JP2001243425A - On-line character recognition device and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that it can not be recognized accurately from only the distance values of a recognition result whether or not the recognition result is a correct character since a conventional on-line character recognition device is so constituted only that its recognition rate is increased. SOLUTION: The character quantity of an input pattern is extracted from time-series information and the character quality information and the distance values of the recognition result are used in combination for decision making, thereby actualizing a device and a method for character recognition which can decide the likelihood of the recognition result with high precision.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition apparatus, and more particularly to an online character recognition apparatus and an online character recognition method capable of accurately determining the likelihood of a recognition result by considering the character quality of written characters. is there.

[0002]

2. Description of the Related Art A conventional on-line character recognition apparatus for determining the likelihood of a recognition result is a conventional on-line character recognition device in which the minimum distance value of a recognition result (the distance value of a pattern that matches the standard pattern of a recognition dictionary) is smaller than a certain threshold value. If the difference between the minimum distance value of the recognition result and the distance value of the second smallest candidate is larger than a certain threshold value (e.g., There is a device that determines that the distance value is probable when the distance value and the distance value of the second candidate are far apart, and a device that uses both of them.

In order to further improve the judgment performance,
A determination device for each similar category (character) is prepared in advance, and at the time of determination, a candidate character of a recognition result is included in which similar category set to determine.
Japanese Patent Application Laid-Open No. 9603/1996), a device for determining using the similar category to which a plurality of candidate characters belong in the device (Japanese Patent Laid-Open No. 06-325213), and a distance value of a plurality of candidate characters as a recognition result. (Japanese Unexamined Patent Application Publication No. 08-212300 and Japanese Unexamined Patent Application Publication No. 08-329194).

FIG. 28 shows, for example, Japanese Patent Application Laid-Open No. H06-259603.
FIG. 1 is a block diagram illustrating a configuration of a conventional device that determines the likelihood of a recognition result disclosed in Japanese Patent Application Laid-Open Publication No. H10-15095. In the figure, reference numeral 1 denotes a feature extracting unit that inputs an input pattern that has been binary-converted for each character by a reading unit such as a scanner or a character cutout device, analyzes the features of the input pattern, and outputs the feature vector to the identification unit 3. Reference numeral 2 denotes a dictionary memory. Reference numeral 3 denotes a standard vector representing a standard feature of a recognition target category extracted from the dictionary memory 2 and a feature vector described above, and the similarity between them is calculated, for example, a distance value is calculated. Are generated in the descending order (if the distance value is smaller, the distance is smaller), and a signal representing category information of a higher candidate and a signal representing distance value information at that time are generated as recognition results. Is an identification unit that outputs.

A coefficient index memory 14 stores a group number (similar category group identification information) of a similar category to which each category to be recognized belongs for all categories of a character pattern to be input. (First memory) 15 is a coefficient memory (second memory) for storing, for each similar category group, a discrimination coefficient value learned using discrimination analysis so as to discriminate between the correct reading tendency and the misreading tendency. is there.

Reference numeral 16 determines a similar category group to which the character pattern to be recognized belongs, based on the group number of the similar category in the coefficient index memory 14 and the category information sent as the recognition result from the identification unit 3.
A coefficient selection unit 17 for outputting the discrimination coefficient value at that time from the coefficient memory 15. Is a difference value creating unit for calculating
Using the discriminant function between the distance value information of the recognition result obtained in step (1) and the plurality of difference value data obtained from the difference value creation unit 17 and the discrimination coefficient value from the coefficient memory 15 selected by the coefficient selection unit 16. A candidate probability calculation unit that calculates the inner product and outputs the discriminant function value as a candidate probability H.

FIG. 29 shows a memory 14 for coefficient index.
20 is a diagram showing the contents of a category number, and 21 is a group number of a corresponding similar category.

FIG. 30 shows the contents of the coefficient memory 15, and 31 is a discrimination coefficient value for each group corresponding to the group number 21 of the similar category.

Next, the operation will be described. First, the feature extraction unit 1 receives an input pattern and outputs a signal representing the feature vector to the identification unit 3. Next, the identification unit 3 collates the feature vector of the input pattern obtained from the feature extraction unit 1 with a standard vector representing the standard feature of the recognition target category extracted from the dictionary memory 2 and obtains distance value information of both. And output category information. That is, when the maximum number of target candidate categories is K,
Category information C is represented by equation (1).

C = {C _k | k = 1, 2,..., K} (1) where C _k is the k-th candidate category

Further, distance value information, that is, a distance value sequence d is represented by equation (2).

D = {d _k | k = 1, 2,..., K} (2) where d _k is the distance value of the k-th candidate category

Therefore, the coefficient selection unit 16 first focuses on the _first candidate category C1, and selects a discrimination coefficient value w to be used when calculating the candidate probability. For example, the first candidate category C
Assuming that the category number of ₁ is 3, referring to FIGS. 29 and 30, the discrimination coefficient value w is expressed as Expression (3).

W = (w21, w22,..., W2k, w2 (2 · k−1)) (3)

Next, the difference value creation unit 17 receives the distance value sequence d and obtains difference value data focusing on the distribution shape of the entire distance value sequence, ie, a difference value data sequence d '. An example of the difference value data sequence d ′ is expressed as in Expression (4).

D ′ = {d ′ _k : d _{k + 1} −d _k | k = 1, 2,..., K} (4) where d ′ _k is the distance value of the k-th candidate category

Further, the candidate probability calculation unit 18 determines the distance value sequence d of the recognition result, the difference value data sequence d ′ obtained by the difference value generation unit 17 and the coefficient memory 15 selected by the coefficient selection unit 16. Based on the discriminant function f (w, d, d ') using the coefficient value w as a parameter, the recognition result candidate accuracy H
Is calculated as in equation (5).

H = f (w, d, d ') = w21d1 + w22d2 + ... w2KdK + w2 (K + 1) d'K + w2 (K + 2) d'2 + ... + w2 (2K-1) d'K-1 (5)

Thereafter, the obtained candidate accuracy H is applied as a quantified value of the certainty of the recognition result to a rejection determination process or a post-processing determination (not shown) at a later stage.

[0020]

As described above,
In a conventional online character recognition apparatus for determining the likelihood of a recognition result, recognition is performed based on a distance value (including a difference value) of the recognition result and a candidate likelihood obtained from a discrimination coefficient value in a similar category set to which the category of the recognition result belongs. Since the likelihood of the result is determined, the discrimination performance is improved when the candidate characters of the recognition result are distributed in a predetermined similar character category, but the discrimination accuracy is reduced for an unknown distribution. There was a problem to do.

Further, since the distributions of the categories overlap each other, if all the categories are classified into any of the similar categories, it is not possible to determine the similar category which must be actually considered, and There is a problem that performance is reduced.

