological feature amount and the rule of the dictionary data are collated (S1410) to determine which candidate character is dominant. (S1412). When it is not possible to determine which is dominant, the process proceeds to S1414. When it is determined that any is dominant, the candidate character determined to be dominant is set as the first candidate character (S1416), and the process ends.

In S1414, it is determined whether there is still another combination of candidate characters. If there is, the process returns to S1402, and if not, the process ends.
With the above processing, when two candidate characters having a high degree of similarity are described in the similar candidate character identification dictionary 102 for N candidate characters for one character image selected by the candidate character selection unit 107, Which candidate character matches the character image more is determined based on the form feature amount.

FIG. 15 is a flowchart for explaining the operation of the individual character verification unit.
When the individual character verification unit 110 receives notification of the first to Nth candidate characters for one character image from the candidate character determination unit 109, the individual character verification unit 110 initializes the counter N to “1” (S1502).
Next, the rule of the Nth candidate character from the individual verification dictionary 103 is read (S1504). The morphological feature amount extraction unit 108 extracts the morphological feature amount described in the rule from the character image (S1506). The extracted morphological feature value is compared and verified with the rule of affirmation / negative condition of the individual verification dictionary 103 (S1508). It is determined whether or not the counter N is 1 (S1510). If it is 1, it is determined whether or not a negative condition is satisfied (S1512). If established, the first candidate character is “rejected” (S1514), and the process is terminated.

When the negative condition is not satisfied, it is determined whether the positive condition is satisfied (S1516). When it is established, the first candidate character is set to “valid” (S1518), and the process ends.
If the positive condition is not satisfied, the process moves to S1520. In S1520, 1 is added to the counter N (S1520). It is determined whether there is an Nth candidate character (S1522), and if there is, return to S1504.

If the counter N is not 1 in S1510, it is determined whether or not a negative condition is satisfied (S1524). When it is satisfied, the process proceeds to S1520. When it is not satisfied, it is determined whether an affirmative condition is satisfied (S1526). If established, the process moves to S1514. If not, the process moves to S1520.
If there is no Nth candidate character in S1522, the first candidate character is set to “valid” (S1528), and the process ends.

In this embodiment, when the first candidate character does not satisfy the negative condition or the affirmative condition, when the negative condition or the affirmative condition does not hold for the second or lower candidate character, the first candidate character is Although it is “valid”, it may be “rejected”.
Next, the operation of the recognition verification unit 112 will be described with reference to the flowcharts of FIGS.
First, the recognition verification unit 112 calculates the average character width / average character height for the character image input from the character image extraction unit 105 (S1602), and calculates the character width / character height of the character of interest. (S1604), and the difference from the average is evaluated (S1606). It is determined whether or not the character size exceeds a predetermined threshold (S1608). If the character size is large, the character of interest is rejected (S1610), and the process proceeds to S1616.

If the character size does not exceed the predetermined threshold, it is further determined whether or not the character size is equal to or smaller than the predetermined threshold (S1612). In the following cases, it is determined whether or not the recognition result is a symbol such as “.” Or “•” (S1614). If negative, the process moves to S1610. If yes, the process moves to S1616.
In S1616, it is determined whether or not the processing has been completed for all characters. If completed, the process proceeds to S1702, and if not completed, the process returns to S1604.

  In S1702, the recognition verification unit 112 calculates an average character position for one line of character string. Next, it is determined whether or not the character size is small (S1704). If the character size is small, the displacement of the target character from the average character position is calculated (S1706), and the displacement is evaluated (S1708). It is determined whether or not the displacement is large (S1710). If it is large, the character of interest is rejected (S1712), and the process proceeds to S1714. Even when the displacement is small, the process proceeds to S1714.

When it is determined that the character size is small, position processing is performed (S1718), and it is determined whether or not processing of all characters has been completed (S1714). If not completed, the process returns to S1704, and if completed, the process proceeds to S1802.
In S1802, when there is a character image recognized as the same character as the first candidate character in the same character string, the authentication verification unit 112 uses the character image as the attention character (C) and the character of the attention character (C). The similarity calculation unit 111 calculates the similarity r from the feature amount of the image and another character image. Among these, the candidate character of the character image with the highest similarity is detected and set as a character (C1) (S1804).

