JP2000090203A - Method and device for recognizing character - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect a character having the possibility of erroneous recognition by finding the reliability of the recognized result while using the relation of character form feature amounts between the concerned character and the other character. SOLUTION: The same character type block extracting part 6 classifies inputted character images into the groups of characters with the same character types and for each block, a character type feature amount extracting part 7 extracts the character type feature amount. Concerning the combination of all the characters in the same character type block, a similarity calculating part 8 calculates the similarity of character type feature amounts and a similarity discriminating part 9 compares the similarity found by the similarity calculating part 8 with a predetermined threshold value and stores the result in a similarity information recording part 10. Concerning each character in the group of characters judged as the same character type, an erroneous recognition detecting part 11 calculates the score while using specified information provided from the relation between the similarity of character types of characters and the recognized result. Thus, the high-reliability score can be defined while considering the relation with the other character and a reject character can be determined.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for calculating the reliability of a character recognition result by comparing a characteristic amount extracted from a character image with a characteristic amount of another character, and using the method.
The present invention relates to an erroneous recognition correction method for correcting a character that may be erroneously recognized.

[0002]

2. Description of the Related Art In character recognition processing, a feature amount is usually extracted from a character image cut out in character units, and the character image is extracted based on a position in a feature amount space in which boundaries between categories are formed in advance by learning. Determine the category.

For example, in the case of handwritten character recognition using a statistical method, a character image filled in by a number of scribes is prepared in advance for each character category, and the feature amount extracted from the character image is determined. , An inter-category boundary is formed so that the categories can be best separated as a whole.

In the vicinity of the category boundary formed in this way, the following erroneous recognition often occurs.

[0005] Since the variation of handwritten characters between different writers is very large, in the training data, a character of a certain category 1 written by a certain writer A is replaced by a character of another category 2 written by another writer B. May be very similar.

Naturally, even in the feature quantity space, feature quantities of two categories are mixed near the boundary between category 1 and category 2.

In such a case, at the time of learning, an inter-category boundary is formed so that the entire character written by a large number of writers is optimal. In other words, categories cannot always be separated at all local boundaries.

Therefore, at the time of recognition, erroneous recognition occurs near the boundary between categories 1 and 2. Regarding such a problem, JP-A-8-50635 and JP-A-10-63785 use the personality of writing to detect erroneous recognition near the boundary between categories.
Attempts have been made to fix it.

The method described in Japanese Patent Application Laid-Open No. 8-50635 is based on the idea that when the character shape of a misread character and the character shape of a correct character are compared, there is some unnaturalness in the misread character.

[0010] In handwritten characters by the same writer,
It is assumed that there are two characteristics, "characters belonging to the same category have the same glyphs" and "characters in different categories have a correlation". The vector is calculated, and the "unnaturalness" of the target character reflecting the above two characteristics is compared with the character shape vector of the target character and the character shape vector of another character.
And use it to detect misreading.

More specifically, the in-category detection of calculating unnaturalness using the relationship between the character of interest and the character recognized in the same category, and the relationship between the character of interest and the character recognized in a different category from the character of interest. Between categories, which calculates unnaturalness using the above.

In the embodiment of this patent publication, an unnaturalness QW as shown in (Equation 1) is defined for detection in a category.

[0013]

(Equation 1)

Further, regarding the detection between categories, (Equation 2)
Is defined as the unnaturalness QB.

[0015]

(Equation 2)

The method disclosed in Japanese Patent Application Laid-Open No. 10-63785 is
It is based on a survey result that "even the same writer writes multiple glyphs in the same character category."

First, clustering of feature vectors used in the recognition process is performed for each category of the character recognition result. Then, a cluster having a small number of elements is extracted as a cluster composed of misread characters. The cluster closest to the misread cluster is found, and when the distance between the clusters is equal to or smaller than the threshold, the character category of the element in the cluster having a small number of elements is corrected to the character category to which the integration partner cluster belongs.

[0018]

SUMMARY OF THE INVENTION However, Japanese Patent Application Laid-Open No. Hei 8-5063
In the method described in Japanese Patent Application Laid-Open No. 5-205, in order for the unnaturalness QW to work well, the number of correct characters in the characters recognized in the category a must be sufficiently large with respect to the number of misrecognized characters. Further, in order to obtain QB, an enormous amount of character data having all character categories for each writer in advance is required.

