JPH08305805A - Text recognition device - Google Patents

Abstract

PURPOSE: To efficiently and accurately narrow down candidates in character recognition processing and character segmentation processing. CONSTITUTION: This text recognition device is equipped with a single character recognizing means 201 which performs the character recognition processing for an image of one character and outputs plural character codes as a character recognition result of one character in the decreasing order of accuracy and a character predicting means 202 which predicts a character that appears next from the character recognition result obtained by the single character recognition means 201 up to a certain point of time and outputs its character code. When there are 41 character code predicted by the character predicting means 202, the character recognition result is outputted after being corrected so that the respective predicted character codes increase in order.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image containing a plurality of characters arranged according to a certain rule or a plurality of characters arranged with a certain statistical property, for example, a sentence such as Japanese or English, or a formatted data string. The present invention relates to a text recognition device that sequentially recognizes each character from the inside and outputs a sequence of a plurality of character codes.

Further, by using the single character recognition means as the online character recognition means, it is possible to
It can also be applied to an online text recognition device that sequentially recognizes sentences such as English or formatted data strings and outputs a sequence of a plurality of character codes.

Further, by using the single-character recognizing means as the phoneme recognizing means, continuously uttered sentences such as Japanese and English,
Alternatively, it can be applied to a continuous voice recognition device that sequentially recognizes a formalized data string and outputs a sequence of a plurality of character codes.

[0004]

2. Description of the Related Art In a conventional text recognition device or character recognition device, when it is attempted to improve the recognition performance by utilizing the property of the arrangement of target characters, for example, the grammatical rules and statistical properties of Japanese text, It is common to perform character recognition processing on a plurality of characters and then perform verification, correction, etc. on the recognition results.
In such a technique, after the character recognition processing for a plurality of characters, such as in a text area, is completed, processing using the character arrangement is performed.

FIG. 1 is a diagram showing an operation flow of text recognition according to a conventional technique.

(S101) Each character is recognized to obtain a recognition result string.

(S102) The recognition result sequence is grammatically verified and corrected.

(S103) The contents of the recognition result sequence are output.

For example, Japanese Examined Patent Publication No. 6-32091 discloses a method of recognizing characters forming a Japanese sentence by using a word dictionary and grammatical connection information for a plurality of recognition candidate character strings. Techniques for detecting and reading results are shown. However, in such a technique, it is difficult to correct the correct character if it is not included in the recognition candidates or if the character is cut out incorrectly.

In order to deal with these cases, it is necessary to extremely increase the number of recognition candidates or to recognize a plurality of character cutting paths that may be correct and save the result. Both the amount and the storage capacity increase.

Further, Japanese Patent Laid-Open No. 6-20105 discloses that
A technique for predicting the next character by using a predictive LR (Left to Right) parser and determining the existence probability of the predicted character by a character collating unit is shown.

However, in this technique, if misidentification occurs in the character collating unit, that is, if the required existence probability is incorrect, all subsequent predictions will be inaccurate and recognition will be difficult. There is. Even if a word that is not included in the word dictionary is being recognized,
After all, all subsequent predictions will be inaccurate.

Further, the case where there are two or more character area candidates that are presumed to be an image of one character is not taken into consideration. For example, in order to recognize general print text, text handwritten at free pitch, etc. Is not practical.

[0014]

[Patent Document 1] Japanese Patent Publication No. 6-320
In the technology typified by Japanese Patent Publication No. 91, since the character arrangement of the target character is used as a test and a correction for the recognition result after a series of character recognition processing, it is used during the character recognition processing. Can not. Therefore, it cannot be used to narrow down the candidates in the character recognition processing or the character cutout processing at the time of the processing.

Therefore, there is a problem in that the character arrangement of the target character cannot be effectively used, or a large number of candidates in the character recognition processing and the character cutting processing must be stored.

Further, in the technique disclosed in Japanese Patent Laid-Open No. 6-20105, a character predicting unit called a predictive LR parser unit is provided to utilize the character arrangement of target characters at the time of performing character recognition processing. However, if the character recognition process or the prediction process is inaccurate at a certain point, it becomes difficult to process subsequent characters. The reason for this is that unpredicted characters are not targeted for character recognition processing, and that characters with a low existence probability obtained by character recognition processing are not used as the basis for prediction of subsequent characters.

Therefore, the present invention is to enable efficient and accurate narrowing down of candidates in the character recognition processing and the character cutout processing.