When the number of similar category groups is increased in order to improve the discrimination accuracy, it is necessary to store (2K-1) dimension coefficient values for each similar category group. Therefore, there is a problem that the memory capacity becomes very large.

In order to obtain the discrimination coefficient with high accuracy, it is necessary to use 10 times as many learning samples as the number of dimensions, and 10 × (2K−1) correct reading samples and error samples are required for each similar category group. Become. Generally, it is relatively easy to prepare correct reading samples, but it is very difficult to prepare error samples on a device with a relatively high recognition rate, and it is difficult to narrow down the number of similar category groups. This necessitates a problem that the discrimination accuracy is reduced.

Further, since the distance value of the recognition result is prepared as a standard pattern for each category, the distance value of the recognition result is not necessarily a distance value reflecting the character quality, and is relatively carefully written. Even in this case, the accuracy of the candidate may not be high, and there is a problem that even if the writing quality of the user is improved, the discrimination accuracy is not improved.

The present invention has been made to solve the above-described problem. In order to determine the certainty of the recognition result, not only the distance value of the recognition result but also the time-series information at the time of input is used. It is an object of the present invention to realize an online character recognition device and an online character recognition method capable of improving the accuracy of discrimination of a recognition result by extracting character quality using the method.

[0026]

A character recognition apparatus according to the present invention includes a character quality extracting unit for inputting an input pattern and category information and extracting character quality information for character quality determination, and determining character quality in advance. A character quality determination information storage memory storing information for performing the recognition, a feature vector of the input pattern, and a feature vector of a standard pattern in each category in the dictionary memory. The identification unit to be output, the distance value information of the recognition result obtained by the identification unit, the character quality information obtained by the character quality extraction unit and the judgment information from the character quality judgment information storage memory are used to determine the accuracy of the recognition result. And a judgment unit that outputs the judgment result.

In the character recognition device according to the present invention, when character quality information is extracted by the character quality extraction unit, character quality information is extracted based on time-series information obtained from an input pattern.

In the character recognition apparatus according to the present invention, when character quality information is extracted by the character quality extraction unit, the character quality is determined based on the number of strokes obtained from the input pattern as time-series information or the normal number of strokes of the candidate character of the recognition result. This is for extracting information.

In the character recognition device according to the present invention, when character quality information is extracted by the character quality extracting unit, the character quality information is calculated based on the number of strokes obtained from the input pattern as time-series information and the number of regular strokes in the candidate character of the recognition result. The character quality information is extracted based on the stroke number variation rate.

In the character recognition device according to the present invention, when character quality information is extracted by a character quality extraction unit, character quality information is extracted based on stroke information obtained from an input pattern as time-series information.

In the character recognition device according to the present invention, when character quality information is extracted by the character quality extraction unit, a stroke rectangle in an input pattern and a stroke rectangle prepared in advance for a candidate character of a recognition result are used as stroke information. Is extracted as character quality information.

In the character recognition apparatus according to the present invention, when the character quality information is extracted by the character quality extraction unit, the stroke direction of the input pattern and the stroke of a stroke prepared in advance for the candidate character of the recognition result are used as the stroke information. The distance value from the direction is extracted as character quality information.

In the character recognition device according to the present invention, when character quality information is extracted by the character quality extraction unit, the direction between successive strokes of the input pattern and the candidate character of the recognition result are prepared in advance as stroke information. The distance value from the direction between successive strokes is extracted as character quality information.

In the character recognition device according to the present invention, when the character quality information is extracted by the character quality extraction unit, the sum of the lengths of the strokes of the input pattern and the candidate characters of the recognition result are prepared in advance as stroke information. The distance value from the total length of the strokes thus extracted is extracted as character quality information.

In the character recognition device according to the present invention, when character quality information is extracted by the character quality extraction unit, stroke type information of an input pattern and stroke candidate information of a recognition result are prepared in advance as stroke information. It is compared with stroke type information and extracted as character quality information.

In the character recognition device according to the present invention, when character quality information is extracted by the character quality extraction unit, character quality information is extracted by using information on a bending point obtained from an input pattern.

According to the character recognition method of the present invention, there is provided a character quality extracting step of extracting information for character quality determination from an input pattern, a feature vector of the input pattern, and a standard for a recognition target character stored in a dictionary memory in advance. A discrimination step of comparing the feature vector of the pattern and outputting category information and distance value information of the recognition result; and the recognition result obtained in this discrimination step, the character quality information obtained in the character quality extraction step, The method further includes a determination step of determining the likelihood of the recognition result using the determination information for determining the character quality stored in the determination information storage memory and outputting the determination result.

[0038]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a circuit block diagram showing a configuration of an online character recognition device according to Embodiment 1 of the present invention. In the figure, reference numeral 1 denotes a feature extraction unit that inputs a pattern that has been binary-converted for each character by a reading unit such as a scanner or a character cutout device, analyzes the features of the input pattern, and outputs a feature vector to the identification unit 3. 2 is a dictionary memory,
Reference numeral 3 compares a standard vector representing a standard feature of the recognition target category extracted from the dictionary memory 2 with the above-described feature vector, calculates a similarity between the two, for example, a distance value, and sorts the similarity in descending order of similarity (distance value (In the order of smaller distances), a category group that becomes a top candidate is generated, and a recognition unit outputs a signal representing category information of a top candidate and a signal representing distance value information at that time. is there. These scanner 1, dictionary memory 2, and identification unit 3 are the same as the components of the conventional apparatus shown in FIG.

Reference numeral 50 denotes a character quality extraction unit for inputting an input pattern and category information to extract character quality information for character quality determination, and 51 stores character quality determination information preliminarily storing information for determining character quality. The memory 52 has a certainty of the recognition result from the character quality information obtained from the character quality extraction unit 50, the distance value information of the recognition result obtained from the identification unit 3, and the character quality judgment information obtained from the character quality judgment information storage memory 51. And outputs a result of the determination.

FIG. 2 is a diagram showing an input pattern. In the figure, reference numeral 60 denotes an input pattern "A" written with seven strokes
It is.

FIG. 3 is a diagram showing a recognition result for the input pattern 60 output from the identification unit 3. In the drawing, reference numeral 61 denotes a rank in a candidate after sorting in the order of smaller distance values (in order of smaller distance from the standard pattern), 62 denotes a category of candidate characters, and 63 denotes a distance value.

FIG. 4 shows a character quality judgment information storage memory 51.
FIG. 6 is a diagram showing a stroke number information table in which the regular stroke numbers (correct stroke stroke numbers) of the recognition target characters stored in.
In the figure, 64 is a category, and 65 is a regular number of strokes in the category.

FIG. 5 is a diagram showing the stroke number variation rate for the input pattern 60 obtained by the character quality extraction unit 50 of the recognition result, and 70 in the figure shows the stroke number variation rate of the candidate character of the recognition result. .