The method described in Japanese Patent Application Laid-Open No. 10-63785 is a necessary condition that "there is a sufficient number of characters for each category in a document image to be recognized".
In other words, such as a transfer slip, account and money
This method cannot be said to be effective because even for handwriting by the same person, there is not always a sufficient number of items per category.

In addition to the above two methods, when a character sample is collected for each writer in advance and a recognition dictionary is created based on the collected character samples, or when a feature of an input character is extracted, a method of writing is used. Some corrections are tailored to the writer, but collecting samples for each writer requires a great deal of trouble, and it is conceivable that the writer's writing characteristics will gradually change over time. So impractical.

According to the present invention, a writing characteristic in a document image can be effectively used even for a document having a small number of digits, such as a handwritten transfer slip, without requiring a sample collection for each writer in advance. It is another object of the present invention to provide a reliability calculation method and an erroneous recognition correction method.

[0022]

Means for Solving the Problems In order to solve the above problems, the present invention employs the following constitution.

According to a first aspect of the present invention, there is provided a character recognition apparatus which receives a character image in a document image including a plurality of character images as input and outputs at least one or more of all categories to be recognized as a recognition result. Performing, extracting a character-shaped feature value obtained by digitizing the shape of the character image, and for any character, the similarity between the character-shaped feature value of the character and the character-shaped feature value of another character; Calculating the reliability indicating the reliability of the character recognition result based on whether the character recognition result and the character recognition result are the same or different.

According to a seventh aspect of the present invention, for a given character, the relationship between the character and another character is the number of characters whose similarity is similar and whose recognition results are different from each other.
The similarity is dissimilar, and the number of characters whose recognition results are equal to each other, and the number of characters whose similarity is similar, and the recognition results are equal to each other, are determined based on the number. Calculate the reliability.

According to an eighth aspect of the present invention, when an arbitrary character has the same recognition result between the character and another character, the similarity between the character shape characteristic amount of the character and the character shape characteristic amount of another character is obtained. When the recognition result between the character and another character is different, a reliability S2 according to the similarity between the character shape amount of the character and the character shape amount of the other character is provided. .

According to a thirteenth aspect of the present invention, for a character whose reliability is smaller than a predetermined threshold, the recognition result is replaced with a correction category which is a category different from the recognition result, and the reliability of a plurality of characters is reduced. Is recalculated, and when the recalculated reliability is larger than the original reliability, the recognition result is determined to be the correction category.

[0027]

Next, embodiments of the present invention will be described with reference to the drawings.

(Embodiment 1) FIG. 1 is a configuration diagram of a character recognition device according to Embodiment 1 of the present invention.

In FIG. 1, an image input unit 1, a preprocessing unit 2, a character image storage unit 3, a feature extraction unit 4, a recognition unit 5, an identical font block extraction unit 6, a character shape feature extraction unit 7, and a similarity calculation unit 8 , A similarity determination unit 9, a similarity information storage unit 10, and an erroneous recognition detection unit 11 constitute a character recognition device.

Next, the operation of such a character recognition device will be described in detail. In this character recognition device, a feature amount is extracted in advance from a character image, which is a large number of learning data, by a feature extraction unit 4, and the recognition unit 5 associates (learns) the feature amount with a category from the feature amount. deep.

At the time of recognition, a document image in which a plurality of characters are printed or handwritten is input by the image input unit 1. FIG. 2 shows an example of the input document (transfer slip). In the pre-processing unit 2, after the noise removal and the frame line removal, a plurality of characters to be recognized are cut out into images in units of one character, and the cut-out character images are stored in the character image storage unit 3.

The feature extracting unit 4 extracts a feature used at the time of recognition from the character image. The recognizing unit 5 performs recognition using the feature amount, and outputs at least one or more (N = 1 in the present embodiment) recognition candidate categories.

In the same font block extraction unit 6, simple information (in this embodiment, character width,
Based on the character height, the input character image is classified into a group of characters having the same font (the same font block). FIG. 3 shows an example of the result of obtaining the same character block for the input character image. Here, the font of the character includes whether the character is a print type or a handwritten character, and if it is a print type, includes the distinction of the print type. The reason for classifying into the same font block is to detect characters that may be erroneously recognized due to the relationship between the similarity of characters and the recognition category at the time of erroneous recognition detection. If the characters are different even if they are print characters, even if they are characters of the same category, the characters are not similar. Because they do not go. In the present embodiment, it is assumed that the character width and height are almost constant if the character is a print character, and that the character width and the height are determined to be print when the dispersion of the height is equal to or less than a predetermined threshold. I have. At this time, it is assumed that handwritten characters written in one document are all entered by the same writer.

