JP3079707B2 - Character recognition method and device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method and apparatus for predicting next character candidates and first performing recognition on them to enable accurate and high-speed character recognition. It is.

[0002]

2. Description of the Related Art In a conventional character recognition method, a character image of an individual character is cut out from a digital image of a character, and various characteristic amounts are extracted from the character image. The character code is determined by comparison. As a method of extracting characters, there are a method of estimating a pitch from a marginal distribution and a method of repeating separation and integration of connected components. As a type of the feature amount, there is a method of paying attention to a direction of a line element, a surrounding shape, a density distribution over the entire pattern, and the like (for example, IEICE Technical Report IE88-6, “Recent Trends in Character Recognition Technology”). Trends "). However, when a language with many types of characters such as Japanese is targeted, the comparison with the feature amount of the target character is performed for all the characters in the feature amount dictionary in terms of processing amount. Was inefficient and problematic. In order to solve this problem, a method of extracting a character candidate by performing a detailed recognition process after performing a large classification using a coarse feature amount is generally performed (for example, “Pattern Recognition”, IEICE). Ed., Pp. 67). However, this method also has a problem because each process takes a lot of time. On the other hand, since high accuracy cannot always be expected only by character recognition based on the feature amount, a method of improving accuracy by performing various post-processing on the result of character recognition has been proposed. For example, for a plurality of candidates obtained by character recognition such as rejected characters,
Method of Word Matching (IEICE Technical Report)
PRL 82-77 "A Study of Word Matching Process in Pattern Recognition"). Furthermore, a method of combining grammatical processing such as morphological analysis and information obtained as a result of recognition processing based on feature amounts has also been proposed (IEICE Transactions Vol. J67-D No. 11 pp. 1348).
"Character post-processing using recognition information and word / phrase information"). There is also a method in which a combination of connectable characters is held as a table, and a connectable character is selected from among a plurality of candidates as a correct answer. It is known that the combination that may be concatenated in the case of kanji is less than 1% of the whole, and if the concatenation processing of three or four characters is performed, the same effect as performing word matching is obtained. (Electronic Communication Society Transactions Vol. J68-D No. 1)
pp. 64 "Determination processing of unreadable character using character connection information-application to character recognition-"). By utilizing this, it is possible to efficiently perform a process of limiting candidates from characters before and after. However, since all of these post-processings are performed as separate processings after the character recognition processing using the feature amount, the processing time becomes longer, which is a problem.

[0003]

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems of the prior art. That is, an object of the present invention is to provide a character recognition method and apparatus for performing character recognition based on feature amounts, in which a target for which feature amounts are compared is narrowed down at high speed and with high accuracy, thereby performing high-speed character recognition without deteriorating accuracy. It is assumed that.

[0004]

According to the character recognition method of the present invention, a first character image and a second character image to be recognized adjacent to each other are recognized.
In recognition of the second character image, the first character image
The second character image is a character connected to the candidate character
Steps to put out as preliminary candidate characters for
(S302), the feature amount of the preliminary candidate character and the second
Preliminary candidate characters are determined based on the distance from the character image features.
Evaluate and evaluate candidate characters for the second character image to be recognized
(S303, S304) and its determination
Based on the relationship between the second character image and the candidate character
Re-determining candidate characters for the first character image
(S306), the first character image determined by the re-determination
Image candidate character and the second image adjacent thereto
Recognition of first character image based on combination with candidate characters
Determining output characters (S308, S701, S
702), only candidate characters connected to the recognized output character are
Redetermined as a new candidate character for the second character image
And step (S309) .
Further, the character recognition device of the present invention, a character image input means for inputting a character image to be read (1 in FIG. 1), Oh
That a second recognition object character first recognized sentence <br/> character before to hold a candidate character as the previous candidate character for the image candidate character holding means adjacent to the image (4), articulated possible characters Union
A connection table that holds the set, based on the connection table,
Characters that connected to the front candidate character for the first recognition target character image and next candidate characters predicting means for predicting a next candidate characters (5-1, 5-2), was predicted by the next candidate character prediction means A candidate character determining means (6) for determining the validity of the character as a next candidate character;
In the set of the next candidate character connected to it, the previous candidate sentence
Between the characteristic amount of the character and the characteristic amount of the first recognition target character image.
A distance, a feature amount of the next candidate character, and the second recognition
Calculate the sum of the distance to the feature of the target character and calculate the sum
The previous candidate character corresponding to the set
Recognition processing means for outputting as recognition output of recognition target characters
(3) and control means (7) for controlling the entire processing.