FIG. 6 is a diagram showing the variation ratio from the normal number of strokes in a large number of handwritten characters, where the horizontal axis represents the stroke number variation and the vertical axis represents the data ratio (%).

FIG. 7 is a diagram showing an input pattern. In the figure, reference numeral 80 denotes an input pattern “A” written with seven strokes upward and to the right.

FIG. 8 is a diagram showing a recognition result for the input pattern 80 output from the identification unit 3. In the figure, reference numeral 81 denotes a rank in a candidate after sorting in the order of smaller distance values (in order of smaller distance from the standard pattern), 82 denotes a category of candidate characters, and 83 denotes a distance value.

FIG. 9 is a diagram showing a stroke count variation rate for the input pattern 80 obtained by the character quality extraction unit 50 of the recognition result. In FIG. 9, reference numeral 84 denotes a stroke count variation rate of the candidate character of the recognition result. .

FIG. 10 is a diagram showing a distribution in the case where the first-ranked recognition result candidate character is a correct character and a distribution in the case of an error character when the distance value of the recognition result is used as a scale. The distribution of correct characters, 86 indicates the distribution of error characters.

FIG. 11 shows the distribution when the first-ranked recognition result candidate character is the correct character and the error character when the first-ranked recognition result candidate character is the scale when the recognition score when the character value is used in combination with the distance value of the recognition result. It is a figure which shows distribution, 87 in a figure.
Shows an example of the distribution of correct characters, 88 shows an example of the distribution of error characters, and 89 shows an example of a threshold value for determining a correct character from an error character.

Next, the operation will be described. Assuming that the recognition unit 3 outputs the recognition result shown in FIG. 3 for the input pattern 60 shown in FIG. 2, the character quality extracting unit 50 extracts character quality information from the input pattern. Here, the stroke number variation ratio di from the regular stroke number in the candidate character of the i-th recognition result is defined as Expression (6) as a scale indicating the character quality in the input pattern.

[0051]

(Equation 1)

Specifically, in order to obtain the stroke number variation rate for the first candidate character in the recognition result shown in FIG. 3, the regular stroke number in the first candidate category “A” is stored in the character quality determination information. The number of strokes is obtained from the stroke number information table stored in the memory 51, and is calculated using Expression (6). The second and third recognition result candidate characters are obtained in the same manner, and a stroke number variation rate 70 is obtained for each candidate character as shown in FIG.
As shown in FIG. 4, the regular stroke number of the second-place candidate character “pot” is 11, and therefore, the stroke count variation rate 70 of this character is
A relatively large value.

Next, the character quality extracting section 50 sends the character quality information (the stroke number variation rate 70 in FIG. 5) of the obtained recognition result to the determining section 52. The determination unit 52 determines the distance value between the sent character quality information and the recognition result sent from the identification unit 3 (63 in FIG. 3).
And the likelihood of the recognition result is determined. The determination of the likelihood of the recognition result is performed by using the distance value of the recognition result (the distance value of the first place, the difference between the second place and the first place), and the character quality information as parameters, and correct character distribution and error characters. The Mahalanobis's general distance to the distribution (distance measurement that compensates for the statistical correlation between features) is obtained, and the value obtained by subtracting the distance Dec to the correct answer from the distance De to the error distribution as shown in Expression (7) is the determination threshold. If it is greater than TH, it is determined as a correct character, otherwise, it is determined as an error character.

[0054]

(Equation 2)

In the first embodiment, the input parameters used for the determination are the distance value of the first candidate character in the recognition result, the second candidate character distance value−the first candidate character distance value, and the character quality information. Assuming that the variation rate of the number of strokes of the first candidate character is n, then n = 3.

The mean vector μc of the correct character distribution and the inverse matrix Σc of the variance / covariance matrix of the correct character distribution
^-1 is calculated using a sample in which the first character of the recognition result is a correct character in a large number of learning samples prepared in advance.

Also, the average vector in the error character,
Similarly, the inverse matrix is calculated using a sample of the learning samples in which the first character of the recognition result has an error. These parameters are stored in the character quality determination information storage memory 51.
Shall be stored in

In a general Mahalanobis generalized distance,
Although the determination is made only based on the magnitude relationship between Dc and De, in the first embodiment, the correct answer rate and the error rate can be adjusted by setting an appropriate threshold (TH). This threshold (T
H) is also calculated in advance by a learning sample and stored in the character quality determination information storage memory 51.

As a result of collecting a large number of handwritten character data samples and analyzing the variation in the number of strokes in the written characters from the regular number of strokes, a square type sample as shown in FIG.
More than a percentage. In a normal character recognition device,
The recognition rate of a square type character sample tends to be higher than that of a character sample in which the number of strokes has changed. Therefore, the stroke number variation rate is effective as a scale for discriminating a correct character from an error character. That is, when the first character in the recognition result is a correct character, the stroke number variation rate tends to be small, and when the character is an error character, it tends to be large. Therefore, the stroke number variation rate is effective as a scale for determining the likelihood of a correct character.

In the first embodiment, the stroke number variation rate 70 shown in FIG. 5 is obtained for the input pattern 60, and the stroke number variation rate 70 in the first recognition result is 0 (input pattern 6).
0 is written with the number of regular strokes of the first character “A” of the recognition result of 7), which is a small value.

Next, it is assumed that an input pattern 80 written upward and to the right as shown in FIG. 7 has been input. Input pattern 60
Similarly to the operation in the case of, it is assumed that the recognition unit 3 obtains the ranking 81, the category 82, the distance value 83 as the recognition result shown in FIG. 8, and the character quality extraction unit 50 obtains the stroke number variation rate 84 shown in FIG. As shown in FIG. 8, the first character in the recognition result is the character “ox” due to the slanted writing, and an error occurs.

However, the distance value is the same as in the case of the input pattern 60 in which the first place is the correct character, and the Mahalanobis distance in the case of the input pattern 60 and the input pattern 80 is only the distance value of the recognition result. Will be the same.
However, even in this case, as shown in FIG. 9, the stroke number variation rate of the first character “ox” becomes a relatively large value of 0.43, and the Mahalanobis general distance to the correct character distribution in the first embodiment. Dc is the case of the correct character (in the case of the input pattern 60), and the case of the error character (the input pattern 80).
), The Mahalanobis' general distance De to the error character distribution is larger in the case of the correct character than in the case of the error character. As a result, even if the certainty of the recognition result cannot be determined only by the distance value of the recognition result, the determination can be correctly performed by applying the rate of change in the number of strokes.

Although the first embodiment has been described above,
In the first embodiment, Mahalanobis' generalized distance is used as a discrimination method, but discrimination may be made by another discrimination method (for example, discrimination by a linear discrimination function).

In the first embodiment, the correct character distribution,
For each error character distribution, the Mahalanobis 'general distance was calculated, but the Mahalanobis' general distance to the correct character distribution was calculated, and if the general distance was smaller than a preset threshold, it was determined as a correct character, and if it was larger, it was determined as an error character. You may. In particular, in the case of a highly accurate recognition method, it is difficult to prepare a large number of error character samples. Therefore, it is preferable to make a determination using only the correct character distribution.

In the first embodiment, the Mahalanobis' generalized distance is obtained by using the inverse matrix of the variance / covariance matrix. However, the inverse matrix of the correlation matrix may be used.

Further, in the first embodiment, the determination is made using both the character quality information and the distance value of the recognition result. However, the determination may be made using only the character quality information.

In the first embodiment, as the information used together with the character quality information, the first distance value and the difference between the second and first distance values are used as the recognition result. You may make it use together only.

Also, in the first embodiment, the determination is made using only the stroke number variation rate of the first character of the recognition result as the character quality information. However, the average value of the stroke number variation rate of a plurality of candidate characters, A difference value between the second and the number-of-number-of-times change rates may be used.

Further, in the first embodiment, the stroke number variation rate is used for all data. However, the learning data is divided into a pattern written with the normal stroke number and a pattern written with a non-regular stroke number. The stroke number variation rate may be used only for a pattern written other than the regular stroke number.

As described above, according to the configuration of the first embodiment, as shown in FIG. 10, when only the distance value of the recognition result is used, the object is to finally increase the recognition rate. Therefore, the total number of patterns that are correct characters (corresponding to the area of the distribution 85 of correct characters) is the total of the number of patterns that are error characters (corresponding to the area of the distribution 86 of error characters).
Although more, the distribution 85 of the correct characters and the distribution 86 of the error characters are overlapped distributions, and often do not provide a correct scale from the viewpoint of correctly determining the correct characters and the error characters.

Therefore, in the present invention, the reliability of the input pattern is improved by introducing a new measure of character quality represented by the degree of variation in the number of strokes from the normal number of strokes, as shown in the first embodiment, for example. It is possible to improve the ratio that can be accurately determined. In other words, in the character recognition device, the lower the character quality of the input pattern is, the more difficult it is to correctly recognize the input pattern and the lower the reliability of the distance value of the recognition result is. Is more reliable.

Therefore, when the character quality is judged to be good by using the scale of the character quality in combination with the distance value, it is judged that the reliability is high even if the distance value is relatively large, so that the character value is relatively large. If the input pattern is determined to be correct and character quality is determined to be poor, it is possible to determine that there is no reliability even when the distance value is relatively small, thereby improving the accuracy of determination of certainty. Can be improved.

As a result, as shown in FIG. 11, although the recognition rate is low, the degree of overlap between the correct character and the error character is small (the separability between the correct character distribution and the error character distribution is high) and the character with few erroneous determinations is obtained. A recognition device can be realized. Here, when the judgment threshold 89 is set, a case where the judgment score is higher than the threshold 89 is judged as a correct answer, and a case where the judgment score is lower than this is judged as an error. In addition, the error character distribution in a portion having a higher discrimination score than the threshold 89 is very small, so that erroneous judgment can be reduced.

Further, in the first embodiment, in addition to the distance value information of the recognition result, the degree of change in the number of strokes of the input pattern is extracted as character quality information and used for determining the certainty of the recognition result. Despite writing in strokes (square style), it is possible to improve the certainty as a correct character in a pattern with a low degree of coincidence with a standard pattern used for recognition, and it is possible to accurately determine a correct character. .

Further, in the first embodiment, the degree of change in the number of strokes is represented by a change ratio from the normal number of strokes. Therefore, a relatively large variation in the number of strokes is allowed. For a character having a small number of strokes, the scale becomes a scale that can be steadily reflected even with a small change, and it is possible to suppress a variation in the determination accuracy in the number of strokes of the input pattern and to accurately determine a correct character.

Further, in the first embodiment, since only the parameters reflecting the two distributions of the correct character distribution and the error character distribution need to be prepared in common for all the categories, the parameter learning is performed using a large number of learning samples. Is possible, and highly accurate discrimination becomes possible.

Since the number of strokes information is used,
The accuracy of determination of similar characters having different stroke counts can be improved, and this can be realized without preparing the determination parameter information for each similar category as in the conventional example, and the device can be realized with a small memory capacity.

Further, since the accuracy of the determination of the certainty of the recognition result is improved, the rejection determination in a hierarchical character recognition device (a device in which a plurality of character recognition devices are configured in series) can be performed with high accuracy. A highly accurate character recognition device can be realized.

If the first character in the recognition result is determined to be the correct character, the number of candidate characters in the recognition result can be reduced to one. Characters can be reduced, and reading accuracy including language processing is improved.

Embodiment 2 Hereinafter, a second embodiment of the present invention will be described with reference to FIG. 1 and FIGS. In the figure, the same or corresponding parts as those in the first embodiment are denoted by the same reference numerals, and redundant description is omitted. FIG. 12 shows a first stroke (a stroke written first in time) and a second stroke (second stroke in time) of a carefully written recognition target character stored in the character quality determination information storage memory 51. A written stroke) is a stroke rectangle information table in which the width and height of the circumscribed rectangle are normalized (normalized to values of 0 to 1) and stored, where 90 is rectangle information of the first stroke, and 91 is rectangle information of the first stroke. This is two-stroke rectangle information.

FIG. 13 shows an example in which the width (W) and the height (H) of the circumscribed rectangle of the first stroke are extracted from the input pattern 60 of FIG. is there.

FIG. 14 is an example in which the width (W) and the height (H) of the circumscribed rectangle of the first stroke are extracted from the input pattern 80 of FIG. 7, and 93 in the drawing indicates the first stroke. is there.

FIG. 15 is a diagram showing the rectangle evaluation value of the first stroke extracted by the character quality extraction unit 50 for the input pattern 60. In FIG. 15, reference numeral 94 denotes a stroke rectangle evaluation value.

FIG. 16 is a diagram showing a rectangle evaluation value of the first stroke extracted by the character quality extraction unit 50 for the input pattern 80. In FIG. 16, reference numeral 95 denotes a stroke rectangle evaluation value.

FIG. 17 is a diagram showing an example of the input pattern 60 from the first stroke extracted by the character quality extraction unit 50 to the second stroke.
It is a figure which shows the direction vector to a stroke,
0 is a direction vector between strokes.

First, the operation of the input pattern 60 shown in FIG. 2 will be described. As in the first embodiment, the identification unit 3
Suppose that a recognition result as shown in FIG. 3 is obtained. Next, the character quality extraction unit 50 obtains the width (W) and the height (H) of the circumscribed rectangle of the first stroke 92 from the input pattern 60 as shown in FIG.

Next, the character quality extraction unit 50 converts the width and height information of the circumscribed rectangle of the first stroke serving as a reference into the character quality determination information for each candidate character of the recognition result output from the identification unit 3. It is determined by referring to the stroke rectangle information table in the storage memory 51. For example, the width and height of the circumscribed rectangle for the first character “A” of the recognition result are 1.0 and 1.0, respectively.
0 and 0.05.

Next, the character quality extracting unit 50 determines the circumscribed rectangle of the first stroke from the circumscribed rectangle information of the first stroke obtained from the input pattern 60 and the circumscribed rectangle information of the first stroke obtained from the recognition result. A rectangle evaluation value di of the first stroke indicating the degree of coincidence is calculated. Specifically, the rectangle evaluation value di is obtained by Expression (8).

[0089]

(Equation 3)

The first stroke rectangle evaluation value 94 specifically obtained by the character quality extraction unit 50 by the equation (8) is shown in FIG.
Shown in Since the first character is the correct character, the stroke rectangle evaluation value is a small value.

Next, as in the case of the first embodiment, the determination unit 52 determines the character quality information (the first stroke rectangle evaluation value of the first recognition result) output by the character quality extraction unit 50,
The Mahalanobis' general distance is calculated using the distance value information 63 of the recognition result output by the identification unit 3 as an input parameter, and it is determined whether the recognition result is a correct character or an error character.

Next, the case of the input pattern 80 shown in FIG. 7 will be described. As in the case of the input pattern 60, FIG.
It is assumed that the rectangle evaluation value 95 of the first stroke shown in FIG. As shown in FIG. 16, the rectangle evaluation value of the first stroke in the first candidate character “ox” is the same as the first stroke 93 of the input pattern and the shape of the first stroke of the character “ox”. The evaluation value is a large value of 0.59.

Next, the determination unit 52 calculates the Mahalanobis general distance in the same manner as in the case of the input pattern 60, and determines whether the recognition result is a correct character or an error character.
In this case, the character is determined to be an error character.

As described above, the rectangle evaluation value of the first stroke is a small value when the recognition result is a correct character, and a large value when the recognition result is an error character if the shape of the first stroke is different. The second embodiment can be applied to the case where the correct character or the error character cannot be determined only by the recognition result.
The correct determination can be made as shown in FIG.

As described above, in the second embodiment, the rectangle evaluation value of the first stroke is used as the character quality information, but the rectangle evaluation value of the second stroke and thereafter may be used. Also, rectangle evaluation values of a plurality of strokes may be used together.

In the second embodiment, the width and height of the circumscribed rectangle are used as the stroke evaluation values. However, a direction vector from the start point to the end point of the stroke may be used.

Further, in the second embodiment, the width and height of the circumscribed rectangle are used as the stroke evaluation values.
Direction vector 1 between successive strokes as shown in FIG.
00 (vector between the middle points of the strokes) may be used.

As described above, in the second embodiment, character quality information is extracted from partial stroke information of an input pattern in addition to distance value information of a recognition result, and is used to determine the certainty of the recognition result. As a result, the character quality can be determined more accurately, and the determination accuracy is improved.

Embodiment 3 Hereinafter, a third embodiment of the present invention will be described with reference to FIG. 1 and FIGS. The same or corresponding parts as those in Embodiments 1 and 2 are denoted by the same reference numerals, and redundant description will be omitted. FIG. 18 is a stroke type table expressing seven types of strokes constituting a character stored in the character quality determination information storage memory 51. In the drawing, reference numerals 100 to 106 denote corresponding strokes.

FIG. 19 shows the character quality judgment information storage memory 5.
4 is a configuration stroke table 110 that indicates the number of configuration strokes in a recognition target character, which is stored in 1.

FIG. 20 shows the number of strokes constituting the input pattern 60 shown in FIG. 2 using the stroke type table. It is a figure showing the absolute value of the difference.

FIG. 21 shows the number of constituent strokes of the input pattern 80 of FIG. 7 using the stroke type table, and compares the number of constituent strokes of the input pattern 80 with the constituent strokes of the candidate character of the recognition result. It is a figure showing the absolute value of the difference.

First, the operation of the input pattern 60 shown in FIG. 2 will be described. As in the first embodiment, the identification unit 3
Suppose that a recognition result as shown in FIG. 3 is obtained. Next, the character quality extraction unit 50 extracts the number of each constituent stroke from the input pattern 60 as shown in FIG. 20, using the stroke type table of FIG.

Next, the character quality extraction unit 50 determines the number of reference strokes of each constituent stroke for each candidate character output from the identification unit 3 in the constituent strokes stored in the character quality judgment information storage memory 51. It is determined with reference to the table 110. For example, the criteria for the first character “A” in the recognition result are three horizontal strokes, three vertical strokes, and one key stroke.

Next, the character quality extraction unit 50 calculates the absolute value of the difference between the constituent strokes in the input pattern 60 and the number of constituent strokes in each candidate character, and calculates the sum of the differences as the distance di when the stroke type is used. And Specifically, the distance di is obtained by Expression (9).

[0106]

(Equation 4)

The difference for each stroke type is shown in FIG.
It is shown in FIG. Since the input pattern 60 is a carefully written character "A", the constituent stroke of the input pattern matches the constituent stroke of the character "A" in the stroke type table. Therefore, for the recognition result "A", the difference is 0 for each stroke, and the final distance d
i also becomes 0.

Next, as in the case of the first embodiment, the determination unit 52 determines the character quality information (the first stroke type evaluation value of the recognition result) output by the character quality extraction unit 50 and the identification unit 3. Calculates the Mahalanobis' generalized distance using the distance value 63 of the recognition result output as an input parameter, and determines whether the recognition result is a correct character or an error character.

Next, the case of the input pattern 80 of FIG. 7 will be described. As in the case of the input pattern 60, FIG.
It is assumed that the difference between the number of strokes of each constituent stroke in the input pattern 80 shown in FIG. 1 and the input pattern 60 of each stroke type for the recognition result character is determined. As shown in FIG. 21, the input pattern 80 is a character “A” written inclining rightward, so that all the horizontal strokes are classified as rightward upward strokes.

Next, when the stroke type evaluation value of the first candidate character "ox" is obtained, since the stroke type evaluation value is a character different from the input pattern, the distance becomes a large value of nine. In this example, the distance is also a large value of 8, even for the second correct character "A" in the recognition result. That is, even if the first place is the correct character, the distance becomes a large value because the input pattern is a character written with an inclination.

Next, the determination unit 52 calculates the Mahalanobis' general distance as in the case of the input pattern 60, and determines whether the recognition result is a correct character or an error character.
In this case, the character is determined to be an error character.

As described above, the stroke type evaluation value is a small value when the input pattern is carefully written and the recognition result is a correct character, and when the recognition result is an error character or the input pattern is not carefully written. Even when the correct character or the error character cannot be determined only by the recognition result, the correct determination can be performed as shown in the third embodiment.

As described above, in the third embodiment, the stroke types are classified into seven types. However, a different number of types may be used.

Further, in the third embodiment, only one stroke type information is provided for the character to be recognized, but a plurality of stroke type tables may be provided for one character.

In the third embodiment, the evaluation is performed using all the strokes constituting the character. However, the stroke type may be evaluated only for a partial stroke.

In the third embodiment, the absolute value of the difference in the number of strokes is used. However, as shown in equation (10), the difference between the number of strokes of the input pattern and the number of strokes of the regular character is calculated as the distance. It may be di.

[0117]

(Equation 5)

As described above, in the third embodiment, in addition to the distance value information of the recognition result, the character quality information is extracted according to the stroke type of the input pattern, and is used to determine the certainty of the recognition result. Therefore, the character quality can be more accurately determined, and the determination accuracy is improved.

Embodiment 4 Hereinafter, a fourth embodiment of the present invention will be described with reference to FIG. 1 and FIGS. The same or corresponding parts as those in Embodiments 1, 2, and 3 are denoted by the same reference numerals, and redundant description is omitted. FIG. 22 is a diagram illustrating a state in which a bending point is extracted from an input pattern using a known technique. In this example, ∠QPR is set to an appropriate threshold θ.
If smaller, the point P is extracted as a bending point.

FIG. 23 is a diagram showing an example in which a bending point is extracted from the input pattern 60. In FIG. 23, 120 is the extracted bending point.

FIG. 24 shows the character quality judgment information storage memory 5.
4 is a bending point number table 131 that stores the number of bending points of the carefully written recognition target character stored in No. 1;

FIG. 25 is a diagram showing an example of calculating the bending evaluation value 132 from the number of bending points in the input pattern 60 and the number of bending points in the recognition result character.

FIG. 26 is a diagram showing an example in which inflection points are extracted from the input pattern 121 in which a decoration has occurred at the start point when writing a character. In the drawing, 122 to 125 indicate the extracted inflection points. .

FIG. 27 is a diagram showing an example in which the bending evaluation value 135 is calculated from the number of bending points in the input pattern 121 and the number of bending points in the recognition result character. 136 in the figure is
This is the distance value of the recognition result for the input pattern 121.

First, the operation of the input pattern 60 shown in FIG. 2 will be described. As in the first embodiment, the identification unit 3
Suppose that a recognition result as shown in FIG. 3 is obtained. Next, as shown in FIG. 22, the character quality extraction unit 50 extracts sample points at regular intervals for each stroke of the input pattern, calculates the angle between the sample point before and after the sample point of interest, and performs the specified bending. A comparison is made with the detected angle θ, and if it is smaller than θ, the point of interest is extracted as a bending point. Since the input pattern 60 is a carefully written pattern, only one bending point 120 is extracted as shown in FIG.

Next, the character quality extracting unit 50 determines the reference bending point number for each candidate character of the recognition result output from the identification unit 3 by using the bending point number table 131 in the character quality determination information storage memory 51. Seek and ask. For example, the reference number of bending points in the first character “A” in the recognition result is one.

Next, the character quality extraction unit 50 calculates the absolute value of the difference between the number of inflection points in the input pattern 60 and the number of inflection points in each of the candidate characters as the inflection evaluation value di using equation (11).

Di = | N−Ni | (11) N: number of bending points of input pattern Ni: number of bending points in i-th candidate of recognition result

FIG. 25 shows the bending evaluation value 132 for the input pattern 60 calculated by the character quality extracting unit 50. Since the input pattern 60 is a carefully written character "A",
The number of inflection points in the input pattern and the number of inflection points in the character “A” in the inflection point number table 131 match. Therefore, the bending evaluation value di in the recognition result “A” is also zero.

Next, as in the case of the first embodiment, the judging section 52 compares the character quality information (the first bending evaluation value of the recognition result) output by the character quality extracting section 50 with the identifying section 3. The Mahalanobis' general distance is calculated using the distance value 63 of the output recognition result as an input parameter, and it is determined whether the recognition result is a correct character or an error character.

Next, the case of the input pattern 121 shown in FIG. 26 will be described. As in the case of the input pattern 60,
It is assumed that a difference between the number of inflection points in the input pattern 121 shown in FIG. 26 and the number of inflection points for the recognition result character has been obtained. As shown in FIG. 26, since the input pattern 121 is a pattern in which decoration is generated at the starting point of the stroke, the number of bending points is four.

Next, when the bending evaluation value of the first candidate character “A” is obtained, the bending evaluation value is 3. That is,
Even if the first place is a correct character, the evaluation value is a large value because decoration is generated in the input pattern.

Next, the determination unit 52 calculates the Mahalanobis' general distance as in the case of the input pattern 60, and determines whether the recognition result is a correct character or an error character.
In this case, the character is determined to be an error character.

As described above, the bending evaluation value is a small value when the input pattern is carefully written and the recognition result is a correct character, and when the recognition result is an error character or the input pattern is not carefully written, Even when the value is large and the correct character or the error character cannot be determined only by the recognition result, the correct determination can be performed as shown in the fourth embodiment.

As described above, in the fourth embodiment, the number of inflection points is extracted from points sampled at equal intervals, but the inflection points may be obtained by other methods.

In the fourth embodiment, only one bending point number table is provided for a character to be recognized. However, a plurality of bending point number tables may be provided for one character.

Further, in the fourth embodiment, the evaluation based on the number of bending points is performed for all strokes constituting a character. However, the number of bending points is evaluated only for a partial stroke or a partial area. You may.

In the fourth embodiment, the absolute value of the difference in the number of inflection points is used. However, as shown in equation (12), the difference between the number of strokes of the input pattern and the number of strokes of the regular character is calculated as the distance. It may be di.

[0139]

(Equation 6)

As described above, in the fourth embodiment, in addition to the distance value information of the recognition result, the character quality information is extracted based on the bending evaluation value of the input pattern, and is used to determine the certainty of the recognition result. So, I wrote it with a regular number of strokes, but it was possible to reduce erroneous judgments on broken handwritten characters, splashes, and habit characters where suppression occurred, and it became possible to judge character quality more accurately, and discrimination accuracy The effect of improving is obtained.

[0141]

As described above, according to the present invention, a character quality extracting unit for inputting an input pattern and category information to extract character quality information for character quality determination and a character quality extracting unit for determining character quality in advance. Using the character quality judgment information storage memory storing the information of the distance, the distance value information of the recognition result obtained by the identification unit, the character quality information obtained by the character quality extraction unit, and the information from the character quality judgment information storage memory. A determination unit that determines the likelihood of the recognition result and outputs the determination result,
Since the configuration is such that the character quality is determined, it is possible to separate a pattern having a reduced character quality from a pattern having a good character quality. In general, the distance distribution between a character whose character quality has deteriorated and a character having good quality has different distributions. If the same distribution is substituted for all the distance distributions, the performance of correct answer / error discrimination deteriorates. Therefore, the performance is improved by learning for discrimination according to the character quality. In addition, there is an effect that performance can be improved by introducing a scale indicating character quality as a discrimination parameter.

Further, according to the present invention, the correctness / error of the character quality is determined by using the time-series information. Is correctly written, it is possible to improve the correctness / error discrimination performance even if the character-like evaluation value is small. The time-series information is information that is easier for the user to write correctly than the character shape, and less burdens on the user than writing characters in a beautiful form. Further, when the recognition is performed only by the character-shaped information, the determination is performed in consideration of the time-series information, so that the time-series information works complementarily, and there is an effect that the determination performance is improved.

Further, according to the present invention, when character quality information is extracted by the character quality extraction unit, the character quality information is obtained based on the number of strokes obtained from the input pattern as time-series information or the normal number of strokes in the candidate character of the recognition result. , The time-series information can easily extract the stroke number information, and the larger the difference between the reference stroke number of each character and the stroke number of the input pattern, the lower the character quality may be. Is a measure of character quality that can be extracted most easily. In addition, there is an effect similar to that in the case of determining the correct answer / error of the character quality using the time-series information.

Further, according to the present invention, when the character quality information is extracted by the character quality extraction unit, the number of strokes calculated based on the number of strokes obtained from the input pattern as the time-series information and the normal stroke number information of the candidate character of the recognition result. Since the character quality information is configured to be extracted based on the change rate, the change in the number of strokes in a character with a large number of strokes and the change in the number of strokes in a character with a small number of strokes can be normalized by applying the change number of strokes. In other words, three strokes are written in two strokes, and 20 strokes are 19
The degree of influence on character quality differs in the case of writing with an image. Therefore, normalizing the difference in the number of strokes with the number of normal strokes has the effect of equalizing the degree of influence, and has the same effect as each of the above items.

Further, according to the present invention, when the character quality information is extracted by the character quality extraction unit, the character quality information is extracted based on stroke information obtained from the input pattern as time-series information. In the case of low-quality characters, the information of strokes (strokes) constituting the characters often collapses, and character quality is determined based on the information. The stroke information is information that can be easily extracted by online character recognition, and can be determined in more detail than the stroke number information. Therefore, the stroke information can be determined with high accuracy, and the same effects as those of the above items can be obtained.

According to the present invention, when character quality information is extracted by the preceding character quality extraction unit, a stroke rectangle in an input pattern and a stroke rectangle prepared in advance for a candidate character of a recognition result are used as stroke information. The width of the
Since the height distance value is extracted as character quality information, circumscribed rectangle information of a stroke that can be easily extracted is used as stroke information. When the character is tilted, the circumscribed rectangle also changes, so that there is an effect that the character can be easily extracted and the character quality can be accurately extracted.

Also, according to the present invention, when the character quality information is extracted by the character quality extraction unit, the stroke direction of the input pattern and the stroke direction prepared in advance for the candidate character of the recognition result are used as the stroke information. Is extracted as character quality information, so the ratio of kanji composed of straight strokes is high, and the stroke direction can be substituted by the direction from the start point of the stroke to the end point of the stroke. Can be extracted. Further, when the character shape is broken, this information is also likely to change, so that there is an effect that the character quality can be accurately extracted.

Further, according to the present invention, when character quality information is extracted by the character quality extraction unit, the direction between successive strokes of the input pattern and the candidate character of the recognition result are prepared in advance as stroke information. Since the distance value from the direction between successive strokes is configured to be extracted as character quality information, the direction between strokes indicates the relative positional relationship between the strokes. Is easy to change. In addition, by setting the interval between consecutive strokes, the stroke order information can be reflected. In addition, there is an effect that it can be easily extracted.

Further, according to the present invention, when character quality information is extracted by the character quality extraction unit, the sum of the lengths of the strokes of the input pattern and the candidate characters of the recognition result are prepared in advance as stroke information. It is configured to extract the distance value with the sum of the lengths of the strokes as character quality information, so the stroke length can be easily extracted by calculating the sum of the distances between sample points in the same stroke. It is. Characters with broken shapes or written quickly
Since the length of the stroke is easily changed, there is an effect that character quality can be accurately extracted.

Further, according to the present invention, when character quality information is extracted by the character quality extraction unit, stroke type information of an input pattern and stroke information prepared in advance for a candidate character of a recognition result are used as stroke information. Since the type information is compared with the type information and extracted as the character quality information, the shape of the stroke constituting the character can be classified into a certain fixed shape. By determining the character quality based on the classification information, there is an effect that the determination can be performed with high accuracy.

Further, according to the present invention, when the character quality information is extracted by the character quality extraction unit, the character quality information is extracted by using the information of the inflection point (point at which the direction changes greatly) obtained from the input pattern. As a result, the inflection point can be easily extracted. In the case of continuous characters, broken characters, and the like, the strokes are combined with each other and a new bending point is generated. Therefore, the bending point information is easily changed, so that there is an effect that character quality can be accurately extracted.

[Brief description of the drawings]

FIG. 1 is a circuit block diagram showing a configuration of an online character recognition device according to Embodiment 1 of the present invention.

FIG. 2 is a diagram showing an input pattern.

FIG. 3 is a diagram showing a recognition result for an input pattern.

FIG. 4 is a diagram illustrating a stroke number information table storing a regular stroke number in a recognition target character;

FIG. 5 is a diagram showing the number-of-images change rate with respect to an input pattern.

FIG. 6 is a diagram showing a variation ratio from a regular number of strokes in handwritten characters collected in large quantities.

FIG. 7 is a diagram showing an input pattern.

FIG. 8 is a diagram showing a recognition result for an input pattern.

FIG. 9 is a diagram showing the number-of-images change rate with respect to an input pattern.

FIG. 10 shows a first example in which the distance value of the recognition result is used as a scale.
It is a figure which shows the distribution in case a recognition result candidate character of a position is a correct character, and the distribution in case of an error character.

FIG. 11 shows a distribution in the case where the first-ranked recognition result candidate character is a correct character and a distribution in the case of an erroneous character when the discrimination score obtained when character quality is used in combination with the distance value of the recognition result is used as a scale. FIG.

FIG. 12 is a stroke rectangle information table in which the width and height of a circumscribed rectangle of a first stroke and a second stroke of a written recognition target character are normalized and stored.