[0018]

SUMMARY OF THE INVENTION The present invention is a text recognition device for sequentially recognizing each character from an image containing a character string and outputting a sequence of character codes, and for a character image. A first single-character recognizing unit that performs character recognition processing and obtains a character recognition result including a character code corresponding to the image for one character and a certainty that is the certainty of the character code;
A character predicting unit that predicts a character code that appears next from a sequence of character recognition results obtained by the first single character recognizing unit up to a certain time point, and a character code obtained by the first single character recognizing unit And a second single character recognizing means for ordering the character codes based on the accuracy and the character code predicted by the character predicting means and obtaining a character recognition result for the one character image. And

Further, the present invention is a text recognition device for sequentially recognizing each character from an image including a character string and outputting a sequence of a plurality of ordered character codes, which is an image area for one character. First character recognition means for calculating a distance or similarity to the standard pattern of the character code, and obtaining a character recognition result including the character code corresponding to the image area for one character and the distance or similarity, A character prediction unit that predicts a character code that appears next from a sequence of character recognition results obtained by the first single character recognition unit up to a certain time point; and the character code that corresponds to the character code predicted by the character prediction unit. A predetermined arithmetic operation is performed on the distance or the similarity calculated by the first single character recognizing means, the character codes are ordered based on the result of the arithmetic operation, and the character recognizing result for the one character image is obtained. And having a Mel second single character recognition means.

Further, the present invention is a text recognition device for sequentially recognizing each character from an image containing a character string and outputting a sequence of a plurality of ordered character codes, and for a character image. And a character recognition result obtained by the single character recognition means up to a certain point of time, the first character recognition means obtaining a character recognition result including a character code corresponding to the image for one character Character predicting means for predicting a character code that appears next from the character string, the character code obtained by the first single character recognizing means, and the character code predicted by the character predicting means.
A second single-character recognizing unit that performs character recognition processing on an image for a character and obtains a character code corresponding to the image for one character by ordering the character codes based on the certainty that is the likelihood of the character code. It is characterized by having.

Further, the present invention is a text recognition device for sequentially recognizing each character from an image including a character string and outputting a sequence of character codes, wherein a character area candidate for one character is cut out from the image. A character recognizing process is performed on the character area candidate for one character cut out by the slicing means and the character slicing means, and character recognition including a character code for the character area candidate for the one character and a certainty indicating its certainty. A first single character recognizing means for obtaining a result;
Score calculating means for calculating a character area score representing the accuracy of the character area candidate for one character based on the character code and the accuracy obtained by the single character recognizing means, and the first single character recognizing by a certain point of time. A character predicting unit that predicts a character code that will appear next from a sequence of character recognition results obtained by the unit; a character code obtained by the first single character recognizing unit and its accuracy; and a score calculating unit that calculates the character code. Determining a character region from the character region candidates based on the determined region score and the character code predicted by the character prediction means, and ordering the character codes to obtain a character recognition result corresponding to the character region. It is characterized by having two single character recognition means.

The first single-character recognizing means and the second single-character recognizing means have the same structure and are realized by one piece of hardware or one piece of software.

[0023]

The text recognition device according to the present invention is provided with the character prediction means, predicts the next character appearing at the time of recognizing each character of the text, and performs the recognition processing and the character segmentation by the single character recognition means. It is possible to efficiently and accurately narrow down the candidates in the character recognition processing and the character cutout processing by using the character area to determine the character area.

Further, since the single character recognizing means also targets the character which is not predicted by the character predicting means, the processing for the subsequent characters is accurately performed even if the character recognizing process or the predicting process is inaccurate. You can continue to.

Further, the first single-character recognizing means and the second single-character recognizing means have the same structure and are realized by one piece of hardware or one piece of software, thereby simplifying the structure.

[0026]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The features of the present invention will be specifically described below based on embodiments with reference to the drawings. Since the first single-character recognizing means and the second single-character recognizing means described in the claims have the same structure and are realized by one piece of hardware or one piece of software, the structure is simplified. The embodiment shows such an example. However, the present invention is not limited to this configuration, and the first single character recognition means and the second single character recognition means may be provided independently.

FIG. 2 is a diagram showing an example of the configuration of an embodiment of the text recognition device according to the present invention. Reference numeral 201 is a single character recognition means, 202 is a character prediction means, 203 is a memory, 20
Reference numeral 4 is a recognition result string stored in the memory 203, and 205 is a predicted character set stored in the memory 203.

The recognition results obtained by the single character recognizing means 201 are sequentially added to the recognition result sequence 204 and stored.
Further, the characters predicted by the character predicting means 202 are stored as a predicted character set 205. Further, the single character recognizing means 201 uses the predicted character set 2 in the course of the processing.
Refer to 05. The character prediction unit 202 refers to the recognition result string 204 in the course of the processing.