[0005]

According to the present invention, by predicting a candidate for a target character from a character before (or after) a character to be recognized, the number of characters that must be recognized without using a candidate character is greatly reduced. I have. For the prediction, the property that the number of consecutive combinations of non-Hiragana characters is relatively small is used. That is, by preparing in advance information of a character that may come next (or come before) for each character, a candidate for the next character can be predicted very quickly, and the number of candidates can be reduced. Can be reduced as compared with the large classification processing.

[0006] The preceding candidate character holding means is provided with a first adjacent character.
And a candidate character for the first recognition target character image in the second recognition target character image is held as a previous candidate character. The concatenation table holds information on combinations of characters (concatenated characters) that may come next to each character. The next candidate character prediction unit predicts a preliminary candidate character (next candidate character) for the second recognition target character image based on the connection table. For example, a case in which each character of a character image string “join” is recognized is taken as an example. As illustrated in FIG. 9, it is assumed that “run”, “luck”, and “speed” are stored in the previous candidate character storage unit as candidates for the recognition result of the first recognition target character image “run”. The next candidate character predicting means, as exemplified in FIG.
By using the connection table, characters connected to "speed"
It is obtained as a preliminary next candidate character of the second recognition target character image. For example, the next candidate character “continuation” “connected” to the previous candidate character “sequence”, the next candidate character “use” connected to the previous candidate character “luck”, and the next candidate character “degree” connected to the previous candidate character “speed” "
Etc. are obtained. The candidate character determination unit determines the validity of the previous candidate character held by the previous candidate character holding unit. This makes it possible to narrow down the number of preceding candidate characters. The determination of the validity of the preceding candidate character is performed, for example, based on the magnitude of the feature distance between the second recognition target image character and the preliminary next candidate character. The recognition processing means includes, in a combination of the previous candidate character and the next candidate character determined to be valid, a distance between the feature amount of the previous candidate character and the feature amount of the first recognition target character image, and the distance of the next candidate character. The sum of the distance between the feature amount and the feature amount of the second recognition target character is obtained, and the previous candidate character corresponding to the set having the smallest total is output as the first recognition target character recognition output.

According to the present invention, the next character is predicted using the connection information of a plurality of characters such as a concatenation table which is information used for post-processing, so that the same recognition accuracy is secured without performing post-processing. And post-processing time can be saved. As described above, according to the present invention, character recognition can be performed at high speed without lowering the accuracy by predicting a character candidate to be recognized from a character before the character (or a character after the character). is there.

【Example】

First, the configuration of the embodiment will be described with reference to FIG. This embodiment comprises the following parts. (1) A character image input unit 1 for inputting a character image to be read. This part is realized by an image input device such as a general image scanner or a facsimile. (2) A character cutout unit 2 that cuts out a character image of each character from an image composed of a plurality of characters. (3) A function of outputting a corresponding character code as a candidate only from a feature amount of a character image, and a distance between a specific character code feature amount and a feature amount of a character image to be recognized (hereinafter simply referred to as distance or distance value and A recognition processing unit 3 having a function of calculating (4) A pre-candidate character holding unit 4 for holding a candidate for a character preceding the recognition target character. (5) As means for predicting the next character, a next candidate character prediction unit 5-1 using inter-character connection information and a character connection table 5-2. (6) Judgment of the validity of the result of recognition by the recognition processing unit 3 on the next candidate character obtained by the next candidate character prediction unit 5-1 and re-establishment of the previous candidate / next candidate based on the result A candidate character determination unit 6 for performing determination; (7) A control unit 7 that performs overall processing and data control.