13 is an explanatory diagram of extracting a width (W) and a height (H) of a circumscribed rectangle of a first stroke with respect to the input pattern of FIG. 2;

14 is an explanatory diagram of extracting a width (W) and a height (H) of a circumscribed rectangle of a first stroke from the input pattern of FIG. 7;

FIG. 15 is a diagram showing a rectangle evaluation value of a first stroke extracted by the character quality extraction unit for the input pattern of FIG. 2;

16 is a diagram showing a rectangle evaluation value of a first stroke extracted by the character quality extraction unit for the input pattern of FIG. 7;

17 is a diagram illustrating a direction vector from a first stroke to a second stroke extracted by the character quality extraction unit with respect to the input pattern of FIG. 2;

FIG. 18 is a stroke type table expressing strokes constituting characters stored in a character quality determination information storage memory in seven types.

FIG. 19 is a configuration stroke table showing the number of configuration strokes in a recognition target character.

20 is a diagram showing an absolute value of a difference between the number of constituent strokes of the input pattern of FIG. 2 and a constituent stroke of a candidate character as a recognition result.

21 is a diagram showing an absolute value of a difference between the number of constituent strokes of the input pattern of FIG. 7 and the constituent strokes of candidate characters as a recognition result.

FIG. 22 is a diagram showing a state of extracting a bending point from an input pattern.

FIG. 23 is a diagram illustrating an example in which a bending point is extracted from an input pattern.

FIG. 24 is a table showing the number of inflection points of a written recognition target character.

25 is a diagram illustrating an example of calculating a bending evaluation value from the number of bending points in the input pattern of FIG. 2 and the number of bending points in a recognition result character.

FIG. 26 is a diagram illustrating an example in which a bending point is extracted from an input pattern in which a decoration has occurred at a starting point during character writing.

FIG. 27 is a diagram illustrating an example in which a bending evaluation value is calculated from the number of bending points in an input pattern in which a decoration has occurred at the start point during character writing and the number of bending points in a recognition result character.

FIG. 28 is a block diagram showing a configuration of a conventional online character recognition device.

FIG. 29 is a diagram illustrating the contents of a coefficient index memory.

FIG. 30 is a diagram showing the contents of a coefficient memory.

[Explanation of symbols]

1 feature extraction unit, 2 dictionary memory, 3 identification unit, 14
Coefficient index memory, 15 Coefficient memory, 16
Coefficient selecting section, 17 difference value creating section, 18 candidate probability calculating section, 20 category numbers, 21 similar category group numbers, 31 group discriminant values, 50 character quality extracting section, 51 character quality judgment information storage memory, 52 Judgment unit, 60 input patterns, 61 ranks, 62 categories, 63 distance values, 64 categories, 65 regular strokes, 70 strokes variation rate, 80 input patterns, 81 ranks, 82 categories, 83 distance values, 84 strokes variation rates, 85 Distribution of correct characters, distribution of 86 error characters,
87 Distribution of correct characters, 88 Distribution of error characters, 89
Threshold value, 90 first stroke rectangle information, 91 second stroke rectangle information, 92 first stroke, 93 first stroke, 94 stroke rectangle evaluation value, 95 stroke rectangle evaluation value, direction vector between 100-106 strokes, 110 Stroke type table, 111
Difference for each stroke type, 112 Difference for each stroke type, 120 Bending point, 121 input pattern,
122 to 125 bending point, 131 bending point number table,
132 bending evaluation value, 133 rank, 134 category, 135 bending evaluation value, 136 distance value.

Claims (12)

[Claims] 1. A dictionary memory in which a feature vector of a standard pattern in a character to be recognized is stored in advance, a feature extraction unit for extracting a feature vector indicating a feature of a pattern from an input pattern, and an input pattern obtained from the feature extraction unit A character recognition device having an identification unit that compares a feature vector of a reference pattern with a feature vector of a standard pattern in each category in the dictionary memory and outputs category information and distance value information of a recognition result. , A character quality extraction unit for extracting character quality information for character quality determination, a character quality determination information storage memory storing information for character quality determination in advance, and a recognition obtained by the identification unit. The resulting distance value information, the character quality information obtained by the character quality extraction unit, and the determination information from the character quality determination information storage memory. Recognition determines the likelihood of a result, the character recognition apparatus characterized by comprising a judgment unit for outputting a determination result using the. 2. The character according to claim 1, wherein the character quality information is extracted by the character quality extracting unit based on time-series information obtained from an input pattern. Recognition device. 3. The character quality information extraction unit extracts character quality information based on stroke number information obtained from an input pattern as time-series information or regular stroke number information on a candidate character of a recognition result. 3. The character recognition device according to 2. 4. A method for extracting character quality information by the character quality extracting unit, based on stroke number information obtained from an input pattern as time-series information and stroke number fluctuation rate calculated from regular stroke number information in candidate characters of a recognition result. The character recognition device according to claim 3, wherein the character quality information is extracted. 5. The character quality information extraction unit extracts character quality information based on stroke information obtained from an input pattern as time-series information.
The character recognition device according to 1. 6. A method according to claim 1, wherein said character quality extracting unit extracts character quality information by extracting, as stroke information, a width and a height of a stroke rectangle in an input pattern and a stroke rectangle prepared in advance for a candidate character as a recognition result. The character recognition device according to claim 5, wherein the distance value is extracted as character quality information. 7. When extracting character quality information by the character quality extracting unit, a distance value between a stroke direction of an input pattern and a stroke direction prepared in advance for a candidate character of a recognition result as stroke information. 6. The character recognition device according to claim 5, wherein the character recognition device extracts the character quality information. 8. When extracting character quality information in the character quality extracting section, the direction between successive strokes of an input pattern and the distance between successive strokes prepared in advance for a candidate character of a recognition result as stroke information. The distance value from the direction is extracted as character quality information.
The character recognition device according to 1. 9. When extracting character quality information in the character quality extracting section, a sum of lengths of strokes of an input pattern as stroke information and a length of a stroke prepared in advance for a candidate character of a recognition result are obtained. 6. The character recognition device according to claim 5, wherein a distance value from the sum of the first character and the second character is extracted as character quality information. 10. When character quality information is extracted by the character quality extraction unit, stroke type information of an input pattern and stroke type information prepared in advance for a candidate character of a recognition result are used as stroke information. The character recognition device according to claim 5, wherein the character recognition device compares and extracts the character quality information. 11. The character recognition apparatus according to claim 2, wherein when the character quality information is extracted by the character quality extraction unit, the character quality information is extracted using information on a bending point obtained from an input pattern. 12. A feature extracting step of extracting a feature vector indicating a feature of a pattern from an input pattern, a feature vector of the input pattern obtained from the feature extracting step, and a standard pattern in a recognition target character stored in a dictionary memory in advance. A character quality extracting step of extracting information for character quality determination from an input pattern, the character quality extracting step comprising: Recognition is performed using distance value information of the recognition result obtained in the identification step, character quality information obtained in the character quality extraction step, and determination information for determining character quality previously stored in a character quality determination information storage memory. A character recognition method comprising: a determination step of determining the likelihood of a result and outputting a determination result.