  The candidate character having the highest similarity among the candidate characters selected as the same candidate character as the target character (C) is set as the character (C2) (S1806). It is determined whether or not the character (C1) and the character (C2) are the same (S1808). If they are the same, it is determined whether or not the character (C1) has been “rejected” (S1810). If not, the character (C) is set to “valid” (S1812), and the process proceeds to S1820.

  If they are not the same in S1808, it is determined whether or not the character (C1) is “valid” and r · R of the character (C1)> r · R of the character (C2) (S1814). When the result is affirmative, the recognition result (candidate character) of the character (C1) is changed to the first candidate character of the target character (C) (S1816), and the process proceeds to S1820. If no in S1814, the first candidate character of the target character (C) is “rejected” (S1818), and it is determined whether or not the processing for all the characters has been completed (S1820). If completed, the process is terminated, and if remaining, the process returns to S1802.

In this embodiment, although the certainty of recognition of the first candidate character is verified using the similarity r between the character images of the character strings on the same line, the similarity between the character images of the same document image as a whole is verified. The degree of recognition of the first candidate character may be verified using the degree r.
Further, in the present embodiment, when there are a plurality of candidate characters selected as the same candidate character in the character image of the same line, the character image of one of the candidate characters is the attention character, but the attention character is the same The first character and the last character of the entire document image can be used as the attention character in order.

Furthermore, if the recognition result of the candidate character is a mixture of printed characters and handwritten characters, or if the recognition result of the candidate character is a combination of alphanumeric characters, kanji, hiragana, and katakana, the attribute of the recognition result Can be verified.
Further, although the classification dictionary 101 is described as a multi-template dictionary in the present embodiment, it may be a single template dictionary in which one reference feature amount is registered for each character.

In the present embodiment, the candidate character selection unit 107 selects N candidate characters, but sets a predetermined similarity and selects a candidate character having a similarity higher than the predetermined similarity. You may make it do. In this case as well, at least one candidate character is selected.
In this embodiment, the recognition verification unit 112 recognizes and verifies the candidate character determined to be “valid” by the individual character verification unit 110, but the first candidate character selected by the candidate character selection unit 107. Alternatively, the modification process may be performed on another first candidate character using the similarity r between the character images directly.

In the present embodiment, the individual character verification unit 110 receives the determination result of the similar characters from the candidate character determination unit 109 and verifies the individual characters. However, for the candidate characters selected by the candidate character selection unit 107, Directly “valid” or “rejected” may be verified.
(Other variations)
Of course, the present invention is not limited to the above-described embodiments, and the following modifications can be implemented.
(1) Although the character recognition device is shown in the block diagram of FIG. 1, specifically, a computer comprising a microprocessor, ROM, RAM, hard disk unit, display unit, keyboard, mouse, optical reader, and the like. System. A computer program is recorded in the RAM or hard disk unit. This computer program is a program that causes a computer to perform the functions of the respective units in the above configuration diagram. The function of each unit is achieved by the microprocessor operating according to the computer program. Here, the computer program is configured by combining a plurality of instruction codes indicating instructions for the computer in order to achieve a predetermined function.
(2) A part or all of the components constituting the character recognition device may be constituted by one system LSI (Large Scale Integration). The system LSI is an ultra-multifunctional LSI manufactured by integrating a plurality of components on a single chip. Specifically, the system LSI is a computer system including a microprocessor, a ROM, a RAM, and the like. . A computer program is stored in the RAM. The system LSI achieves its functions by the microprocessor operating according to the computer program.
(3) A part or all of the components constituting the character recognition device may be constituted by an IC card that can be attached to and detached from the device or a single module. The IC card or the module is a computer system including a microprocessor, a ROM, a RAM, and the like. The IC card or the module may include the super multifunctional LSI. The IC card or the module achieves its function by the microprocessor operating according to the computer program. This IC card or this module may have tamper resistance.
(4) The present invention may be a character recognition method of a character recognition device as described in its operation. Further, the present invention may be a computer program that realizes this method by a computer, or may be a digital signal composed of the computer program.

  The present invention also provides a computer-readable recording medium such as a flexible disk, hard disk, CD-ROM, MO, DVD, DVD-ROM, DVD-RAM, BD (Blu-ray Disc). ), Recorded in a semiconductor memory or the like. The digital signal may be recorded on these recording media.

In the present invention, the computer program or the digital signal may be transmitted via an electric communication line, a wireless or wired communication line, a network represented by the Internet, a data broadcast, or the like.
The present invention may be a computer system including a microprocessor and a memory, wherein the memory stores the computer program, and the microprocessor operates according to the computer program.