In the above configuration, as shown in FIG. 2, the recognition processing section 3 includes a feature quantity calculating section 31 for calculating the feature quantity of each character image cut out by the character cutout section 2, A distance calculation unit 32 for calculating a distance, and a feature amount-character code match king unit 33 for determining a character code by performing large classification based on the feature amount and matching based on the detailed feature amount.
And The distance here is a distance in an N time space having N feature amounts as axes. Also,
The feature-character code matching unit 33 includes a table-based feature dictionary (recognition dictionary) 332 in which the feature of each character image is associated with a character code, and a large classification based on coarse feature and an address of the feature dictionary 332. And a hash table 333 for associating a character code with an address of a feature dictionary. Note that the recognition processing unit 3 treats a character image input by the character image input unit 1 that is cut out by the character cutout unit 2 and becomes a character image for each character as input.

One of the functions of the recognition processing unit 3 is a general character recognition function for determining a character code by performing a large classification based on feature amounts and matching based on detailed feature amounts with respect to an input character image. . Another function is a function of recognizing a character image using a feature dictionary of a specific designated character code and outputting a distance between the character image and the designated character. In order to realize these two functions, as described above, the large classification table 331 has a configuration capable of double lookup so that the feature dictionary 332 can be looked up from either side.
And a hash table 333. When the character code is subtracted from the feature value, the large classification table 121 is used, and when the feature value is obtained from the character code, the hash table 123 is used, so that a high-speed search can be performed.

Next, the overall processing flow will be described.
FIG. 3 shows the overall processing flow. (1) Predicting next candidate character It is determined whether or not there is a previous candidate character (step S30).
1). (1-1) When there is no candidate character in the previous candidate character holding unit When there is no previous candidate character as a result of the determination in step S301, character recognition using only the feature amount is performed to generate a candidate (step S301). S310), and stores the distance value between these and the recognition target character image in the preceding candidate character storage unit 4 (step S310).
311).

(1-2) When there is one or more candidate characters in the preceding candidate character holding unit If the result of the determination in step S301 is that there are preceding candidate characters, the next candidate character is predicted for each preceding candidate character. (Step S302). In the case of this embodiment, the prediction of the next candidate character uses the connection relation between the characters. FIG. 4 is a flowchart showing a processing procedure of the next candidate character prediction using the connection relation. When a certain character is observed, a character that may be connected to the character is stored as a connected character table as shown in FIG. 7, and for each character held in the previous candidate character holding unit 4, Using the character concatenation table, all characters that may be concatenated are set as next candidate characters.

That is, in FIG. 4, first, one preceding candidate character is extracted (step S401), and the concatenation table (FIG. 9) is searched using the preceding candidate character as a key (step S402). All connected characters obtained as a result of the search are output as next candidates (step S403). It is determined whether all the previous candidates have been seen (step S40).
4) If not, the next preceding candidate character is extracted (step S405), and the processing of steps S402 to S404 is repeated. When the processing has been completed for all of the preceding candidate characters, the next candidate character prediction processing in FIG. 4 ends.

(2) Recognition Processing Using Feature Amounts The recognition processing unit 3 uses the function of a feature amount distance calculation unit 32 for calculating the distance between the feature amount of a specific character code and the feature amount of a character image. The distance between each next candidate character predicted by the candidate character prediction and the recognition target character is calculated (step S303). The distance value between the next candidate character thus obtained and the recognition target of each next candidate character is sent to the candidate character determination unit 6.

(3) Next candidate character determination In the candidate character determination unit 6, only those whose distance between the characteristic amount of the character image and the characteristic amount of the recognition dictionary is equal to or less than a certain threshold value are set as the next candidate character, and the others are set as the next candidate character. It is deleted from the character (step S304). FIG. 5 is a flowchart showing details of the processing in step S304. In FIG. 5, first, one next candidate character is extracted (step S501). Then, it is determined whether the distance is equal to or less than the threshold (step S).
502), if it is less than the threshold, it is deleted from the next candidate (step S503). It is determined whether all the next candidate characters have been seen (step S504). If not, the next next candidate character is extracted (step S505), and the processing of steps S502 to S504 is repeated. When the processing has been completed for all of the next candidate characters, the candidate character determination processing of FIG. 5 ends.