For each block that is determined to be the same font block by the same font block extraction unit 6, the character shape feature amount is extracted by the character shape feature extraction unit 7 for all the characters included in the block.

The similarity calculation unit 8 calculates the similarity between the character-shaped feature amounts for all combinations of characters in the same font block. For example, assuming that M characters are included in the same font block, M · (M−1) / 2 similarities, which are all combinations of M characters, are calculated. In the present embodiment, a similarity such as (Equation 3) is adopted as the similarity.

[0036]

(Equation 3)

Next, the similarity judging section 9 judges "similarity" if the similarity calculated by the similarity calculating section 8 is larger than a predetermined threshold T1, and judges "dissimilar" if it is smaller than the predetermined threshold T2. A determination process is performed, and the result is stored in the similar information storage unit 10.

In addition, the erroneous recognition detecting section 11 performs the following processing on each character in the group of characters determined to have the same font. Number of characters determined to be similar and having the same character category as the recognition result B. Number of characters that are determined to be similar and have different character categories in the recognition result Count the number of characters in the same group that are judged to be dissimilar and meet the conditions for the number of characters with the same character category in the recognition result. Here, as shown in FIG.
Let us consider a case in which the result of recognizing a handwritten handwriting of the string of numbers by the same writer is [006120]. (The third character is erroneously recognized as 6 for a correct answer of 0).
FIG. 5 shows an example in which the number of characters that meet the conditions B and C is counted.

When the result of FIG. 5 is obtained by the similarity determination section 9, the two characters out of the first, second, third and sixth characters are determined to be similar to each other, and the fourth The character and the first, second, third, and sixth characters are determined to be dissimilar to each other. Other combinations are combinations that are not determined to be similar or dissimilar. Here, assuming that A is a positive score and B and C are negative scores, the one with the lowest total score is the third character (the score is -3), and the fourth and fifth characters are the scores. The 0th, 1st, 2nd, and 6th characters are in the order of score 1. Therefore, based on this score, a character that is likely to be erroneously recognized can be detected (rejected) (the third character is actually erroneously recognized).

As described above, by calculating the score using the information A, B, and C obtained from the relationship between the degree of similarity between character glyphs and the result of recognition, reliability based on the relationship with other characters is considered. Highly likely scores can be defined. In addition, by determining the rejected character with reference to the score, it is possible to detect a character having a possibility of erroneous recognition with high accuracy.

If the characters to be recognized in the input document are all printed in the same typeface only, or
When all are handwritten by the same writer, the same font block extraction unit 6 can be omitted.

In the present embodiment, the character-shaped feature amount is different from the characteristic amount used in the recognition unit in the character-shaped feature extraction unit, but may be the same as the characteristic amount used in the recognition unit. Alternatively, the dimension of the feature amount used in the recognition unit may be reduced using principal component analysis or the like.

Although the similarity calculation unit calculates the similarity as shown in (Equation 3) between the character-shaped feature quantities, another similarity may be used. Alternatively, another distance scale such as a Euclidean distance, a city distance, or a Mahalanobis distance between feature values may be used.

In the case where the similarity calculation unit obtains a distance instead of a similarity, the similarity determination unit uses a predetermined threshold T1.
If it is smaller than this, it is determined to be “similar” and a predetermined threshold T
If it is larger than two, it may be determined that "dissimilar".

The score may be calculated by weighting and adding the numerical values of A, B, and C, or by using a cutout score obtained by the preprocessing unit (character cutout unit) and a recognition score obtained by the recognition unit. Add weights and scores
It may be used as an overall score.

(Embodiment 2) FIG. 6 is a configuration diagram of a character recognition device according to Embodiment 2 of the present invention. Embodiment 1 is the same as Embodiment 1 except that there is no similarity determination unit 9 in Embodiment 1.
This is the same configuration as.

Next, the operation of such a character recognition device will be described in detail. The operation up to the similarity calculation unit 8 is the same as that of the first embodiment.

That is, for each character block determined to have the same font by the same font block extraction unit 6, the character shape feature extraction unit 7 extracts the character shape feature amount for all the characters included in the block, and the similarity The calculation unit 8 calculates the similarity between all the characters included in the block according to the equation (Equation 3). FIG. 7 shows the similarities obtained for the six character images included in the same font block as shown in FIG. This similarity is stored in the similarity information storage unit 10.