An existing technique can be used for the character predicting means 202. For example, as in the technique disclosed in Japanese Patent Laid-Open No. 6-20105, LR for predicting the next character.
It is possible to prepare a table in advance, refer to the LR table from the recognition result sequence obtained so far, and obtain the character that can appear next. Further in the present invention,
Since the recognition result string up to that point can be referenced, if parsing fails due to the occurrence of unregistered words or misrecognition, reset the stack and restart parsing from the required position, Can be predicted.

As another technique for realizing the character predicting means, N-gram statistics shown in, for example, Proceedings of the 1994 Spring Conference of the Institute of Electronics, Information and Communication Engineers, 7-223, "Japanese Document Recognition Using Double Markov Model". The data can be used. When N-gram statistical data is used, the probability that each character will appear next can be calculated from the (N-1) characters at the end of the recognition result sequence.

FIG. 3 is a diagram showing a schematic operation flow of an embodiment of the text recognition device according to the present invention.

(S301) The recognition result sequence 204 and the predicted character set 205 are cleared.

(S302) With reference to the predicted character set 205, one character in the text to be recognized is recognized by the single character recognition means 2
Recognize by 01.

(S303) The obtained recognition result is added to the recognition result column 204.

(S304) With reference to the recognition result sequence 204, the character predicting means 202 predicts a character that may appear next.

(S305) The obtained predicted character is written in the predicted character set 205.

(S306) It is checked whether all the characters in the text to be recognized have been recognized. If all the recognitions have not been completed, the process is repeated from (S302). When all the recognition is completed, the process proceeds to (S307).

(S307) The contents of the recognition result sequence 204 are output.

The configuration and operation of the above-mentioned single character recognizing means 201 will be described below in detail.

FIG. 4 is a diagram showing an example of the configuration of the single character recognition means 201 in the text recognition device. Reference numeral 401 is a feature extracting means, 402 is a distance calculating means, 403 is a standard feature storing means, 404 is a distance value storing means, and 405 is a distance value sorting means.

Various features can be used as the features extracted by the feature extracting means 401. Eg mesh features,
Peripheral features (see Hashimoto: “Introduction to Character Recognition”, The Telecommunications Association, March 20, 1982, pp. 83-85), direction line element features, etc. can be used. Moreover, you may use a some characteristic together.

Various distance calculation methods can be used as the distance calculation means 402. For example, Euclidean distance, Mahalanobis distance, pseudo Bayes function, etc. can be used.

In the standard feature storage means 403, features extracted from previously collected character images by the same method as the feature extraction means 401 are stored in association with character numbers. The character number is stored in association with each character code, and the character code is uniquely determined by referring to the character number.

The distance value storage means 404 stores the distance value calculated by the distance calculation means 402 in association with the character number.

The distance value sorting means 405 sorts the distance values stored in the distance value storage means 404 together with the character numbers, and outputs the character numbers corresponding to the plurality of sorted higher distance values as the character recognition result. To do. Further, not only the character number but also the distance value itself may be output so that the character predicting means 202 or the like can refer to the value.

FIG. 5 is a diagram showing an operation flow of the single character recognizing means 201 in the embodiment of the text recognizing device according to the invention.

(S501) The feature extraction means 401 extracts a feature vector for recognition from a character image of one character. This feature vector is called X.

(S502) 0 is assigned to the integer variable i representing the character number being processed.

(S503) Feature vector X and standard feature F
The distance calculation means 402 calculates the distance to i, and the value is stored in the distance value storage means 404. Let this value be Di. The standard feature Fi is a feature vector stored in the standard feature storage means 403 in association with the character number i.

(S504) It is checked whether the character corresponding to the number i is included in the predicted character set 205. If it is included, (S505) is executed. If not included, the process proceeds to (S506).

(S505) The distance value Di is multiplied by a predetermined constant 0.9, and the value is stored in the distance value storage means 404.

(S506) 1 is added to the variable i.

(S507) It is checked whether the variable i is smaller than the total number of characters I. If it is smaller, the process is repeated from (S503). If it is not smaller, the process proceeds to (S508).

(S508) The distance value sorting means 405 sorts the distance values Di stored in the distance value storage means 404 in ascending order, that is, in ascending order of the value, and outputs the obtained upper character number.

The loop from (S503) to (S507) described above is a loop for performing the same processing on the character corresponding to the number i. For example, a plurality of distance calculating means 402 are prepared and the processing is parallelized. As a result, the processing time can be shortened.