(4) Determination of Candidate Character Next, it is determined whether or not the next candidate character remains as a result of the next candidate character determination (step S305). (4-1) In the case where no next candidate character remains as a result of the next candidate character determination, character recognition by prediction is unsuccessful. Then, the one with the smallest distance value is output as the recognition result of the preceding candidate character (step S307). And for the characters to be recognized,
Character recognition using only general features is performed (step S
310), the obtained candidate and the distance value are stored in the previous candidate character holding unit 4
(Step S311).

(4-2) When one or more next candidate characters remain as a result of the next candidate character determination First, the previous candidate character is re-determined from the remaining next candidate characters (step S306). Specifically, step S306
Consists of the procedure shown in the flowchart of FIG. That is, the previous candidate character having no character connected to the next candidate character group remaining by the next candidate character determination is deleted from the previous candidate. That is, in FIG. 6, first, one preceding candidate character is extracted (step S601), and it is checked whether or not a character connected to the preceding candidate character is in the next candidate character (step S60).
2). If not, the previous candidate character is deleted from the list of previous candidate characters (step S603). If it is in the next candidate character, it is not deleted. Next, it is determined whether all the previous candidates have been viewed (step S604). If not, the next previous candidate character is extracted (step S605), and the processes of steps S602 to S604 are repeated. When the processing has been completed for all of the preceding candidate characters, the preceding candidate re-determination processing of FIG. 6 ends.

Next, the previous candidate is determined and output (step S308). FIG. 7 is a flowchart showing details of the procedure for determining and outputting the previous candidate. That is, for each pair of the remaining previous candidate character and the next candidate character connected to the remaining candidate character, the sum of the distance value of the previous candidate character and the distance value of the next candidate character is calculated for each pair. (Step S701). Then, the set of the preceding candidate characters having the smallest value is output as the recognition result of the preceding candidate character (step S702). Further, only the next candidate character connected to the output preceding candidate character is stored as a new preceding candidate character (step S309) in the preceding candidate character holding unit (step S311). As a result, the next candidate character that is not connected to the output preceding candidate character is deleted from the candidates.

Step S309 comprises the procedure shown in the flowchart of FIG. That is, in FIG. 8, first, one next candidate character is extracted (step S80).
1) It is checked whether the character is connected to the output next candidate character (step S802). If not, the next candidate character is deleted from the next candidate character list (step S80).
3). If they are connected, they are not deleted. Next, it is determined whether all the next candidates have been viewed (step S80).
4) If not, the next next candidate character is extracted (step S805), and the processing of steps S802 to 804 is repeated. When the processing has been completed for all of the next candidate characters, the previous candidate re-determination processing of FIG. 8 ends.

(5) Final character processing When there is no character image to be recognized next to the character to be recognized, the character is called a final character. Step S31 in FIG.
In step 2, it is determined whether the character is the last character. (5-1) When the recognition target character is the last character Among the remaining candidates, the one with the smallest distance value is output as the recognition result. (5-2) When the Character to be Recognized is Not the Last Character Repeat from the processing of (1).

Next, a specific example will be described. FIG.
Is a diagram for explaining the operation of this specific example. Consider recognizing an image of a character string “join”. Here, as for the word "ren", "ren, luck,
It is assumed that the distance value between the candidate “speed” and each character with respect to the recognition target character “sequence” is known, and the table (A) of these characters and the distance values is stored in the previous candidate character holding unit. It is assumed that processing starts from this state.

First, since there is a candidate character in the previous candidate character holding unit 4, the next candidate character is predicted. A connection table as shown in FIG. 9 is prepared for each next candidate character, and the next candidate character is predicted with reference to these tables. In the case of this embodiment, the connection table does not include hiragana. This is because there are many combinations of hiragana and non-hiragana characters that are extremely connected, and if hiragana is included, many hiragana are included in the predicted candidates. As a result of the prediction, all the characters listed in these connection tables are the next candidate characters.

Next, the recognition processing unit 3 performs recognition of all the next candidate characters obtained in this manner by the feature amount, and obtains the distance between each candidate character and the recognition target character "tangent". As a result, a table (B) having a distance to the next candidate character, the preceding candidate character connected to the next candidate character, and the recognition target character “tangent” of each character is created. This table (B) is sent to the candidate character determination section 6. The candidate character determining unit 6 determines each distance value, and deletes only the characters that are equal to or less than a certain threshold value and deletes other characters. As a result, only “contact, carrying, carrying” remains as shown in the table (C).

