JP2003296661A - Character recognition device, character recognition method, its execution program and recording medium recording it - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To resolve a problem in conventional character recognition wherein a recognition candidate character that should be normally excluded from an object of post processing is outputted as a highly reliable object of post processing, a recognition candidate character that should normally be the object of post processing is considered unreliable and excluded from the object of post processing, and recognition accuracy is deteriorated due to determination of reliability. <P>SOLUTION: The character recognition device has a character recognition means for recognizing a sequence of coordinate points of a character inputted by handwriting, a feature extracting means for calculating an average writing speed of the sequence of coordinate points of the character inputted by handwriting as a feature amount for calculating reliability of a determination object recognition candidate character group outputted by the character recognition means, a reliability calculating means for calculating the reliability of the determination object recognition candidate character group on the basis of the feature amount from the feature extracting means and a statistical trend of sample data, and a post processing control means for controlling the post processing of the determination object recognition candidate character group on the basis of the reliability from the reliability calculating means. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device, a character recognition method for recognizing a character input by handwriting by determining the reliability (probability) of a character recognition result, an execution program thereof, and a recording thereof. Recording medium.

[0002]

2. Description of the Related Art In a conventional character recognition method, for example, a feature amount of a character input by handwriting is extracted and compared with a feature amount in a recognition dictionary, and the similarity between them is high or the distance between them is high. Small value (collectively, for convenience,
The recognition candidate character is output as "high confidence". However, in such character recognition for each character, a relatively accurate recognition result can be obtained when the written character is close to the feature amount of the recognition dictionary, but a character far from the feature amount of the recognition dictionary is written. In this case, a proper recognition result cannot be easily obtained.

Therefore, in addition to the character recognition for each character, the concatenation probability or co-occurrence probability of the preceding or following characters or words or phrases is detected, and the character string is matched from the certainty factor for each character and these probabilities. The so-called post-processing is performed to calculate the degree and output the recognized character string candidates for the entire character string according to the degree of matching. However, if too many recognition candidate characters are targeted during such post-processing, the post-processing calculation processing time will increase. Further, when the recognition candidate character having a low certainty factor is targeted, there is a possibility that a wrong character string candidate may be output as a result of the post-processing.

Therefore, in order to improve the accuracy of the post-processing while suppressing the increase of the calculation time, various methods of limiting the candidate characters to be the target of the post-processing have been proposed. For example, (a) a method of directly using the certainty factor of each candidate character obtained by character recognition, (b) estimating the reliability of the recognition result from the certainty factor of each candidate character, A method of controlling the number of candidate characters accordingly, a method of (c) estimating the reliability of the recognition result from the linguistic connection between adjacent recognition candidate character groups, and the like.

(A) is the simplest method. That is, this is a method in which the certainty factor of each candidate character is compared with a predetermined threshold value, and only candidate characters having a certainty factor higher than this threshold value are subjected to post-processing. In addition, as an example of the method of obtaining the reliability of the recognition result in the method of (b), there is Japanese Patent Laid-Open No.
-259226 gazette "a recognition method evaluation method and a recognition apparatus" is mentioned. This is the certainty of the 1st place candidate character and 2
This is a method of obtaining a difference value of the certainty factor of the rank candidate character and using the linear sum of this difference value and the certainty factor of the first rank candidate character as a measure of the correctness of the recognition result. This method focuses on the tendency that the confidence of the 1st place candidate character is relatively high when the recognition result is correct and the difference between the 1st place candidate character and the 2nd place candidate character is relatively large. Is. In addition to this method, a method that uses the ratio of the certainty factor of the first-ranked candidate character and the certainty factor of each of the second and lower candidate characters, the certainty factor of each candidate character is regarded as a multidimensional probability distribution, and Various methods have been proposed, such as a method for obtaining reliability.

Further, as the method (c), the reliability of the recognition result is statistically evaluated by paying attention to the connection probability between each character included in the recognition candidate character group to be judged and the recognition candidate character group immediately before or after the recognition candidate character group. It is a method to estimate it.

[0007]

However, generally, the certainty factor of each candidate character obtained as a result of character recognition does not necessarily properly reflect the correctness of the candidate character.
For example, in the case of relatively disorderly written characters, the certainty factor tends to be low even if the candidate character is the correct answer. On the other hand, if the candidate character and the correct character are similar,
Even if the answer is incorrect, the certainty factor may be high. Therefore, it is difficult to accurately limit the candidate characters by the method (a).

When the combination of the first-ranked candidate character and the second-ranked candidate character is a similar character such as "tsu"-"fu", "yuki"-"e", regardless of whether the recognition result is correct or not. , The difference value of these certainty factors tends to be small. If you want to recognize all hiragana, katakana, kanji, and alphanumeric characters,
Since such combinations of similar characters frequently occur, it is difficult to accurately estimate the reliability of the recognition result even with the method (b).

On the other hand, in the method (c), the method (a)
Problems due to the belief property as seen in (b) and (b) are avoided. However, in this method, in the case of a character string having a low concatenation probability between characters (for example, technical terms or proper nouns that are rarely used), the reliability of the recognition result may not be obtained accurately. Therefore, according to the above-mentioned conventional technique, the recognition candidate character that should be excluded from the target of the post-processing is output as the target of the post-processing with high reliability, and conversely, the recognition candidate character that should be the target of the post-processing is output. Results in excluding from post-processing as unreliable,
The result of the reliability judgment is that the recognition accuracy is rather lowered. The common point of these prior arts is that the feature amount used for calculation of reliability determination is extracted only from the character recognition result.

By the way, in so-called on-line handwritten character recognition for recognizing characters handwritten by a stylus or the like, the time-series coordinate point sequence, which is input data, is
It may contain information useful for estimating the reliability of the recognition result. For example, randomly written characters tend to have a faster writing speed, while carefully written characters tend to have a slower writing speed. Since misrecognition is likely to occur in randomly written characters, the reliability of the recognition result is considered to decrease. Also,
In general, when the number of writing strokes is small, many similar characters are present, so that erroneous recognition is likely to occur.

Further, when the number of writing strokes is significantly smaller than the number of normal strokes of the candidate character obtained as a result of the recognition processing, for example, the number of writing strokes is one, whereas the number of regular strokes of the first-ranked candidate character is ten. In the case of a stroke, it is highly likely that a character with a large number of strokes is written in one stroke, that is, an extreme continuous character. Even in such a case, erroneous recognition is likely to occur. Therefore, the mutual relationship between the writing stroke number and the regular stroke number of the candidate character also reflects the reliability of the character recognition result.

Therefore, the present invention solves the problems of the prior art by paying attention to the features useful for estimating the reliability obtained from the handwritten input data, such as the certainty factor of candidate characters and the connection probability between candidate characters. A character recognition device, a character recognition method, and an execution program thereof that can relatively accurately calculate the reliability even when the reliability of the recognition result cannot be appropriately estimated only by the features obtained from the character recognition result. And a recording medium storing the same.

[0013]

The invention according to claim 1 is
Character recognition means for recognizing the coordinate point sequence of characters input by handwriting and outputting a recognition candidate character group, and a feature amount for calculating the reliability of the judgment target recognition candidate character group output from the character recognition means. , The feature extraction means for calculating the average writing speed of the coordinate point sequence of the characters input by handwriting, the feature amount from the feature extraction means, and the statistical tendency of the sample data, and the determination target recognition candidate character A reliability calculation means for calculating reliability of the group, and a post-processing control means for controlling post-processing of the judgment target recognition candidate character group based on the reliability from the reliability calculation means. .

According to a second aspect of the present invention, in the first aspect of the present invention, the feature extraction means determines the average writing speed and the number of writing strokes in the coordinate point sequence of the handwritten input character as the determination target. It is characterized in that it is extracted as a feature amount for calculating the reliability of the recognition candidate character group. Claim 3
The invention according to claim 1 is the invention according to any one of claims 1 and 2, wherein the feature extraction unit further sets the number of regular strokes of upper N characters (N ≧ 1) in the determination target recognition candidate character group,
It is characterized in that it is extracted as a feature amount for calculating the reliability of the judgment target recognition candidate character group.

The invention according to claim 4 relates to claims 1 to 3.
In the invention described in any of 1,
Furthermore, the upper M characters (M
The certainty factor of ≧ 1) is extracted as a feature amount for calculating the reliability factor of the determination target recognition candidate character group. According to a fifth aspect of the present invention, in the invention according to any one of the first to fourth aspects, the feature extracting means further includes each recognition candidate character in the determination target recognition candidate character group and immediately before handwriting input immediately before that. In order to calculate the reliability of the determination target recognition candidate character group, the value of the connection probability with the recognition candidate character group and / or the value of the connection probability with the immediately subsequent recognition candidate character group for the handwriting input immediately after that. Is extracted as a feature amount of.

According to a sixth aspect of the present invention, in the invention according to the fifth aspect, the feature extracting means includes each recognition candidate character in the determination target recognition candidate character group and the highest belief in the immediately preceding recognition candidate character group. The value of the concatenation probability with the recognition candidate character of the degree and / or the value of the concatenation probability with the recognition candidate character of the highest confidence in the immediately following recognition candidate character group are the features of the determination target recognition candidate character group. It is characterized in that it is extracted as a quantity.

According to a seventh aspect of the present invention, in the invention according to the fifth aspect, the feature extraction means selects one recognition candidate character in the determination target recognition candidate character group and a recognition candidate character group immediately before or after the recognition candidate character group. Among the connection probabilities with each recognition candidate character, the highest connection probability is set as the connection probability between the one recognition candidate character and the immediately preceding or following recognition candidate character group.

The invention according to claim 8 relates to claims 1 to 7.
In any one of the inventions, the reliability calculation means includes a judgment score calculation means for calculating the certainty of one recognition candidate character in the judgment target recognition candidate character group from the characteristic amount as a judgment score, The reliability is calculated based on the discrimination score. According to a ninth aspect of the present invention, in the invention according to any one of the first to eighth aspects, the post-processing control means sets the post-processing target based on the reliability calculated by the reliability calculation means. The feature is that recognition candidate characters are limited.

According to a tenth aspect of the present invention, a character recognition step of recognizing a coordinate point sequence of characters input by handwriting and outputting a recognition candidate character group, and a determination target recognition candidate character group output from the character recognition step. As a feature amount for calculating the reliability of, the feature extraction step of calculating the average writing speed of the coordinate point sequence of the characters input by handwriting, the feature amount from the feature extraction step, and the statistical tendency of sample data A reliability calculation step of calculating the reliability of the judgment target recognition candidate character group based on the following, and after controlling the post-processing of the judgment target recognition candidate character group based on the reliability from the reliability calculation step And a processing control step.

According to an eleventh aspect of the present invention, in the invention of the tenth aspect, in the feature extracting step, the average writing speed and the number of writing strokes in the coordinate point sequence of the character input by handwriting are used as the determination target. It is characterized in that it is extracted as a feature amount for calculating the reliability of the recognition candidate character group. According to a twelfth aspect of the present invention, in the invention according to any one of the tenth to eleventh aspects, the feature extracting step further includes the upper N characters (N ≧) in the determination target recognition candidate character group.
The feature is that the regular stroke number of 1) is extracted as a feature amount for calculating the reliability of the determination target recognition candidate character group.

According to a thirteenth aspect of the present invention, in the invention according to any one of the tenth to twelfth aspects, the feature extraction step further includes the upper M characters (M ≧ 1) in the judgment target recognition candidate character group. It is characterized in that the certainty factor is extracted as a feature amount for calculating the reliability of the judgment target recognition candidate character group. The invention according to claim 14 is claim 10
In the invention described in any one of (1) to (13), the feature extraction step further includes the concatenation probability of each recognition candidate character in the determination target recognition candidate character group and the immediately preceding recognition candidate character group for the immediately preceding handwriting input. It is characterized in that the value and / or the value of the concatenation probability between the immediately subsequent recognition candidate character group and the immediately subsequent handwritten input is extracted as a feature amount for calculating the reliability of the determination target recognition candidate character group. .

According to a fifteenth aspect of the present invention, in the invention according to the fourteenth aspect, the feature extracting step includes each recognition candidate character in the determination target recognition candidate character group and the highest confidence in the immediately preceding recognition candidate character group. The value of the concatenation probability with the recognition candidate character of the degree and / or the value of the concatenation probability with the recognition candidate character of the highest confidence in the immediately following recognition candidate character group are the features of the determination target recognition candidate character group. It is characterized in that it is extracted as a quantity.

According to a sixteenth aspect of the present invention, in the invention according to the fourteenth aspect, the feature extracting step includes one recognition candidate character in the determination target recognition candidate character group and a recognition candidate character group immediately before or after the recognition candidate character group. Among the connection probabilities with each recognition candidate character, the highest connection probability is set as the connection probability between the one recognition candidate character and the immediately preceding or following recognition candidate character group.

According to a seventeenth aspect of the present invention, in the invention according to any one of the tenth to sixteenth aspects, the reliability calculation step is based on the feature amount and is one recognition candidate character in the judgment target recognition candidate character group. It is characterized by including a discriminant score calculation step for calculating the certainty of as a discriminant score, and calculating the reliability based on the discriminant score. The invention according to claim 18 is the invention according to any one of claims 10 to 17, wherein the post-processing control step is based on the reliability calculated from the reliability calculation step,
The feature is that the recognition candidate characters that are the target of the post-processing are limited.

The invention according to claim 19 is a program for causing a computer to execute each processing step in the character recognition method according to any one of claims 10 to 18. The invention according to claim 20 is a computer-readable recording medium recording a program for causing a computer to execute each processing step in the character recognition method according to any one of claims 10 to 18.

[0026]

BEST MODE FOR CARRYING OUT THE INVENTION <First Embodiment> A first embodiment of the present invention will be described below with reference to the drawings.
First, FIG. 1 is a circuit block diagram of a handwritten character recognition device according to the first embodiment.

In FIG. 1, an input unit 1 generates and outputs handwriting character information from handwriting input by handwriting on a tablet or the like. A character recognition unit 2 compares the handwritten character information supplied from the input unit 1 with the character feature amount of the character recognition dictionary 3, and the proximity (confidence) between the two is in the recognition dictionary from the first rank to the Nth rank. The character is output as a recognition candidate character of the handwriting character.
Reference numeral 3 is a character recognition dictionary in which candidate characters are stored together with their character feature amounts.

A feature extraction unit 4 calculates an average writing speed from the handwritten character information obtained by the input unit 1 and obtains the number of writing strokes. Further, by referring to the regular stroke number table 5, the regular stroke number of the upper N characters of the recognition candidate character group of the determination target supplied from the character recognition unit 2 is extracted. A regular stroke number table 5 stores the regular stroke number of the recognition target character in association with each recognition target character.

A discrimination score calculation unit 6 processes the feature amount obtained by the feature extraction unit 4 to calculate the true / false discrimination score of the recognition candidate character group to be judged. A recognition reliability calculation unit 7 compares the discrimination score from the discrimination score calculation unit 6 with the discrimination score-reliability conversion table 8 and outputs the reliability of the recognition candidate character group of the determination target. Reference numeral 8 denotes a discrimination score-reliability conversion table, which stores the relation between the discrimination score and the reliability as a table.

Reference numeral 9 is a recognition candidate number control unit, which is a reliability from the recognition reliability calculation unit 7, and a reliability-cumulative correct reading rate table 10.
And are compared to limit the number of recognition candidates of the determination target character. 10 is a reliability-cumulative correct reading rate table,
The relationship between the reliability and the correct reading rate is stored as a table. A language processing unit 21 performs post-processing on the number of recognition candidate characters set by the recognition candidate number control unit 9 and outputs a character string candidate.

The "character recognizing means" in the claims corresponds to the character recognizing unit 2 and the character recognizing dictionary 3 in FIG. 1 in the embodiment. The “feature extraction means” in the claims corresponds to the feature extraction unit 4 and the regular stroke number table 5 of FIG. 1 in the embodiment. The “reliability calculation means” in the claims corresponds to the discrimination score calculation unit 6, the recognition reliability calculation unit 7, and the discrimination score-reliability conversion table 8 of FIG. 1 in the embodiment. The “post-processing control means” in the claims corresponds to the recognition candidate number control unit 9 and the reliability-cumulative correct reading rate table 10 of FIG. 1 in the embodiment.

Next, details of the processing of each unit shown in the circuit block diagram will be described. First, the processing of the feature extraction unit 4 will be described with reference to FIG. FIG. 2 shows a sequence of coordinate points when the character “i” is written from the input unit 1. The feature extracting unit 4 calculates the average writing speed from the coordinate point sequence (xi, yi, ti, pi) (i = 1 to K) obtained by the input unit 1 and obtains the number of writing strokes. Here, K is the number of coordinate points, and xi and yi are i
The x coordinate and y coordinate of the th coordinate point and ti are the times when the i th coordinate point occurred. Further, pi indicates the state of the input pen at time ti, and has a value of pi = 1 when the pen is in contact with the tablet and pi = 0 when the pen is away from the tablet.

The writing speed vi when the pen moves from the i-th coordinate point to the (i + 1) -th coordinate point is expressed by equation 1.

[0034]

[Equation 1]

Therefore, the average writing speed V of the coordinate point sequence is calculated by the equation 2.

[0036]

[Equation 2]

The number of writing strokes S is calculated by counting the number of times the value of pi indicating the input pen state changes from 1 to 0.
inp is obtained. Further, the normal stroke number Sn1, Sn2, ..., SnN can be obtained by referring to the regular stroke number table 5 for the upper N characters of the recognition candidate character group. Next, with reference to FIG. 3, the processing in the discrimination score calculation unit 6 will be described.

The average writing speed V, the number of writing strokes Sinp, the number of regular strokes Sn1, Sn2, ...
The SnN set can be expressed as a predetermined vector (feature vector) in a (2 + N) -dimensional vector space. The discrimination score calculation unit 6 similarly extracts a feature vector as learning data for a sample in which the first-ranked recognition candidate character is correctly read or erroneously read, and extracts it as a feature vector of the recognition candidate character group to be judged. And the discrimination score of the recognition candidate character group is calculated.

For example, it is assumed that the learning data of the correct reading and misreading feature vectors and the feature vector of the recognition candidate character group to be judged are in a state as shown in FIG. FIG. 3 illustrates the case where N = 1, that is, the number of dimensions of the feature vector is 3, for simplification. The discrimination score calculation unit 6 obtains the average value (centroid vector) of the feature vectors of both classes and the covariance matrix in advance from the distribution of the feature vector of each set (class) of correct reading and erroneous reading, and stores this. ing.
Then, the Mahalanobis distances DMc and DMe between the center of gravity of each of these classes and the feature vector of the character group to be determined are obtained, and the ratio of these values, the logarithm of the ratio, or the difference is used as the determination score.

Here, the Mahalanobis distance DM is calculated as follows. That is, assuming that the center of gravity vector of the class C1 is m1 and the covariance matrix of the class C1 is Σ1, the Mahalanobis 2 from the predetermined feature vector x to m1.
The riding distance DM1 is defined by Equation 3.

[0041]

[Equation 3]

Here, the covariance matrix Σ1 is an n × n square matrix (n: the number of dimensions of the feature space), and its (i, j) element is the co-variance of the i-th feature quantity and the j-th feature quantity. The variance, ie Σ1 (i, j) = σij. In the above, the discriminant score was calculated using the Mahalanobis distance DM, but instead of this, a linear discriminant function was obtained by linear discriminant analysis from the distribution of the characteristic vectors of each class of correct reading and misreading,
The linear discriminant function may be applied to the feature vector of the recognition candidate character group to be determined to obtain the discrimination score.

Further, using the feature vector extracted from the learning sample of correct reading or erroneous reading as learning data, the neural network trained so as to judge whether the target characteristic vector is correct reading or erroneous reading is used, and the characteristic vector to be judged. The output value of the neural network with respect to may be used as the discrimination score. Next, referring to FIG. 4, the recognition reliability calculation unit 7
The process will be described.

For example, in calculating the discrimination score,
When the discriminant score is the ratio of the Mahalanobis distances DMc and DMe or the logarithm of the ratio, the relationship between the discriminant score obtained from each of the correct reading and misreading learning samples and the reliability is as shown in FIG. Here, the reliability is calculated from the learning sample by Bayes' theorem. That is, the discrimination score y
P (y | X1 = C), which is the ratio of the number of correct reading samples with respect to the total number of correct reading samples, and p is the ratio of the number of misreading samples with the discrimination score y to the total number of misreading samples.
(Y | X1 = E), the ratio of the total number of correct reading samples to the total number of samples is P (X1 = C), and the ratio of the total number of misreading samples to the total number of samples is P (X1 = E). The reliability of the recognition candidate character having the highest certainty factor of the recognition candidate character group having the score y can be calculated by the following equation.

P (X1 = C | y) = p (y | X1 = C)
・ P (X1 = C) / [p (y | X1 = C) ・ P (X1 =
C) + p (y | X1 = E) · P (X1 = E)] Here, X1 represents the first candidate character for recognition, X1 = C,
X1 = E means that the first-ranked recognition candidate characters are correct and incorrect, respectively. A discriminant score-reliability conversion table showing the relationship between the discriminant score and the reliability is created in advance from this equation and stored in the discriminant score-reliability conversion table 8. The recognition reliability calculation unit 7 is the discrimination score calculation unit 6
And the score of the determination score-reliability conversion table 8 are compared, and the corresponding reliability is output as the reliability of the first recognition candidate character of the recognition candidate character group.

Next, the processing of the recognition candidate number control section 9 will be described with reference to FIG. The table shown in the upper part of FIG. 5 shows the reliability with respect to the recognition candidate character having the first confidence in the determination target,
It shows the relationship with the cumulative probability that a correct reading character is included in the N-th recognition candidate characters to be judged. The probabilities in the table are calculated in advance based on the learning samples of correct reading and erroneous reading.

This table is stored in the reliability-cumulative correct reading rate table 10. Then, the recognition candidate number control unit 9
Compares the reliability of the recognition candidate character group to be determined with the reliability level in the table, and refers to the cumulative probability of the corresponding reliability level up to the maximum number of recognition candidate characters in the language processing unit 21. Decide whether to output. Here, up to what is output is determined by, for example, whether or not the cumulative probability of the corresponding reliability level reaches a predetermined threshold value. At this time, the threshold value to be set may be uniform for all reliability levels, or may be set individually for each reliability level.

Alternatively, the number of output candidates for each reliability level may be preset based on the table in the upper part of FIG. 5, and this may be stored in the reliability-cumulative correct reading rate table 10. Figure 5
The table below shows an example of the case where the reliability level and the number of output candidates are preset. When such a table is stored in advance in the reliability-cumulative correct reading rate table 10, the recognition candidate number control unit 9 reads the corresponding output candidate number from the table,
Accordingly, the recognition candidate characters output to the language processing unit 21 are limited.

In the above embodiment, since the feature amount useful for estimating the reliability of the recognition result is extracted from the handwritten input data, the reliability of the recognition result cannot be properly estimated from the certainty factor of the candidate character. Also in this case, the reliability can be calculated relatively accurately, and thus the recognition candidate character having a high correct reading rate can be output to the language processing unit. <Second Embodiment> Next, a second embodiment according to the present invention will be described below.

In this embodiment, the feature extraction processing in the feature extraction unit 4 is changed. First, FIG. 6 shows a circuit block diagram of the handwritten character recognition apparatus according to the present embodiment. The difference from FIG. 1 shown in the first embodiment is that the “character-character connecting probability dictionary 11” is added.

In the present embodiment, as the feature extraction processing in the feature extraction unit 4, in addition to the processing content shown in the first embodiment, the judgment target recognition candidate is obtained by referring to the intercharacter concatenation probability dictionary 11. A process of extracting a concatenation probability between the recognition candidate character in the character group and the recognition candidate character group immediately before and after the recognition candidate character is added. That is, in the first embodiment, the average writing speed, the number of writing strokes, and the top N of the judgment target recognition candidate character group are used as the feature amount used for reliability calculation.
Although the normal stroke number of characters is adopted, in the second embodiment,
In addition to these, the value Pbk (k = 1 to L) of the connection probability between the recognition candidate characters up to the top L in the judgment target recognition candidate character group and the recognition candidate character group immediately before it, and the judgment target recognition candidate The value Pfk (k = k) of the concatenation probability between the recognition candidate character up to the upper L rank in the character group and the recognition candidate character group immediately after it.
1 to L) are adopted as feature quantities used for reliability calculation.

Here, the kth character in the recognition target recognition candidate character group
In the present embodiment, the value Pbk of the connection probability between the rank candidate character and the immediately preceding recognition candidate character group is the connection between the kth candidate character and the immediately preceding 1st to Jth candidate characters. Maximum probability. Similarly, Pfk is the maximum value of the concatenation probability between the kth candidate character and the immediately following 1st to Jth candidate characters.

For example, in the example of FIG. 7, Pb1 for the 1st character "day" of the recognition candidate character to be judged is the "day" and the preceding 1st character "To" to the Jth character "hu". Among the respective connection probabilities P (C1 | Cbk), the maximum connection probability is adopted. In addition, P for the first character "day"
f1 is the concatenation probability P (Cfk | C1) of the "day" and the immediately preceding 1st character "mo" to the Jth character "death."
Among them, the highest connection probability is adopted. Similarly, Pb2 and Pf2 for the second character "month" of the recognition candidate character to be determined
Uses the maximum value of the concatenation probability for the character groups immediately before and after respectively.

Here, C1 represents the first character of the recognition candidate to be judged, and Cbk and Cfk represent the character of the kth candidate immediately before and after the recognition candidate, respectively. And P (Cj | Ci)
Represents the concatenation probability that the character Cj appears after the character Ci. In the second embodiment, the discriminant space shown in FIG. 3 has (2 + N + 2L) dimensions. Also, in the case of correct reading / wrong reading samples, in addition to the average writing speed, the number of writing strokes, the number of normal strokes of the upper N characters of the recognition candidate character group of the sample, the concatenation probability Pb with respect to the upper L characters of the recognition candidate character group of the sample
k and Pfk are feature extraction elements, and the discrimination score-reliability conversion table 8 and reliability-according to the sample data.
A table stored in the cumulative correct reading rate table 10 is set.

In the second embodiment, in addition to the feature quantity adopted in the first embodiment, a concatenation probability between characters included in adjacent recognition candidate character groups is adopted as a feature quantity for reliability determination. Therefore, the reliability determination can be performed with higher accuracy than in the first embodiment. As still another embodiment, in addition to the concatenation probabilities Pbk and Pfk, the certainty factor (similarity or distance value) of the recognition candidate characters up to the Mth position is added as a feature element, and (2 + N + 2L +
The feature vector of the recognition candidate character group may be extracted in the M) -dimensional vector space. In such a case, the discriminant space shown in FIG. 3 also has (2 + N + 2L + M) dimensions. In the second embodiment, not only the concatenation relation but also the certainty factor is taken into consideration, so that the reliability determination can be performed with higher accuracy.

By the way, in the above-described embodiment, a series of processing flows are explained by dividing the processing into each block in FIG. 1, but it is also possible to execute this processing flow by the CPU according to the control program. In such cases,
The processing flow is stored in the ROM or the RAM as a control program. Further, the character recognition dictionary 3, the regular stroke number table 5, the discrimination score-reliability conversion table 8, the reliability-cumulative correct reading rate table 10, and the inter-character connection probability dictionary 1
The reference data of 1 is also stored in the ROM or the RAM. C
The PU executes the above process according to the control program while referring to the reference data.

FIG. 8 shows a flow of the control program. Here, step S101 is processing in the input unit 1, step S102 is processing in the character recognition unit 2, step S103 is processing in the feature extraction unit 4, step S104 is processing in the discrimination score calculation unit 6, and step S105 is recognition reliability calculation. Processing in the unit 7, step S106 is processing in the recognition candidate number control unit 9.

The "character recognition step" in the claims corresponds to step S102 of FIG. 7 in the embodiment. The “feature extraction step” in the claims corresponds to step S103 in FIG. 7 in the embodiment. The “reliability calculation step” in the claims corresponds to steps S104 and S105 of FIG. 7 in the embodiment.
The control program and various reference data can be traded via a recording medium such as a flexible disk or a transmission medium such as the Internet. FIG. 9 shows an example of a file structure of data traded via a recording medium or a transmission medium. Data having such a file structure is recorded on the recording medium. Further, in the transaction via the transmission medium, the data having such a file structure is supplied via the transmission medium.

Although the embodiments according to the present invention have been described above, the present invention is not limited to the embodiments and various modifications can be made. For example, in the above-described embodiment, an example is shown in which the reliability is calculated using the average writing speed, the number of writing strokes, the number of normal strokes of candidate characters, the connection probability between adjacent recognition candidate character groups, and the certainty of candidate characters as a feature amount. However, it is also possible to select and employ only the feature amount that is particularly useful in the individual implementation devices from these various feature amounts.

Further, the processing content in the recognition candidate number control unit 9 of FIG. 1 may be such that the recognition result is rejected (invalid) according to the reliability, instead of limiting the number of recognition candidates. Furthermore, although the target of the handwriting input is a character as an example, it is needless to say that it is not limited to this and may be a figure.

In addition, the method for calculating the discrimination score in the discrimination score calculation unit 6 in FIG. 1 and the method for calculating the recognition reliability in the recognition reliability calculation unit 7 in FIG. 1 are the same as the Mahalanobis distance DM shown in the above embodiment. Alternatively, a method other than the method using Bayes' theorem can be adopted. The embodiments of the present invention can be appropriately modified in various ways within the scope of the technical idea of the present invention.

Further, the above-described embodiment is, to the last,
This is one embodiment of the present invention, and the meanings of the terms of the present invention or each constituent element are not limited to those described in the embodiment.

[0063]

As described above, according to the present invention, it is possible to obtain from the character recognition result such as the certainty factor of the candidate character and the concatenation probability between the candidate characters by using the feature useful for the reliability estimation obtained from the handwritten input data. Even if the reliability of the recognition result cannot be properly estimated with only the information, the reliability can be calculated with relatively high accuracy. Therefore, when this is adopted in the character recognition device, the accuracy of character recognition can be improved. Like

[Brief description of drawings]

FIG. 1 is a diagram showing a circuit block diagram according to a first embodiment.

FIG. 2 is a diagram for explaining a process of a feature extraction unit according to the first embodiment.

FIG. 3 is a diagram for explaining a process of a discrimination score calculation unit according to the first embodiment.

FIG. 4 is a diagram for explaining a process of a recognition reliability calculation unit according to the first embodiment.

FIG. 5 is a diagram for explaining a process of a recognition candidate number control unit according to the first embodiment.

FIG. 6 is a diagram showing a circuit block diagram according to a second embodiment.

FIG. 7 is a diagram for explaining a process of a feature extraction unit according to the second embodiment.

FIG. 8 is an execution flowchart according to the first and second embodiments.

FIG. 9 is a file structure of an execution program and reference data according to the second embodiment.

[Explanation of symbols]

1 ... Input section 2 ... Character recognition part 3 ... Character recognition dictionary 4 ... Feature extraction unit 5 ... Regular stroke number table 6 ... Discrimination score calculation unit 7 ... Recognition reliability calculation unit 8 ... Judgment score-reliability conversion table 9 ... Recognition candidate number control unit 10 ... Reliability-cumulative correct reading rate table 11 ... Character probability dictionary

Continued front page    F-term (reference) 5B064 AB04 BA05 DD05 DD07 EA18                 5B068 AA01 BD02 BD17 CC19 CD02                       CD06

Claims (20)

[Claims] 1. A character recognition means for recognizing a coordinate point sequence of characters input by handwriting and outputting a recognition candidate character group, and a reliability of a judgment target recognition candidate character group output by the character recognition means. As a feature amount for the feature extraction means for calculating the average writing speed of the coordinate point sequence of the characters input by handwriting, the feature amount from the feature extraction means, and the statistical tendency of the sample data, based on the It has reliability calculation means for calculating the reliability of the judgment target recognition candidate character group, and post-processing control means for controlling post-processing of the judgment target recognition candidate character group based on the reliability from the reliability calculation means. A character recognition device characterized by the above. 2. The invention according to claim 1, wherein the feature extraction unit determines the average writing speed and the number of writing strokes of the coordinate point sequence of the character input by handwriting as the reliability of the determination target recognition candidate character group. A character recognition device characterized by extracting as a feature amount for calculating a degree. 3. The invention according to any one of claims 1 and 2, wherein the feature extraction means further sets the number of regular strokes of upper N characters (N ≧ 1) in the determination target recognition candidate character group,
A character recognition device characterized by being extracted as a feature amount for calculating the reliability of the judgment target recognition candidate character group. 4. The invention according to any one of claims 1 to 3, wherein the feature extracting means further determines the certainty factor of the upper M characters (M ≧ 1) in the determination target recognition candidate character group. A character recognition device characterized by extracting as a feature amount for calculating the reliability of a target recognition candidate character group. 5. The invention according to claim 1, wherein the feature extracting means further includes each recognition candidate character in the determination target recognition candidate character group and the immediately preceding recognition candidate character for handwriting input immediately before the recognition candidate character. A feature value for calculating the reliability of the determination target recognition candidate character group by determining the value of the connection probability with the group and / or the value of the connection probability with the immediately subsequent recognition candidate character group for the handwriting input immediately after that. Character recognition device characterized by extracting as. 6. The invention according to claim 5, wherein the feature extracting means includes each recognition candidate character in the determination target recognition candidate character group and a recognition candidate character with the highest certainty factor in the immediately preceding recognition candidate character group. And / or extracting the value of the connection probability with the recognition candidate character having the highest certainty factor in the immediately following recognition candidate character group as the feature amount of the determination target recognition candidate character group. Characterized character recognition device. 7. The invention according to claim 5, wherein the feature extracting means includes one recognition candidate character in the determination target recognition candidate character group and each recognition candidate character immediately before or after the recognition candidate character group. A character recognition apparatus, wherein the highest connection probability among the connection probabilities between the one recognition candidate character and the immediately preceding or immediately following recognition candidate character group is used as the connection probability. 8. The invention according to claim 1, wherein the reliability calculation means determines the certainty of one recognition candidate character in the judgment target recognition candidate character group from the feature amount. A character recognition device comprising: a discrimination score calculating means for calculating the reliability, and calculating the reliability based on the discrimination score. 9. The invention according to any one of claims 1 to 8, wherein the post-processing control means, based on the reliability calculated by the reliability calculation means, is a recognition candidate character to be subjected to post-processing. A character recognition device characterized by restricting. 10. A character recognition step of recognizing a coordinate point sequence of characters input by handwriting and outputting a recognition candidate character group, and calculating reliability of a judgment target recognition candidate character group output from the character recognition step. As a feature amount for the feature extraction step of calculating the average writing speed of the coordinate point sequence of the handwritten input characters, the feature amount from the feature extraction step, and the statistical tendency of the sample data, based on the A reliability calculation step of calculating the reliability of the judgment target recognition candidate character group; and a post-processing control step of controlling post-processing of the judgment target recognition candidate character group based on the reliability from the reliability calculation step. A character recognition method characterized by the above. 11. The invention according to claim 10, wherein in the feature extraction step, the average writing speed and the number of writing strokes in the coordinate point sequence of the character input by handwriting are calculated as the reliability of the determination target recognition candidate character group. A character recognition method characterized by extracting as a feature amount for calculating the degree. 12. The invention according to claim 10, wherein in the feature extracting step, the number of regular strokes of upper N characters (N ≧ 1) in the determination target recognition candidate character group is further determined. A character recognition method characterized by extracting as a feature amount for calculating the reliability of a target recognition candidate character group. 13. The invention according to claim 10, wherein in the feature extracting step, the certainty factor of the upper M characters (M ≧ 1) in the determination target recognition candidate character group is further determined. A character recognition method characterized by extracting as a feature amount for calculating the reliability of a target recognition candidate character group. 14. The invention according to claim 10, wherein the feature extraction step further includes each recognition candidate character in the determination target recognition candidate character group and the immediately preceding recognition candidate character for the handwriting input immediately before the recognition candidate character. A feature value for calculating the reliability of the determination target recognition candidate character group by determining the value of the connection probability with the group and / or the value of the connection probability with the immediately subsequent recognition candidate character group for the handwriting input immediately after that. Character recognition method characterized by extracting as. 15. The invention according to claim 14, wherein in the feature extraction step, each recognition candidate character in the determination target recognition candidate character group and the recognition candidate character with the highest certainty factor in the immediately preceding recognition candidate character group are included. And / or extracting the value of the connection probability with the recognition candidate character having the highest certainty factor in the immediately following recognition candidate character group as the feature amount of the determination target recognition candidate character group. Character recognition method that features. 16. The invention according to claim 14, wherein in the feature extracting step, one recognition candidate character in the judgment target recognition candidate character group and each recognition candidate character in the recognition candidate character group immediately before or after the recognition candidate character group are recognized. A character recognition method, wherein the highest connection probability among the connection probabilities between the recognition candidate characters is the connection probability between the one recognition candidate character and the immediately preceding or following recognition candidate character group. 17. The invention according to claim 10, wherein the reliability calculation step determines a certainty of one recognition candidate character in the determination target recognition candidate character group from the feature amount. A character recognition method, comprising: a discrimination score calculating step of calculating as; and calculating the reliability based on the discrimination score. 18. The invention according to claim 10, wherein the post-processing control step is based on the reliability calculated from the reliability calculation step, and is a recognition candidate character to be post-processed. A character recognition method characterized by limiting. 19. The computer according to any one of claims 10 to 18.
A program for executing each processing step in the character recognition method described in any one of 1. 20. The computer according to any one of claims 10 to 18.
A computer-readable recording medium recording a program for executing each processing step in the character recognition method according to any one of 1.