The erroneous recognition detection unit 11 calculates the following score for each character in the group of characters determined to have the same font.

For all the characters having the same recognition result as the target character, the score is calculated by a function as shown in FIG. 8 (representing the relationship between the similarity and the score), and the average Sa is obtained. For all the characters having different recognition results, the score is calculated by a function as shown in FIG. 9 (representing the relationship between the similarity and the score), and the average Sb is calculated. The sum S of the score Sa and the score Sb is calculated. For example, the score S of the first character, which is the score of the target character, is such that the characters having the same recognition result are the second and sixth characters, and the characters having different recognition results are the third, fourth and fifth characters. Because it is a letter
It looks like below.

Sa = {(200 × 0.92 × 0.92−100) + (200 × 0.91 × 0.91−100)} / 2 = 67.45 Sb = [{− 400 × (0.94−0.5) × (0.94−0.5)} + 0 + {-400 x (0.62-0.5) x (0.62-0.5)] / 3 = -23.73 ... S = Sa + Sb = 67.45 + (-43.737) = 39.72 ... Thus, for each character When the score is obtained, FIG.
(However, the decimal places are rounded off).

Thus, based on the score, a character that is likely to be erroneously recognized can be detected (rejected) (the third character is actually erroneously recognized).

As described above, by calculating the score reflecting the degree of similarity between the character shapes, it is possible to define a highly reliable score in consideration of the relationship with other characters. In addition, by determining the rejected character with reference to the score, it is possible to detect a character having a possibility of erroneous recognition with high accuracy.

If the characters to be recognized in the input document are all printed in the same typeface only, or
When all are handwritten by the same writer, the same font block extraction unit 6 can be omitted.

In the present embodiment, the character-shaped feature amount is different from the characteristic amount used in the recognition unit in the character-shaped feature extraction unit, but may be the same as the characteristic amount used in the recognition unit. Alternatively, the dimension of the feature amount used in the recognition unit may be reduced using principal component analysis or the like.

Although the similarity calculation unit calculates the similarity as shown in (Equation 3) between the character-shaped feature quantities, another similarity may be used. Alternatively, another distance scale such as a Euclidean distance, a city distance, or a Mahalanobis distance between feature values may be used.

In the case where the similarity calculator calculates the distance instead of the similarity, the erroneous recognition detector may prepare the function of the distance and the score instead of the function of the similarity and the score. Further, the functions of the similarity and the score need not necessarily be those of FIGS. 8 and 9 and may be other appropriate functions as long as the following conditions are satisfied.

If the recognition result is the same character, a monotonically increasing function that gives a small score if the degree of similarity is small, and a large score if the degree of similarity is high. A monotonically decreasing function that gives a large score if the score is small and a small score if the degree of similarity is large. The score calculated in the present embodiment is calculated by the cutout score obtained by the preprocessing unit (character cutout unit) and the recognition unit. The obtained recognition score or the like may be weighted and added to be used as an overall score.

(Embodiment 3) FIG. 11 is a configuration diagram of a character recognition apparatus according to Embodiment 3 of the present invention. The configuration is the same as that of the first embodiment except that the erroneous recognition detection unit 11 is replaced with the erroneous recognition correction unit 12.

Next, the operation of such a character recognition device will be described in detail. In the operation up to the similarity information storage unit 10, the recognition unit 5 performs the recognition using the feature amount extracted by the feature extraction unit 4. This is the same as the first embodiment, except that a plurality of (N = 3 in this embodiment) recognition candidate categories are output.