On the basis of the result, the preceding candidate character re-determination processing is performed. That is, in this case, since there is no character connected to the character "speed" among the characters remaining in the next candidate character, this character is deleted from the previous candidate character. As a result, only the two characters “ream and luck” are left as previous candidate characters.
Then, a set of characters connected to these two characters, that is, “concatenation”
The distance value of each character is summed up for each of "cooperation" and "transportation", and is set as the distance of the character set. As a result, a table (E) having a set of connected characters and a total distance value of each set is created. From these, select the one with the smallest total distance value, and replace the character of the part that is the previous candidate of that pair with
It is output as a recognition result for "ren". In this case, since the distance value of “joint” is the smallest, the character “joint” in this group is selected and output. Then, all the previous candidate characters are deleted from the previous candidate character holding unit 4. Next, the two characters of the character "", which are connected to the output character "", are stored in the previous candidate character storage unit 4. At this time, the character "luck" connected to the character "luck" which has not been output is excluded from the candidates. As a result, the preceding candidate character holding unit 4 stores two characters of "contact, carrying" as shown in the table (F).

Next, similarly, when the next candidate character is predicted for these two characters, the characters in the concatenation table are obtained as the next candidate characters as shown in Table (G). In this case, since the connection table does not include hiragana, the correct answer "su" does not enter the candidates. Recognition of these candidates by the feature amount and determination of candidate characters indicate that there is no character that satisfies the threshold at a distance from “su”. Therefore, all characters are deleted from the next candidate character, and the next candidate character prediction is performed. Fail. Therefore, character recognition is performed using the characteristic amount of a general character image, and as a result, a candidate such as the table (I) is obtained for “su”. Since the determination of the next candidate character has failed, the next candidate character cannot be used to determine the character in the previous candidate character holding unit 4. Therefore, the character “contact”, which is a character having a small distance value, is simply output, and the preceding candidate character holding unit 4 is emptied. Then, "su," obtained by character recognition using the feature amount
“Seng, gen” is stored in the previous candidate character holding unit 4. Further, in the same manner, the next candidate character is predicted using “S, Dimension, Talent”. In this case, since the hiragana is excluded from the connection table, the next candidate character for "su" cannot be predicted. Therefore, the next candidate character prediction is performed only for “dimension”, but if the same processing as above is performed on the candidate obtained as a result, all the next candidate characters satisfy the condition. Otherwise, the next candidate character prediction fails. Therefore, character recognition using the feature amount is performed.

By repeating such processing, character recognition processing using the next candidate character prediction is performed until there is no more character image to be recognized.

In the present embodiment described above, the next character is predicted by the next candidate character predicting unit 5-1 with reference to the preceding candidate character holding unit 4, thereby reducing the number of next character candidates. And perform character recognition at high speed. Further, since the same effect as performing post-processing can be expected by predicting the next candidate character, a high recognition accuracy can be obtained without performing post-processing, and time required for post-processing can be saved.

In the embodiment, the recognition processing unit 3 performs the recognition using the characteristic amount. However, the recognition processing unit 3 may be configured to perform the recognition without using the following characteristic amount. For example, a recognition method using a neural network that receives an image can be adopted. This is not to take the feature amount of the character image, but to train the image itself in the neural network and input the character image to be recognized, and output the corresponding character. Further, template matching and the like, which are recognized by superimposing character images, can also be cited as examples of a recognition method that does not use a feature amount. In this method, a template of a character image itself is prepared as a dictionary, and the characters represented by the template that best matches each other are compared for each pixel as a recognition result of the character.

As the next candidate character prediction method, the following method can be used in addition to the next candidate character prediction method of the above embodiment. (1) A process using a concatenation table of three or more characters is used together. For example, when a three-character concatenation table is used, the previous candidate holding unit holds two characters before, and predicts the next candidate character using the concatenation table having two-character entries. For the character predicted in this manner, the evaluation value is set higher than that of the character predicted from the two-character concatenation table, so that the character can easily remain as the next candidate. Although the processing amount increases,
Candidates with high possibility can be left, and accurate prediction can be made.