By the process of (S505), the characters predicted by the character prediction unit 202 and included in the character prediction set 205 are changed to have a smaller distance value from the standard feature. Constant 0.9 multiplied by Di in the above (S505)
May have other values. It is effective to set the constant to a value smaller than 1. The optimum value changes depending on the accuracy of the character predicting means 202. It is also possible to change the value of the constant each time according to the number of characters included in the character prediction set 205. Furthermore, it is also effective to change the value of the constant according to the character type of the character corresponding to the number i. For example,
0.9 if the character corresponding to the number i is hiragana
5, 0.93 for katakana and 0.9 for kanji.

In the above (S505), a constant may be added instead of multiplying the distance value Di by a constant. In this case, it is effective to make the constant smaller than 0, that is, negative.

Next, the effect of the present invention will be described in comparison with the prior art.

FIG. 6 is a diagram showing a state of recognition in a conventional text recognition device. 601 is a part of the text image. 602 is the content of the recognition result sequence,
Corresponding to each character, the first, second, and third recognition candidates are shown in order from the top.

In this example, noise is added to the character corresponding to "su" in the text "as a business", and the correct answer "su" is not included in the recognition candidates up to the third place. , The correct answer cannot be obtained even by the grammatical test and correction after the recognition of each character is completed.

As a method for making the case like this example a correct answer by improving the conventional technique, the number of candidates to be stored as a recognition result sequence is increased, and characters that are not included in the recognition candidates are taken into consideration in the grammatical verification and correction. Although there is a possibility to put it in, there are drawbacks such as an increase in storage capacity, an increase in calculation processing amount, and an increase in false positives due to side effects.

FIG. 7 is a diagram showing a state of recognition in the embodiment of the text recognition device according to the present invention.

702, 704, 706, 708, 710
Is the contents of the recognition result sequence 204. 703, 705, 7
07, 709, and 711 are the contents of the predicted character set 205.

Reference numeral 702 indicates the contents of the recognition result sequence 204 at the time when the first character "Bi" is recognized by the single character recognition means 201, and reference numeral 703 indicates the character prediction set 205 obtained by the character prediction means 202 with reference to 702. Content, 704 is 703
Is the content of the recognition result sequence 204 at the time when the second character “ji” is recognized by the single character recognition means 201 with reference to, and so on. Recognition result string 70 that has recognized up to the third character
Since the character prediction set 707 “S” is obtained from 6, the distance value Di corresponding to the character “S” is changed and becomes the first candidate.

As described above, according to the present invention, it is possible to make a correct answer a character that cannot be determined as a correct answer by using a grammatical rule in the prior art. Further, even when "su" is not the first candidate but the second or third candidate, the grammatical test after completion of character recognition similar to the prior art in the present invention,
It is also possible to make it the first-ranked candidate by using the modification together.

Further, in the text recognition apparatus according to the present invention, the predicted character set 2 obtained by the character prediction means 202 is used.
At 05, a prediction score is assigned to each character in the predicted character set 205. The prediction score is a value indicating the degree to which each character is predicted, and is set to a larger value, for example, as the character is strongly predicted to appear next. When the above-described N-gram statistical data is used as the character prediction unit 202, the statistical probability that each character will appear next can be obtained, so that value can be used as the prediction score as it is.

In the text recognition device according to the present invention which uses the predictive score, a function of the predictive score is used instead of using the constant in the process of FIG. 5 (S505). For example, when the prediction score is p,

[Equation 1] To Di.

If the function of the prediction score is used for the correction of the distance value, the correction degree of the distance value can be adjusted according to the degree of prediction of each character. That is, the grammatical property can be more accurately reflected in the character recognition result by changing the distance value to a smaller value so that it is strongly predicted that the character will appear next.

With the configuration and operation described with reference to FIGS. 4 to 7, the characters predicted by the character prediction unit 202 and included in the character prediction set 205 have a smaller distance value Di with respect to the standard feature Fi. Since it is changed, there is an effect that it is positioned higher in the character recognition result. Furthermore, the degree to which the predicted character moves to the upper position can be adjusted by changing the value of the constant.

FIG. 8 is a diagram showing another example of the configuration of the single character recognizing means 201 in the embodiment of the text recognizing device according to the present invention. In the figure, 801 is a first feature extraction means, 802 is a first distance calculation means, 803 is a first standard feature storage means, 804 is a first distance value storage means, 805.
Is a first distance value sorting means, 806 is an intermediate candidate adding means, 807 is a second feature extracting means, 808 is a second distance calculating means, 809 is a second standard feature storing means, and 810 is a second distance. A value storage unit 811 is a second distance value sorting unit.

The first feature extracting means 801 and the second feature extracting means 807 extract the same or different features from the character image. When the same features are used, the configuration can be simplified by implementing the two means with one hardware or one software module. In this case, the configuration can be simplified by further implementing the first standard feature storage means 803 and the second standard feature storage means 809 as one hardware or one memory area.