The erroneous recognition correction unit 12 performs the following operations on each character in a character block determined to have the same font, as in the operation of the erroneous recognition detection unit 11 of the first embodiment. Number of characters determined to be similar and having the same character category as the recognition result B. Number of characters that are determined to be similar and have different character categories in the recognition result Count the number of characters in the same group that are judged to be dissimilar and meet the conditions for the number of characters with the same character category in the recognition result. In the case of the input image and the recognition result as shown in FIG.
FIG. 13 shows an example in which B and C are counted. Here, FIG.
2 is a result of showing a sequence of numbers [000120] handwritten and written by the same writer together with the outputs of the first to third candidates of the recognition unit, and the third character is erroneously recognized as 6. All others are first candidates and are correct. FIG. 13 shows the result of the recognition of the first candidate of each character as A, B,
The number of characters that meet the condition of C is counted.
Incidentally, the result of the determination by the similarity determination unit is that any two of the first, second, third and sixth characters are determined to be similar to each other, and the fourth character and the first, second, third and sixth characters The characters are determined to be dissimilar to each other. Other combinations are combinations that are not determined to be similar or dissimilar.
The number of characters that meet the respective conditions of A, B, and C is counted. When A is considered as a positive score and B and C are considered as negative scores, the one with the lowest total score is the third one. Characters (score is -3), the fourth and fifth characters are score 0, and the first, second, and sixth characters are score 1.

If the score with the lowest score is equal to or less than a predetermined threshold value T (T = -1 in this embodiment), the recognition result of the character is replaced with that of the second candidate, and A, B , C are counted. FIG. 14 shows the result. By changing the recognition result of the third character to “0” as the second candidate, the negative score is eliminated and the overall score (for example, the sum of the scores of each character) is improved. In other words, it can be considered that the combination obtained by adopting the second candidate only for the third character obtained this time has increased the certainty of the recognition, rather than the combination of the recognition results of the simple first candidates. If the score does not improve even if it is changed to another recognition candidate category, the character can be rejected.

In this way, the relationship with other characters (similarity,
By replacing the recognition result of a character having a low score obtained from (recognition result), a character that is likely to be erroneously recognized can be corrected.

When the characters to be recognized in the input document are all printed in the same typeface only, or
When all are handwritten by the same writer, the same font block extraction unit 6 can be omitted.

In the present embodiment, the character-shaped feature amount is obtained by the character-shaped feature extracting unit different from the characteristic amount used by the recognizing unit, but may be the same as the feature amount used by the recognizing unit. Alternatively, the dimension of the feature amount used in the recognition unit may be reduced using principal component analysis or the like.

Although the similarity calculating unit calculates the similarity as shown in (Equation 3) between the character-shaped feature quantities, other similarities may be used. Alternatively, another distance scale such as a Euclidean distance, a city distance, or a Mahalanobis distance between feature values may be used.

When the similarity calculation unit obtains a distance instead of a similarity, the similarity determination unit uses a predetermined threshold T1.
If it is smaller than this, it is determined to be “similar” and a predetermined threshold T
If it is larger than two, it may be determined that "dissimilar".

The score may be calculated by adding weights to the numerical values of A, B, and C, or by using a cutout score obtained by the preprocessing unit (character cutout unit) and a recognition unit obtained by the recognition unit. Add weights and scores
It may be used as an overall score.

In this embodiment, the recognition result of a character having a low score is replaced by using a recognition candidate of a recognition unit that outputs a plurality of recognition results. May be used as a similar character table that holds the information as information.

(Embodiment 4) FIG. 15 is a configuration diagram of a character recognition device according to Embodiment 4 of the present invention. The configuration is the same as that of the second embodiment except that the erroneous recognition detection unit 11 is replaced by the erroneous recognition correction unit 12.

Next, the operation of such a character recognition device will be described in detail. In the operation up to the similarity information storage unit 10, the recognition unit 5 performs recognition using the feature amount extracted by the feature extraction unit 4. The second embodiment is the same as the second embodiment except that a plurality of (N = 3 in this embodiment) recognition candidate categories are output.

The misrecognition correction unit 12 assigns the following score to each character in the group of characters determined to have the same font, similarly to the operation of the misrecognition detection unit 11 of the second embodiment. calculate. Here, the recognition result indicates one of a plurality of (number of candidates N = 3) recognition candidate categories,
Initially, all first candidates are used.

For all characters having the same recognition result as the target character, a score is calculated by a function as shown in FIG. 8 (representing the relationship between similarity and score), and the average Sa is obtained. For all the characters having different recognition results, the score is calculated by a function as shown in FIG. 9 (representing the relationship between the similarity and the score), and the average Sb is calculated. The sum S of the score Sa and the score Sb is calculated. Here, it is assumed that a set of character images as shown in FIG. 12 is specified by the same font block extraction unit 6, and the similarity calculation unit 8 calculates the similarity between characters as shown in FIG. Suppose it was done.