In addition, the program or the digital signal is recorded on the recording medium and transferred, or the program or the digital signal is transferred via the network or the like, and executed by another independent computer system. It is good.
(5) The above embodiment and the above modifications may be combined.

  The character recognition device according to the present invention is used as a character input device of a personal computer equipped with an image reading device, and is used as an input device for various typeface documents and particularly handwritten characters having a large variation in character shape. Used in office work as an input device.

It is a block diagram of one Embodiment of the character recognition apparatus which concerns on this invention. It is a figure which shows the data structure of the classification dictionary of the said embodiment. It is a figure which shows an example of the classification dictionary of the said embodiment. It is a figure which shows an example of the similar candidate character identification dictionary of the said embodiment. It is a figure for demonstrating the content of the similar candidate character identification dictionary of the said embodiment. It is a figure which shows an example of the individual verification dictionary of the said embodiment. It is a figure for demonstrating the content of the separate verification dictionary of the said embodiment. It is a figure which shows an example of the document image received by the document image input reception part of the said embodiment. It is a figure which shows an example of the several candidate character selected in the candidate character selection part of the said embodiment. It is a figure for demonstrating the rejection process in the recognition verification part of the said embodiment. It is a figure which shows an example of the character image notified to the recognition verification part of the said embodiment. It is a figure for demonstrating the processing content in the recognition verification part of the said embodiment. It is a flowchart explaining operation | movement of the document image input reception part of the said embodiment, a character image extraction part, a feature-value extraction part, and a candidate character selection part. It is a flowchart explaining operation | movement of the candidate character determination part of the said embodiment. It is a flowchart explaining operation | movement of the separate character verification part of the said embodiment. It is a flowchart (the 1) explaining operation | movement of the recognition verification part of the said embodiment. It is a flowchart (the 2) explaining operation | movement of the recognition verification part of the said embodiment. It is a flowchart (the 3) explaining operation | movement of the recognition verification part of the said embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Classification dictionary 102 Similar candidate character identification dictionary 103 Individual verification dictionary 104 Document image input reception part 105 Character image extraction part 106 Feature-value extraction part 107 Candidate character selection part 108 Morphological feature-value extraction part 109 Candidate character determination part 110 Individual character verification part 111 Similarity Calculation Unit 112 Recognition Verification Unit 113 Recognition Result Output Unit

Claims (4)

A character recognition device that receives input of a document image and recognizes each character included in the document image,
A classification dictionary in which the reference feature values of each character are registered;
A character image extracting means for extracting a character image from a document image;
Feature amount extraction means for extracting the feature amount from the extracted character image;
A candidate character selection means for calculating a similarity between the extracted feature quantity and the reference feature quantity, and selecting a character having a high similarity as a candidate character;
Similarity calculation means for calculating the similarity between character images in the same document using the feature quantity extracted by the feature quantity extraction means;
A recognition verification means for verifying the certainty of recognition as a candidate character based on the similarity as a candidate character and the similarity between the character images ;
The recognition verification means includes
A calculation unit for calculating a product of the similarity r with the verification target character calculated by the similarity calculation unit and the similarity R calculated by the candidate character selection unit;
A recognition result changing unit that changes the recognition result of the verification target character to the other candidate character when there is another candidate character having a product larger than the product of the same candidate character selected by the candidate character selection unit; character recognition apparatus according to claim Rukoto to have a. An individual verification dictionary that describes, as rules, an affirmative condition indicating morphological feature quantities to be provided as candidate characters for each character and a negative condition indicating morphological feature quantities that should not be provided as candidate characters,
Extracts morphological features from the character image, determines whether the selected candidate character meets an affirmative condition or a negative condition, and rejects the selection of the candidate character when it is determined that the negative condition is met The character recognition apparatus according to claim 1, further comprising individual character verification means. A similar candidate character identification dictionary describing two character feature quantities having high similarity and rules for identifying the two characters;
When candidate characters selected from one character image are registered in the similar candidate character identification dictionary, a morphological feature amount is extracted from the character image, checked against the rule, and any of the two characters is selected as a candidate character. The character recognition device according to claim 1, further comprising candidate character determination means for determining whether or not Wherein the configuration characteristic quantity, character recognition loop position or number average linear density of the predetermined range, the convex end positions or number, claim 2 or 3, wherein it contains at least one of the concave end point position or the number apparatus.

Images