It is highly accurate to make the features used by the first feature extraction means 801 extractable in a relatively short time and to make the features used by the second feature extraction means 807 relatively accurate. This is effective for obtaining a high recognition result at high speed. An example of an effective combination is that the first feature extraction unit 801 uses mesh features, and the second feature extraction unit 807 uses mesh features, peripheral features, and direction line element features together.

The distance calculation methods used by the first distance calculation means 802 and the second distance calculation means 808 are the same or different distance calculation methods. When the same distance calculation method is used, the configuration can be simplified by implementing the two means with one hardware or one software module.

The distance calculation method used by the first distance calculation means 802 can be calculated in a relatively short time, and the distance calculation method used by the second distance calculation means 808 can be made relatively accurate. However, it is effective for obtaining a highly accurate recognition result at high speed. As an example of valid combinations,
The first distance calculating means 802 may use the Euclidean distance, and the second distance calculating means 808 may use the pseudo Bayes function.

First distance value storage means 804 and second distance calculation means 808, first distance value sorting means 805 and second distance value storage means 805.
The distance value sorting means 811 can be configured with one hardware or one software module to simplify the configuration.

The first distance value sorting means 805 sorts the distance values stored in the first distance value storage means 804 together with the character numbers, and obtains the character numbers corresponding to the plurality of sorted upper distance values. , As an intermediate candidate.

FIG. 9 is a diagram showing an operation flow of the single character recognizing means in the text recognition device shown in FIG.

(S901) The first feature extraction means 80 obtains the first feature vector for recognition from the character image of one character.
Extract by 1. This feature vector is called X.

(S902) 0 is assigned to the integer variable i representing the character number being processed.

(S903) The distance between X and the first standard feature Fi is calculated by the first distance calculation means 802, and the value is stored in the first distance value storage means 804. This value is D
i. The first standard feature Fi is a feature vector stored in the first standard feature storage means 803 in association with the character number i.

(S904) 1 is added to the variable i.

(S905) It is checked whether the variable i is smaller than the total number of characters I. If it is smaller, the processing from (S903) is repeated. If it is not smaller, the process proceeds to (S906).

(S906) First distance value sorting means 80
5, the distance values Di stored in the first distance value storage unit 804 are sorted in ascending order, that is, in ascending order of the values, and the obtained upper N characters are used as intermediate candidates.

(S907) If there is a character that is not included in the intermediate candidates and is included in the predicted character set 205, the intermediate candidate adding means 806 adds it to the intermediate candidates.

(S908) The number of intermediate candidates is substituted into the integer variable J.

(S909) The second feature vector for recognizing from the character image of one character is extracted by the second feature extracting means 80.
Extract according to 7. This feature vector is called X '.

(S910) 0 is assigned to the integer variable j representing the character number being processed.

(S911) The character number of the jth intermediate candidate is assigned to the integer variable i.

(S912) X'and the second standard feature F'i
The distance to and is calculated by the second distance calculation means 808, and the value is stored in the second distance value storage means 810. This value is Dj. The second standard feature F′i is a feature vector stored in the second standard feature storage unit 809 in association with the character number i.

(S913) 1 is added to the variable j.

(S914) It is checked whether the variable j is smaller than the number J of intermediate candidates. If it is smaller, the process is repeated from (S911). If it is not smaller, the process proceeds to (S915).

(S915) Second distance value sorting means 81
According to 1, the distance values Dj stored in the second distance value storage means 810 are sorted in ascending order, that is, in ascending order of the values, and the obtained upper character number is output.

For example, 10 is used as the value of N used in the process of (S906). The value of N may be in the range of equal to or smaller than the total number of characters I and equal to or larger than the number of candidates to be finally output. The larger the value of N, the higher the accuracy of the final recognition candidate, but the amount of calculation tends to increase.

The process of (S907) is a feature of the present invention. According to the present invention, even when the target character image has noise or deformation and the recognition result is falsely recognized as a result in which the correct character does not remain in the intermediate candidate in the conventional technique, the effect of the intermediate candidate adding unit 806 is achieved. Thus, if a correct character is included in the predicted character set 205, it will be included in the intermediate candidate. Compared with the first feature extracting means 801 and the first distance calculating means 802, the second feature extracting means 807 and the second distance calculating means 808 perform high-precision processing and therefore are affected by noise and deformation. Since it is difficult, the probability that the correct character added by the intermediate candidate adding means 806 is higher in the recognition result is high, and the accuracy of the recognition result can be increased accordingly.

Although FIG. 8 and FIG. 9 show an example having two candidate extraction stages, the present invention can be applied even if there are two or more candidate extraction stages. By providing the intermediate candidate adding means for each stage, it is possible to prevent the correct character from being missing from the intermediate candidates at that stage.

As described above, due to the configuration and operation described with reference to FIGS. 8 and 9, in the single character recognition process having two or more candidate extraction stages, the candidate correct answer character is missing in the middle. By doing so, most cases where misidentification occurs can be remedied, and the accuracy of final text recognition can be improved.

FIG. 10 is a diagram showing still another example of the configuration of the embodiment of the text recognition apparatus according to the present invention. In the figure, 1001 is a single character recognizing unit, 1002 is a character predicting unit, 1006 is a character cutting unit, 1003 is a memory,
1004 is a recognition result sequence stored in the memory 1003,
Reference numeral 1005 is a predictive character set stored in the memory 1003, 1007 is a character area candidate stored in the memory 1003, and 1008 is a temporary recognition result stored in the memory 1003.

The recognition result obtained by the single character recognizing means 1001 is stored in the temporary recognition result 1008. The characters predicted by the character prediction unit 1002 are stored as a predicted character set 1005. Character cutting means 100
6 is the character area candidate 100.
Stored in 7. The recognition result corresponding to the character area determined by the character cutting means 1006 is sequentially added to the recognition result sequence 1004 and stored.

The single character recognizing means 1001 refers to the character area candidate 1007 in the process of the processing. The character prediction unit 1002 recognizes the recognition result string 1 in the process of the processing.
Please refer to 004. The character cutout unit 1006 refers to the predicted character set 1005 and the temporary recognition result 1008 in the process of the processing.

FIG. 11 is a diagram showing an operation flow of the text recognition device shown in FIG.

(S1101) The recognition result sequence 1004 and the predicted character set 1005 are cleared.

(S1102) The character area candidate 1007 and the temporary recognition result 1008 are cleared.

(S1103) Character cutting means 1005
Thus, the character area candidate that is estimated to be one character in the text to be recognized is extracted and stored in the character area candidate 1007.

(S1104) The single character recognition means 1001 recognizes the image corresponding to one of the character area candidates.

(S1105) The obtained recognition result is added to the temporary recognition result 1008.

(S1106) It is checked whether all the characters of the character area candidate have been recognized. If all the recognitions have not been completed, the process is repeated from (S1104). If all have been recognized, the process proceeds to (S1107).

(S1107) Character cutting means 1006
Thus, the character area score corresponding to each character area candidate is calculated with reference to the temporary recognition result 1008. If any of the characters included in the temporary recognition result corresponding to each character region is included in the predicted character set 1005, the prediction score is multiplied by a constant 2. The character area candidate having the largest prediction score is determined as the character area.

(S1108) The recognition result corresponding to the confirmed character area is changed from the temporary recognition result 1008 to the recognition result sequence 1
Add to 004.

(S1109) The character predicting means 1002 predicts a character that may appear next.

(S1110) The obtained predicted character is written in the predicted character set 1005.

(S1111) It is checked whether all the characters in the text to be recognized have been recognized. When all the recognition is not completed, the process is repeated from (S1102). When all recognition is completed, the process proceeds to (S1112).

(S1112) The contents of the recognition result column 1004 are output.

In the above operation flow, the above (S110
The constant 2 by which the prediction score is multiplied in 7) may be another value. It is effective to set the constant to a value greater than 1. The optimum value changes depending on the accuracy of the character prediction unit 1002.
It is also possible to change the value of the constant each time according to the number of characters included in the predicted character set 1005. Furthermore, it is also effective to change the value of the constant according to the character type of the character commonly included in the temporary recognition result 1008 and the predicted character set 1005. For example, the temporary recognition result 100
8 and the character commonly included in the predicted character set 1005 is 1.5 when the character is hiragana, and 1. when the character is katakana.
7. If it is Kanji, it is 2. Furthermore, it is also effective to change the value of the constant in accordance with the rank of the characters included in both the temporary recognition result 1008 and the predicted character set 1005 in the recognition candidates. For example, the temporary recognition result 1008
If the first-ranked candidate is included in the predicted character set 1005, it is set to 2, if the second-ranked candidate is included, 1.3, and if the third-ranked candidate is included, 1.1 is set.

In the above embodiment, the condition for multiplying the constant is the case where any of the characters included in the temporary recognition result 1008 is included in the predicted character set 1005, but other conditions may be used. For example, the first rank candidate of the temporary recognition result 1008 may be included in the predicted character set 1005.

FIG. 12 is a diagram showing an example of the effect of the text recognition device shown in FIG.