For example, in the score S of the first character, the characters having the same recognition result are the second and second characters, and the characters having different recognition results are the third, fourth and fifth characters. From, it becomes like below

Sa = {(200 × 0.92 × 0.92−100) + (200 × 0.91 × 0.91−100)} / 2 = 67.45 Sb = [{− 400 × (0.94−0.5) × (0.94−0.5)} + 0 + {-400 x (0.62-0.5) x (0.62-0.5)] / 3 = -23.73 ... S = Sa + Sb = 67.45 + (-43.737) = 39.72 ... Thus, for each character When the score is obtained, FIG.
(However, the decimal places are rounded off). The one with the lowest score S is the third character (score is -44).

If the score with the lowest score is equal to or smaller than a predetermined threshold value T (T = −20 in this embodiment), the recognition result of the character is replaced with that of the second candidate, and the score is calculated again. .

For example, for the first score S, since the second candidate “0” is adopted as the recognition result of the third character, the characters having the same recognition result are the three characters of the second, third, and sixth characters. Since the characters with different recognition results are the second and fourth characters, the result is as follows.

Sa = {(200 × 0.92 × 0.92−100) + (200 × 0.91 × 0.91−100) + (200 × 0.94 × 0−.94−100)} / 3 = 70.54 Sb = [0+ {-400 x (0.62-0.5) x (0.62-0.5)] / 2 = -2.88 S = Sa + Sb = 67.66 + (-2.88) = 67.66 Thus, when a score is obtained for each character, FIG.
become that way. The result is shown in FIG. By changing the recognition result of the third character to “0” as the second candidate, not only the score of the third character is raised, but also the overall score (for example, the sum of the scores of each character) is improved. In other words, it can be considered that the combination obtained by adopting the second candidate only for the third character obtained this time has increased the certainty of the recognition, rather than the combination of the recognition results of the simple first candidates. If the score does not improve even if it is changed to another recognition candidate category, the character can be rejected. When there are a plurality of characters with low scores, the recognition candidates may be changed one character at a time and the change in the score may be observed.

In this way, the relationship with other characters (similarity,
By replacing the recognition result of a character having a low score obtained from (recognition result), a character that is likely to be erroneously recognized can be corrected.

When the characters to be recognized in the input document are all printed in the same typeface only, or
When all are handwritten by the same writer, the same font block extraction unit 6 can be omitted.

In the present embodiment, the character-shaped feature amount is different from the characteristic amount used in the recognition unit in the character-shaped feature extraction unit, but may be the same as the characteristic amount used in the recognition unit. Alternatively, the dimension of the feature amount used in the recognition unit may be reduced using principal component analysis or the like.

Although the similarity calculating section calculates the similarity as shown in (Equation 3) between the character-shaped feature values, another similarity may be used. Alternatively, another distance scale such as a Euclidean distance, a city distance, or a Mahalanobis distance between feature values may be used.

When the similarity calculation unit obtains the distance instead of the similarity, the misrecognition detection unit may prepare the function of the distance and the score instead of the function of the similarity and the score. Further, the functions of the similarity and the score need not necessarily be those of FIGS. 8 and 9 and may be other appropriate functions as long as the following conditions are satisfied.

If the recognition result is the same character, a monotonically increasing function that gives a small score if the degree of similarity is low, and a large score if the degree of similarity is high. A monotonically decreasing function that gives a large score if the score is small and a small score if the degree of similarity is large. The score calculated in the present embodiment is calculated by the cutout score obtained by the preprocessing unit (character cutout unit) and the recognition unit. The obtained recognition score or the like may be weighted and added to be used as an overall score.

In this embodiment, the recognition result of a character having a low score is replaced by using a recognition candidate of a recognition unit that outputs a plurality of recognition results. May be used as a similar character table that holds the information as information.

[0086]

As described above, according to the present invention (the effect of the first aspect of the present invention), the reliability indicating the reliability of the character recognition result can be calculated not only from the target character but also from the target character and other characters. By using the relationship between the extracted glyph features, it is possible to obtain a reliable result if the recognition results are different although the characters are similar, or if the recognition results are the same but the characters are not similar. Characters that may be erroneously recognized due to the low degree can be detected.

Further, there is no need to collect character samples for each writer in advance, and the operation is performed even if the number of characters per category is not sufficiently large in the document to be recognized.