(2) When hiragana is not included in the concatenation table, an error may occur if a kanji or katakana whose shape is similar to hiragana is predicted as the next candidate character. In response to this, a set of similar non-hiragana characters and hiragana characters is stored in advance as a table, and if such a character is predicted, an error-prone hiragana character may be added to the next candidate character. Can be By doing so, it is possible to prevent to some extent an error in which a kanji similar to Hiragana is regarded as the correct answer. For example, it is assumed that the character "red" is included in the preceding candidate character in the character string image "red also". And "hair" on the connection table of "red"
Is output in the next candidate character. Next, when the distance between the character "hair" and the character image of the character "mo" is calculated, the two characters are similar characters and thus meet the threshold, and the character "hair" becomes the next candidate character. As a result, character recognition using the feature amount is not performed. As a result, the character "mo" is not recognized and the character "hair" is erroneously output. Therefore, a similar character table as shown in FIG. 11 is used to prevent this. In other words, for characters such as "hair" that are easily mistaken for hiragana, the character "mo" is stored in the similar character table as a similar character, and if the character "hair" is predicted, Is added to the next candidate character (FIG. 12). In this case, "mo" is also the next candidate character in addition to "hair". As a result, since the recognition distance of "mo" is short, "mo" which is the correct answer is output.

(3) When the hiragana is not included in the connection table, a method of adding all the hiragana to the next candidate together with the character predicted by the connection table may be considered. As a result, the number of candidates to be predicted increases by the number of all hiragana characters (several tens), but it is still smaller than the number obtained as a result of performing a large classification based on coarse feature amounts. In addition, the size of the table can be reduced as compared with the case where the hiragana is included in the concatenation table. By adopting such a method, the hiragana can be included in the next candidate, so that the hiragana is not mistakenly recognized as a non-hiragana character as in (2), and the accuracy can be improved.

(4) As in (3), not all hiragana characters are uniformly added to the next candidate, but it is also conceivable to add hiragana characters, symbols and kanji which are particularly frequently used to the candidates. Of the hiragana, if the most frequently used characters are added to the top half, it is possible to predict a significant part of the hiragana. In addition, by adding frequently used symbols and kanji uniformly and not adding infrequently used hiragana, the number of candidate characters to be added can be reduced, and the prediction ability can be improved. it can.

(5) The connection character table may have each connection probability instead of simply indicating whether or not connection is performed. Then, by weighting the distance value based on the feature value with the connection probability, it is possible to reduce the evaluation value for the connection that rarely occurs, and reduce errors in the connection table.

(6) Application to languages other than Japanese (Chinese, Korean, etc.) is also possible. As a target language, a language having a large number of characters and a small number of consecutive combinations of individual characters is effective.

As a candidate character determining method, the following method can be used in addition to the candidate character determining method of the above embodiment. (1) The determination of the preceding candidate character is performed not by the total value of the distance between the consecutive candidate character and the next candidate character but by the distance value of only the preceding candidate character. This means that the previous candidate character is determined immediately when the candidate appears. Therefore,
There is only one previous candidate character, and no previous candidate re-determination or next candidate re-determination is performed. As a result, although the accuracy is reduced, the processing is simplified and the processing time is shortened.

(2) A normalized distance value is used instead of a simple distance value as a selection criterion for a set of a preceding candidate character and a next candidate character. As a specific normalization method, it is conceivable to divide the distance value for other candidate characters by the distance value of the first candidate based on the smallest distance among the candidate character groups. By doing so, it is possible to eliminate erroneous evaluation due to variation in the recognition distance for each character.

(3) In addition to a method of uniformly determining the threshold value of the recognition distance used for determining the next candidate character, a method of specifying a different value for each character is also conceivable. It is conceivable that the threshold value of a character that is likely to be erroneous with many similar characters is set to be strict, and the threshold value of a character that is small is set to be loose. For example, there is a method of determining the distance value based on the distance to the character having the smallest distance from the target character. According to this method, the possibility that a similar character is erroneously determined to be correct is reduced, and accurate determination can be realized.

(4) All the previous candidate characters until the next candidate character prediction fails are left, and at the time the next candidate character prediction fails, all combinations of the previous candidate characters that can be connected are used. Select the one with the highest evaluation value from.
This makes it possible to evaluate a wider range of concatenated character strings, and is expected to improve the accuracy.

(5) In the embodiment, prediction is made from the front of the sentence,
Although recognition was performed, it is also possible to perform recognition from the back of the sentence. In this case, a connection table predicted from the rear is required. Prediction from the rear to the front may be more restrictive in some cases, and is effective in narrowing down the candidates.

[0040]

According to the present invention, by predicting a character candidate to be recognized from a character in front of a character to be recognized, the number of characters that need to be recognized using the feature amount is greatly reduced. Therefore, character recognition can be performed at high speed without lowering the accuracy.

Further, according to the present invention, as a method for predicting, the fact that the number of contiguous combinations of non-hiragana characters is relatively small is utilized, and for each character, Since the information is prepared as a concatenation table between characters, it is possible to predict the next character candidate at a very high speed and to reduce the number of candidates as compared with the conventional large classification processing.

Further, since the next character is predicted by using a connection table or the like which is information used for post-processing, the same recognition accuracy can be secured without performing post-processing and post-processing time can be saved. be able to.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of an embodiment of the present invention;

FIG. 2 is a diagram illustrating a configuration of a recognition processing unit;

FIG. 3 is a diagram showing a flow of overall processing in the embodiment;

FIG. 4 is a diagram showing a flow of processing of next candidate character prediction;

FIG. 5 is a diagram showing a flow of a candidate character determination process;

FIG. 6 is a diagram showing a flow of a process of re-determining a previous candidate;

FIG. 7 is a diagram showing a flow of processing of final candidate determination and output,

FIG. 8 is a diagram showing a flow of a process of re-determining a previous candidate;

FIG. 9 is a diagram showing an example of a character connection table;

FIG. 10 is a diagram for explaining the operation of the embodiment using a specific example;

FIG. 11 is a diagram showing an example of a similar character table.

FIG. 12 is a diagram illustrating an example of adding a similar character.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Character image input part, 2 ... Character extraction part, 3 ... Recognition processing part, 31 ... Feature amount calculation part, 32 ... Feature amount distance calculation part, 33 ... Feature amount-character code matching part, 4 ... Previous candidate character Holding unit, 5-1: next candidate character prediction unit, 5-2: connection table, candidate character determination unit.

Claims (2)

(57) [Claims] 1. A first character image to be recognized which is adjacent to each other.
In the recognition of the image and the second character image, the candidate characters for the first character image are connected
Characters as preliminary candidate characters for the second character image
And keep out step, wherein the characteristic amount and the second character image of the preliminary candidate character
Evaluate and recognize preliminary candidate characters based on their distance from the volume
Step of determining candidate characters for the target second character image
Between the step and the candidate character for the determined second character image
Redetermines candidate characters for the first character image based on
Steps and candidate characters of the first character image determined by the re-determination
A set of candidate characters for the second image adjacent thereto
The recognition output character for the first character image based on the
And outputting only candidate characters connected to the recognized output character to a second character
Re-determining as a new candidate character for the image
Character recognition method characterized by comprising and. 2. A character image input means for inputting a character image to be read, and a candidate character for a first recognition target character image adjacent to a certain second recognition target character image as a preceding candidate character. before candidate character holding means holds, articulated table for holding a combination of articulating possible characters, based on the concatenated table, said first recognition target character articulating the prior candidate character for the image next candidate characters and next candidate characters predicting means for predicting as, connecting the next candidate characters and the candidate characters determination means for determining the validity of the next climate <br/> auxiliary character predicted characters by the prediction means, and it the front candidate character Next candidate character set
The feature amount of the preceding candidate character and the first recognition target character
Distance to image feature and feature of next candidate character
And the total distance between the feature amount of the second recognition target character and
The previous candidate character corresponding to the pair having the smallest sum
Sure to be output as the recognition output of the first recognition target character
A character recognition device comprising: recognition processing means; and control means for controlling overall processing.