In the figure, 1201 is a part of a text image. 1202 and 1209 are defined character areas. 1203 and 1210 are the contents of the recognition result sequence 1004. 1204 is the content of the character prediction set 1005.
1205 and 1206 are character area candidates 1007. 1
207 and 1208 are the contents of the temporary recognition result 1008.

Reference numeral 1202 shows the state of the character area at the time when the part 1201 of the text image is recognized up to "Vect", and the confirmed character area is shown by a black rectangle.
At this point, the contents of the recognition result sequence 1004 are as shown in 1203. Referring to 1203, character prediction means 1002
The content of the predicted character set 1005 obtained by
It becomes as shown in.

When the vertical projection is seen, the next character "Lu" has two parts, and character region candidates 1205 and 12
06 is obtained. In the figure, the character area candidates 1007 are shown by gray rectangles.

Single character recognition means 1 for each character area candidate
Temporary recognition result 100 obtained by performing recognition at 001
The contents of 8 are 1207 and 1208. The left columns are the first, second, and third candidates in order from the top, and the numerical values on the right show the character region scores. The character area score is a value indicating how likely the character area is to be a character, and for example, the reciprocal of the distance of the first candidate in the character recognition process is used.

The reciprocal of the distance of the first rank candidate is "0.172" for the character area candidate 1205, and the character area candidate 12
It is "0.155" for 06.

Since the character segmenting means 1006 determines the character region candidate having the largest character region score as the character region, when the value of the reciprocal of the distance of the first place candidate is used as it is, the character region candidate 1205 becomes the character region. Will be determined as, causing an error in character cutting.

However, in the present invention, the character area candidate 1206
No. 1 candidate of the temporary recognition result 1208 corresponding to
Is included in the predicted character set 1204, the value of the reciprocal of the distance corresponding to the character area candidate 1206 is “0.155”.
The value "0.310" obtained by multiplying by is a character area score. On the other hand, the temporary recognition result 1207 corresponding to the character area candidate 1205 includes the predicted character set 120
Since there is no character included in 4, the character area score is “0.172” as it is.

As a result of such an operation, the character area candidate 12
The character area score “0.310” corresponding to 06 becomes the maximum, the character area candidate 1206 is determined as the character area, the character area becomes like 1209, and the correct character cutout result is obtained.

Character area candidate 12 confirmed as a character area
The contents of the temporary recognition result 1207 corresponding to 05 are written in the recognition result column 1004, and the contents of the recognition result column 1004 are as shown in 1210.

As described above, the embodiment shown in FIG. 10 corresponds to the character predicted by the character prediction means and included in the character prediction set by the configuration and operation described with reference to FIGS. 10 to 12. Since the character area candidate is changed to have a larger character area score, it has the effect of being easily determined as a character area. Furthermore, the degree to which the predicted character is likely to be fixed can be adjusted by changing the value of the constant.

The above-described embodiments can be combined with each other, and by using them in combination, the character recognition accuracy can be further improved.

[0127]

As described above, according to the present invention,
When there is one or more character codes predicted by the character predicting means, the character recognition result is corrected so that each predicted character code is in the higher order, and then output. It is possible to efficiently and accurately narrow down the candidates in the character cutting process.

Even if the character recognition process or the prediction process is inaccurate, the process for the subsequent characters can be continued accurately.

[Brief description of drawings]

FIG. 1 is a diagram showing an operation flow of text recognition according to a conventional technique.

FIG. 2 is a diagram showing an example of a configuration of an embodiment of a text recognition device according to the present invention.

FIG. 3 is a diagram showing a schematic operation flow of an embodiment of a text recognition device according to the present invention.

FIG. 4 is a single character recognition means 2 in the text recognition device.
It is a figure which shows an example of a structure of 01.

FIG. 5 is a diagram showing an operation flow of a single character recognition means 201 in an embodiment of a text recognition device according to the present invention.

FIG. 6 is a diagram showing a state of recognition in a conventional text recognition device.

FIG. 7 is a diagram showing a state of recognition in the embodiment of the text recognition device according to the present invention.

FIG. 8 is a diagram showing another example of the configuration of the single character recognition means in the embodiment of the text recognition device according to the present invention.

9 is a diagram showing an operation flow of a single character recognition means in the text recognition device shown in FIG.

FIG. 10 is a diagram showing still another example of the configuration of the embodiment of the text recognition device according to the present invention.

11 is a diagram showing an operation flow of the text recognition device shown in FIG.

FIG. 12 is a diagram showing an example of effects of the text recognition device shown in FIG.

[Explanation of symbols]

201 ... Single character recognition means, 202 ... Character prediction means, 20
3 ... Memory, 204 ... Recognition result sequence, 205 ... Predicted character set, 401 ... Feature extraction means, 402 ... Distance calculation means, 4
03 ... Standard feature storage means, 404 ... Distance value storage means, 4
05 ... Distance value sorting means, 601 ... Part of text image, 602 ... Contents of recognition result sequence, 702, 704, 70
6, 708, 710 ... Contents of recognition result sequence, 703, 70
5, 707, 709, 711 ... Predicted character set contents, 8
01 ... first feature extraction means, 802 ... first distance calculation means, 803 ... first standard feature storage means, 804 ... first distance value storage means, 805 ... first distance value sorting means, 8
06 ... Intermediate candidate adding means, 807 ... Second feature extracting means, 808 ... Second distance calculating means, 809 ... Second standard feature storing means, 810 ... Second distance value storing means, 811
... second distance value sorting means, 1001 ... single character recognition means, 1002 ... character prediction means, 1003 ... memory, 10
04 ... Recognition result sequence, 1005 ... Predicted character set, 1006
... character cutting means, 1007 ... character area candidates, 100
8 ... Temporary recognition result, 1201 ... Part of text image,
1202, 1209 ... Confirmed character area, 1203, 1
210 ... Contents of recognition result sequence, 1204 ... Contents of character prediction set, 1205, 1206 ... Character region candidates, 1207,
1208 ... Contents of temporary recognition result

Claims (6)

[Claims] 1. A text recognition device for sequentially recognizing each character from an image containing a character string and outputting a sequence of character codes, wherein character recognition processing is performed on an image for one character, A first single-character recognizing means for obtaining a character recognition result including a character code corresponding to an image of characters and a certainty that is the certainty of the character code; and Character predicting means for predicting a character code that appears next from the sequence of the character recognition results, the character code obtained by the first single character recognizing means and its accuracy, and the character code predicted by the character predicting means. And a second single character recognizing means for determining the character recognition result for the image of one character based on the above. 2. A text recognition device for sequentially recognizing each character from an image including a character string and outputting a sequence of a plurality of ordered character codes, wherein the character code is for an image area of one character. A first single character recognition means for calculating a distance or similarity with respect to the standard pattern of, and obtaining a character recognition result including the character code corresponding to the image area for one character and the distance or the similarity; A character prediction unit that predicts a character code that will appear next from a sequence of character recognition results obtained by the first single character recognition unit; and the first character that corresponds to the character code predicted by the character prediction unit. A second operation is performed in which a predetermined operation is performed on the distance or the similarity calculated by the character recognition means, the character codes are sequenced based on the operation result, and the character recognition result for the one character image is obtained. A text recognition device comprising a character recognition means. 3. A text recognition device for sequentially recognizing each character from an image including a character string and outputting a sequence of a plurality of ordered character codes, the character recognizing process for an image of one character. And a first single character recognition means for obtaining a character recognition result including a character code corresponding to the image for one character, and a sequence of character recognition results obtained by the single character recognition means up to a certain time. A character predicting means for predicting a character code appearing in, a character code obtained by the first single character recognizing means, and a character code predicted by the character predicting means based on the image for one character. And a second single character recognizing means for performing character recognition processing for the character code, and ordering and determining character codes corresponding to the image for one character based on the certainty which is the certainty of the character code. A text recognition device for collection. 4. A text recognition device for sequentially recognizing each character from an image including a character string and outputting a sequence of character codes, and a character slicing means for slicing a character region candidate for one character from the image. , Character recognition processing is performed on the character area candidate for one character cut out by the character cutting means, and a character recognition result including a character code for the character area candidate for one character and a probability indicating the certainty is obtained. A first single-character recognizing means, and a score calculating means for calculating a character area score indicating the accuracy of the character area candidate for the one character based on the character code and the accuracy obtained by the first single-character recognizing means. A character predicting unit that predicts a character code that appears next from a sequence of character recognition results obtained by the single character recognizing unit up to a certain time point; and the first single character recognizing unit. A character code determined by the means and its accuracy, based on the area score calculated by the score calculation means and the character code predicted by the character prediction means, while determining the character area from the character area candidates And a second single-character recognizing unit that orders character codes and obtains a character recognition result corresponding to the character area. 5. The accuracy is a value corresponding to a distance between a feature extracted from a character image and a standard feature, and the accuracy is any one of claims 1, 3, and 4. The text recognition device according to paragraph. 6. The first single-character recognizing means and the second single-character recognizing means have the same configuration and are realized by one piece of hardware or one piece of software. The text recognition device according to claim 5.