Further, according to the present invention (the effect of the seventh aspect of the present invention), when determining the reliability of the target character, the character shape feature values are determined to be similar to the target character, and the recognition result is not obtained. The number of different characters and the glyph features are determined to be dissimilar, and the number of characters with the same recognition result is counted as negative reliability, and the glyph features are determined to be similar and recognized. By counting the number of characters with the same result as positive reliability, the reliability using the relationship between the glyph features is defined by a simple method, and the possibility of erroneous recognition based on the reliability is determined. Certain characters can be detected.

Further, according to the present invention (the effect of the invention described in claim 8), when the reliability of the target character is obtained, if the recognition result is the same for the target character, the character shape feature amount If the reliability S1 according to the similarity is given, and if the recognition result is a different character, the reliability S2 according to the similarity between the glyph feature amounts is given, so that the degree of similarity reflecting the magnitude of the similarity is improved. It is possible to define a high degree of reliability and detect a character that may be erroneously recognized based on the degree of reliability.

Further, according to the present invention (the effect of the thirteenth aspect), for a character whose reliability is smaller than a predetermined threshold value, the recognition result is replaced with a correction candidate category and the reliability is recalculated. Do.

As for the recalculation of the reliability, the recalculation is performed not only for the character whose recognition result has been replaced but also for other characters.

When the correction candidate category is the correct answer category, not only the reliability of the character whose recognition result has been replaced is increased, but also the reliability of other characters is improved due to the relationship between the recognition result and the glyph feature. It is easy to judge the correctness of the candidate category.

Therefore, it is possible to correct a character category which may be erroneously recognized with high accuracy.

Therefore, by using the reliability calculation method and the erroneous recognition correction method of the present invention in a form recognition apparatus, highly accurate recognition becomes possible.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a first embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of a document image input according to the first embodiment of the present invention;

FIG. 3 is a diagram illustrating an example of a determination result in the same font block extraction unit according to the first embodiment of the present invention;

FIG. 4 is a diagram illustrating an example of a character image cut out by a preprocessing unit and a character category of a recognition result output by a recognition unit according to the first embodiment of the present invention;

FIG. 5 is a diagram showing an example of a relationship between similarities between character shapes and recognition results in a similar information storage unit according to the first embodiment of the present invention;

FIG. 6 is a block diagram showing a configuration of a second embodiment of the present invention.

FIG. 7 is a diagram showing a list of similarities calculated by a similarity calculator according to the second embodiment of the present invention;

FIG. 8 is a diagram showing a function between the similarity and the score when the recognition result is the same, which is used in the score calculation in the erroneous recognition detection unit according to the second embodiment of the present invention.

FIG. 9 is a diagram showing a function between similarity and score when recognition results are different, which is used in score calculation by an erroneous recognition detection unit according to the second embodiment of the present invention.

FIG. 10 is a diagram illustrating a score of each character calculated by an erroneous recognition detection unit according to the second embodiment of the present invention.

FIG. 11 is a block diagram showing a configuration of a third embodiment of the present invention.

FIG. 12 is a diagram illustrating an example of a character image cut out by a preprocessing unit and a character category of a recognition result output by a recognition unit according to the third embodiment of the present invention;

FIG. 13 is a diagram illustrating a similarity between characters and a score obtained from the recognition result when all the recognition results of the first candidates are used in the similar information storage unit according to the third embodiment of the present invention;

FIG. 14 is a diagram illustrating a similarity between character shapes when the recognition result of the second candidate is used only for the third character and the recognition result of the first candidate is used for all other characters in the similar information storage unit according to the third embodiment of the present invention. Diagram showing the score obtained from the degree and the recognition result

FIG. 15 is a block diagram showing a configuration of a fourth embodiment of the present invention.

FIG. 16 is a diagram showing a similarity between characters and a score obtained from the recognition result when all the recognition results of the first candidates are used in the similar information storage unit according to the fourth embodiment of the present invention.

FIG. 17 shows a similarity between character shapes when the recognition result of the second candidate is used for only the third character and the recognition result of the first candidate is used for all other characters in the similar information storage unit according to the fourth embodiment of the present invention. Diagram showing the score obtained from the degree and the recognition result

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image input part 2 Preprocessing part 3 Character image storage part 4 Feature extraction part 5 Recognition part 6 Identical character block extraction part 7 Character shape extraction part 8 Similarity calculation part 9 Similarity judgment part 10 Similar information storage part 11 False recognition detection part 12 Error recognition and correction unit

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Takeshi Megumi 1006 Kadoma Kadoma, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. F-term (reference) 5B064 AA01 AB02 AB03 AB13 EA08 EA26

Claims (17)

[Claims] A step of inputting the character image in a document image including a plurality of character images as input and performing character recognition by outputting at least one or more categories among all categories to be recognized as a recognition result; Extracting a character-shaped feature amount obtained by digitizing the shape of the character image; and extracting one character from the plurality of character images and one from the remaining character images.
Calculating a reliability indicating the reliability of the recognition result of the character based on the degree of similarity between the character-shaped feature amounts of the character or more and the difference of the recognition result. Method. 2. The method according to claim 1, wherein the step of performing the character recognition includes extracting a characteristic amount from the character image by some method and determining a category to be output based on the characteristic amount. 2. The character recognition method according to claim 1, wherein a characteristic amount or a part of the characteristic amount is output as a character-shaped characteristic amount. 3. The step of extracting the character-shaped feature value comprises:
3. The character recognition method according to claim 2, wherein a component obtained by performing a principal component analysis on the feature value or a part of the feature value and performing dimensional compression is output as a character-shaped feature value. 4. The method according to claim 1, wherein the similarity is represented by one of a first state of similarity, a second state of dissimilarity, and a third state of neither of them. Or the character recognition method according to 3. 5. If the degree of similarity between the character-shaped feature amount of the character and the character-shaped feature amount of another character is greater than a predetermined threshold value 1 with respect to an arbitrary character, the similarity is determined to be similar and the predetermined degree is determined. 5. The character recognition method according to claim 4, wherein the similarity is determined to be non-similar if the difference is smaller than the second threshold. 6. For a given character, if the distance between the character-shaped feature value of the character and the character-shaped feature value of another character is smaller than a predetermined threshold value 1, the similarity is determined to be similar and the predetermined similarity is determined. 5. The reliability calculation method according to claim 4, wherein the similarity is determined to be non-similar if the value is larger than the second threshold. 7. The number of characters whose similarity is similar to an arbitrary character and whose recognition result is different,
The number of characters whose similarity is dissimilar and the recognition result is equal, and the number of characters whose similarity is similar and the recognition result is equal are determined, and based on the number, reliability is determined. 5. The degree is calculated.
The character recognition method according to any one of claims 1 to 6. 8. A reliability according to a similarity between a character shape characteristic amount of the character and a character shape characteristic amount of another character when a recognition result of the character and another character is the same for an arbitrary character. S1 is given, and when the recognition result between the character and another character is different,
The character recognition method according to claim 1, wherein a reliability S <b> 2 is provided according to a degree of similarity between the character-shaped feature amount of the character and the character-shaped feature amount of another character. 9. When a given character has the same recognition result as that of another character, the reliability S1 corresponding to the distance between the character feature of the character and the character feature of another character is determined. And when the recognition result of the character is different from that of another character, a reliability S2 according to a distance between the character shape amount of the character and the character shape amount of another character is provided. ,
4. The character recognition method according to 2 or 3. 10. A character reliability is calculated by combining reliability obtained by the character recognition method according to any one of claims 1 to 9 with reliability obtained from other processing. Character recognition method. 11. A step of outputting a set of character images written by the same person and a set of character images printed in the same typeface as the same character type set,
The character recognition method according to claim 1, wherein character reliability is calculated for character images included in the same character type set. 12. A rejected character determination method, wherein a rejected character is determined based on the reliability obtained by the character recognition method according to claim 1. Description: 13. The character recognition method according to claim 1, wherein a recognition result of a character whose reliability is smaller than a predetermined threshold is classified into a category different from the recognition result. Replacing the reliability of a plurality of characters by replacing with a certain correction category, and determining the recognition result in the correction category when the recalculated reliability is larger than the original reliability. Misrecognition and correction method. 14. The method according to claim 11, wherein the character recognition step outputs a plurality of candidate categories as a recognition result, wherein the correction category is a category of a candidate different from the candidate for which the current reliability is calculated. The erroneous recognition correction method according to claim 13, wherein 15. An erroneously recognized character table in which a relationship between characters that are easily erroneously recognized is stored in advance, and the correction category is
14. The erroneous recognition correction method according to claim 13, wherein the erroneous recognition character table is a character. 16. A character recognition device for executing all or a part of the function of each step according to claim 1. Description: 17. A recording medium storing a program for causing a computer to execute all or a part of the functions of each of the steps according to claim 1. Description: