JPH08212300A - Method and device for recognizing character - Google Patents

Abstract

PURPOSE: To obtain a character recognition device leading out the evaluation value of a new recognition result for obtaining accurate and stable post processing result without depending on the kind of a character category. CONSTITUTION: The character recognition device collates the feature vector of a recognition object character pattern and a standard vector to obtain distance values and outputs a character category group making a character a high-ordered candidate in order of the smaller distance value and the string of distance values corresponding to individual character category. A difference value data string is prepared from the string of the distance values by a difference value preparing part 4 and a discrimination coefficient value is extracted from a coefficient memory 5 storing a discrimination coefficient value for discriminating correct/wrong reading in each character category. And, concerning this discrimination coefficient, a candidate probability calculation part 6 determines candidate probability to a first candidate category by inner-product- calculating the difference value data string and the string of the distance values with the discrimination coefficient value. In addition, a candidate probability calculation part 7 determines candidate probability to an i-th candidate character by multiplying first candidate probability by the ratio of a first candidate distance value to i-th (i>1) candidate distance value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for determining the certainty (hereinafter, candidate certainty) of a recognition result in character units in a character recognition device, and particularly to post-processing such as character string collation based on the recognition result. The present invention relates to a character recognition method and a character recognition device that contribute to stabilization of characters.

[0002]

2. Description of the Related Art As shown in FIG. 3, a conventional general character recognition apparatus includes a feature extraction unit 1, a dictionary memory 2, and an identification unit for identifying the output of the feature extraction unit 1 using the dictionary memory 2. Three
And have. The feature extraction unit 1 extracts a feature of a character pattern that is binary quantized character by character by a reading means such as a scanner or a character slicing device (not shown) of a character string to be recognized on a form, and its output is It becomes the feature vector of the character pattern. It should be noted that the term “character” used in the present specification includes, in addition to Chinese characters, kana, etc., numbers, symbols, etc., unless otherwise specified. The dictionary memory 2 stores in advance standard vectors representing standard features of character categories. The identification unit 3 collates the standard vector extracted from the dictionary memory 2 with the feature vector input from the feature extraction unit 1 and calculates the similarity between them, for example, a distance value (candidate distance value, hereinafter, distance value unless otherwise specified). Will be abbreviated), and the character categories that will be the upper candidates will be sorted in descending order of similarity (in order of increasing distance value, increasing distance value). Then, category information of higher candidates and distance value information including a column of distance values at that time are output as recognition results.

As an example of this recognition result, FIG. 4 shows a sequence relating to distance value information when a Mahalanobis distance (distance measure that compensates for statistical correlation between features) is used, that is, a distance value sequence. In FIG. 4, the x-axis direction shows the candidate ranks of the recognition results (from the first rank to the 10th rank), and the y-axis direction shows the magnitude of the distance value at that time in each candidate rank. It means that the candidate category is likely. Further, pattern A shows an example of a pattern in which similar categories exist, and pattern B shows an example of a pattern of small stroke number characters.

Some character recognition devices use a distance value as the evaluation value of the recognition result and perform post-processing such as word matching and character string matching based on this evaluation value. For example, in the "word reading method" of Japanese Patent Application Laid-Open No. 63-121989 and the "character recognition result correction method" of Japanese Patent Application No. 6-50865, a value indicating the certainty of each candidate character in the recognition result is used. It describes how post-processing is performed using the distance value.

[0005]

Usually, when a candidate category sequence and a distance value sequence at that time are obtained for a character pattern to be recognized, differences such as presence / absence of similar categories, kanji / non-kanji, etc. Depending on the category, the distribution shape of the distance values, such as the average value and the variance value, differs. For example, in the example of FIG. 4, when the similar category exists like pattern A, the variance value of the distance values is small, and in the case of small stroke number characters like pattern B, the distance value tends to be small overall. .

That is, the distance value as a recognition result tends to have various distribution shapes depending on the type of character category. Therefore, in the conventional method of performing the post-processing based on the distance value, there is a problem that the correct reading rate and the erroneous reading rate vary depending on the type of the character category, and a stable post-processing result cannot be obtained.

In view of such a problem, an object of the present invention is to obtain a new evaluation value of a recognition result for deriving a highly accurate and stable post-processing result without excessively depending on the type of character category. To provide.

[0008]

The candidate accuracy is used as an evaluation value indicating whether or not the result itself output from the character recognition device of FIG. 3 is reliable, and how reliable each of the plurality of candidate categories is. There is a viewpoint of using it as an evaluation value that represents From the former point of view, the candidate accuracy that is output is one for the character recognition result, and can be a reference for reject determination, for example. The applicant of the present invention has proposed this in "Character recognition method and character recognition device using the same" in Japanese Patent Application No. 5-75071, and the desired effect is obtained. On the other hand, in the present invention, in order to stably perform the post-processing performed based on the candidate certainty, the candidate certainty is given to each candidate category in the character recognition result. This is from the latter point of view.

That is, in the present invention, the distance vector for each character category is calculated by comparing the feature vector representing the characteristics of the recognition target character pattern with the standard vector prepared in advance for each character category, and the distance value is calculated. After outputting the category information for the character categories that are upper candidates in the ascending order and the distance value information including the distance value group corresponding to each character category as the recognition result, the difference value data between the adjacent distance values is calculated from the distance value information. create. Then, the column of the created difference value data and the distance value information are internally producted by the discrimination coefficient value for distinguishing the correct reading tendency and the misreading tendency of each category to obtain the candidate accuracy for the first-ranked candidate category of the recognition result. decide. Further, when determining the candidate certainty for the i-th candidate category, for example, the candidate certainty of the first-ranked candidate category is multiplied by the ratio of the first-ranked candidate distance value and the i-th candidate distance value. Since the present invention mainly focuses on determining the optimal candidate probability for each candidate character, the candidate probability for the i-th candidate character may be determined by a method other than the above calculation.

Such a method can be implemented by the character recognition device having the following configuration. (1) A feature extraction unit that creates a feature vector representing a feature of a recognition target character pattern; (2) a dictionary memory that stores a standard vector for each character category in advance; (3) the feature vector and the dictionary memory. An identification unit that outputs, as a recognition result, category information for character category groups that are upper candidates in the ascending order of distance values obtained by collating with a standard vector, and distance value information that is a sequence of distance values corresponding to individual character categories. (4) A coefficient memory that stores the discrimination coefficient value for each category for distinguishing the correct reading tendency and the erroneous reading tendency of each character category, (5) Adjacent that expresses the distribution shape of the distance value string from the distance value information. Difference value data creating means for creating difference value data between distance values, (6) the created character string of the difference value data and the distance value information in the corresponding character category in the coefficient memory First candidate probability determining means for determining the candidate probability of the recognition result for the first-rank candidate category by inner-product with the discrimination coefficient value of (7), the first-rank candidate distance value and the i-th candidate distance for the first-rank candidate probability. A second determining a candidate probability for the i-th candidate character of the recognition result based on a ratio with a value;
Candidate probability determination means. The difference value data creation means and the first and second candidate probability determination means are realized by, for example, a programmed information processing device.

[0011]

According to the present invention, the discrimination coefficient value for each category for distinguishing the correct reading tendency and the erroneous reading tendency is stored in the coefficient memory on the basis of the parameters collected in advance. Then, the difference value data between the adjacent distance values is created from the distance value information output from the identification unit, and the discrimination coefficient value corresponding to the first-ranked candidate category is selected from the coefficient memory from the category information of the recognition result. The coefficient value is inner-producted with the distance value information and the sequence of difference value data (hereinafter, difference value data sequence) by the first candidate certainty determining means to determine the candidate certainty for the first-ranked candidate category. At this time, the difference value data string represents the distribution shape of the entire distance value string. Then, the second candidate probability determining means determines the candidate probability for the i-th candidate character by multiplying the first-rank candidate probability by the ratio of the first-rank candidate distance value and the i-th candidate distance value. As described above, the correctness of the first-rank candidate probability can be increased by reflecting the correct reading / misreading tendency for each category and handling the distribution shape of the distance value sequence in a composite manner. As a result, the candidate probability of the second and subsequent candidate characters is obtained using this first-rank candidate probability, which is more valid and stable than the first-rank candidate distance value, and as a result, a more optimal candidate for each candidate character. The accuracy can be derived.

[0012]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a block diagram showing the configuration of a character recognition apparatus according to an embodiment of the present invention. The same elements as those of FIG. 3 showing the configuration of the conventional apparatus are designated by the same reference numerals. The character recognition device according to the present embodiment has an identification unit 3
The first candidate probability calculating unit 6 is provided in the subsequent stage. In the output of the identification unit 3, the category information is directly led to the distance value information, and the distance value information is led to the first candidate probability calculation unit 6 via the difference value creation unit 4. The coefficient memory 5 is connected to the first candidate certainty calculating unit 6, and the distance value information and the candidate certainty of the output thereof are guided to the second candidate certainty calculating unit 7, This second
The distance value information and the candidate accuracy output from the candidate accuracy calculation unit 7 and the category information output from the identification unit 3 are the final recognition result by the character recognition device.

Here, the difference value creating section 4 creates the distance value information output from the identifying section 3, that is, a plurality of difference value data representing the distribution shape of the entire distance value sequence. The discriminant coefficient value learned by using discriminant analysis so as to discriminate the correct reading tendency and the erroneous reading tendency for each category is stored in advance. In addition, the first candidate probability calculating unit 6 uses a predetermined discriminant function from the discriminant coefficient values determined according to the first-order candidate categories of the plurality of difference value data strings created by the difference value creating unit 6 and the recognition result. The inner product is calculated and the discriminant function value is output as the first-order candidate probability. The second candidate probability calculating unit 7 calculates the second and subsequent candidate probabilities based on the first candidate probability and the candidate distance value sequence.

FIG. 2 is a diagram showing an example of the contents of the coefficient memory 5. Reference numeral 11 is a category column, and 12 is a discrimination coefficient column for each category. It is assumed that the coefficient value in the discrimination coefficient column 12 is preset. Next, a candidate certainty determination process by this character recognition device will be described with a specific example. However, here, it is assumed that the identification unit 3 outputs category information C and a distance value sequence (distance value information) d as shown below.

[0015]

## EQU1 ## C = {C _i | i = 1, 2, ..., I} d = {d _i | i = 1, 2, ..., I}

However, C _i is the i-th candidate category, d _i is the i-th candidate distance value, and I is the maximum number of target candidate categories. First, the difference value creating unit 4 receives the distance value sequence d and obtains the difference value between the individual adjacent distance values, that is, the difference value data sequence d ′. An example of this difference value data string d'is shown below.

[0017]

## EQU2 ## d '= {d _i ': d _{i + 1} −d _i | i = 1, 2, ...
., I-1}

However, d _i 'is difference value data corresponding to the i-th candidate category. Next, the first candidate probability calculation unit 6
At, the discriminant coefficient value w used for the calculation of the first-rank candidate probability is selected and extracted from the coefficient memory 5 by focusing on the first-rank candidate category C ₁ . For example, if the first-ranked candidate category C ₁ is “A”, the discrimination coefficient value w is represented by the following equation from FIG.

[0019]

## EQU3 ## w = (w ₄₁ , w ₄₂ , ..., W _4m , w _{4 (2I-1)}

Then, the discriminant function f (w, d, d ') using the distance value sequence d of the recognition result, the distance value sequence d'obtained by the difference value creating section 4, and the discriminant coefficient value w as parameters. Then, the 1st place candidate probability of the recognition result is calculated as follows.

[0021]

[Number 4] # 1 candidate confidence = f (w, d, d ') = w 41 · d 1 + w 42 · d 2 + ··· + w 4I · d I' + w 4 (I + 1) · d 1 ' + w _{4 (I + 2)}・ d ₂ '+ ・ ・ ・ + w _{4 (2I-1)}・ d _I-1 '

It should be noted that the quoting of the distance value and the creation of the difference value data string d'as described above are performed on a sample basis, and other embodiments are also conceivable. Then, the second candidate probability calculating section 7 calculates the second and subsequent candidate probabilities as follows from the first-rank candidate probability and the distance value sequence d.

[0023]

Equation 5] i-position candidate confidence = 1-position candidate confidence × (d ₁ / d _i) where, 2 ≦ i ≦ I

After that, the 1st to Ith candidate probabilities are applied to the post-processing in the subsequent stage as the quantified values of the probabilities of the individual candidate characters.

As described above, in this embodiment, when quantifying the certainty of the recognition result from the discriminant function, the discriminant coefficient w corresponding to the tendency of each candidate category C _i is used, and further,
Since the distance value sequence d and the difference value data sequence d ′ are obtained and the distance value information is expressed in a composite manner so that the distribution shape of the entire distance value sequence can be understood, only the distance value information that excessively depends on the category as in the past. The relevance and stability of the candidate accuracy are significantly improved compared to the case expressed by. Thus, for example, in post-processing such as character matching that uses the distance value information as it is in the character recognition result, by using the individual candidate accuracy derived above instead of the distance value information, the correct reading rate is improved and the misreading rate is increased. Can be reduced. Therefore, when the method of the present embodiment is applied to the character recognition device used in the data entry work in a private institution or a public institution, the correction time of the operator is shortened and the manual intervention is reduced. .

Since the above-mentioned citation of the distance value sequence d and the creation of the difference value data sequence d'are carried out on a sample basis,
It is needless to say that the contents of the present embodiment are not limited to the contents of the present embodiment and may be other contents.

[0027]

As is apparent from the above description, according to the present invention, there is an effect that the optimum candidate probability can be obtained for each candidate category for the character pattern to be recognized. Further, since the distance value information is treated in a composite manner so that the distribution shape of the entire distance value sequence can be grasped, there is an effect that the most appropriate candidate probability can be stably calculated for the candidate category. Thereby, a new evaluation value for obtaining a highly accurate and stable post-processing result can be derived without depending on the category, and the conventional problems can be solved.

[Brief description of drawings]

FIG. 1 is a block configuration diagram of a character recognition device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of an example of contents of a coefficient memory included in the character recognition device according to the present embodiment.

FIG. 3 is a basic configuration diagram of a conventional character recognition device.

FIG. 4 is an actual measurement diagram showing a relationship between candidate ranks and distance values related to character recognition processing by a conventional character recognition device.

[Brief description of reference numerals]

 1 Feature Extraction Unit 2 Dictionary Memory 3 Discrimination Unit 4 Difference Value Creation Unit 5 Coefficient Memory 6 First Candidate Probability Calculation Unit 7 Second Candidate Probability Calculation Unit

Claims (4)

[Claims] 1. A candidate distance value for each category is calculated by collating a feature vector representing a feature of a recognition target character pattern with a standard vector prepared in advance for each character category.
In a character recognition method, which includes a step of outputting, as a recognition result, category information for character category groups that are upper candidate categories in ascending order of candidate distance values and distance value information that is a sequence of candidate distance values corresponding to individual character categories. , A step of creating difference value data between adjacent candidate distance values from the distance value information, and distinguishing the created column of the difference value data and the distance value information from the correct reading tendency and the misreading tendency of each character category. For determining the candidate probability for the highest (hereinafter, first) candidate category of the recognition result by inner product with the discriminant coefficient value for the character recognition method. 2. The character recognition method according to claim 1, wherein the candidate probability of the determined first-rank candidate category is multiplied by a ratio between the first-rank candidate distance value and the i (i> 1) -rank candidate distance value. And a step of deriving a candidate probability of the i-th candidate category of the recognition result. 3. A feature extraction unit that creates a feature vector representing a feature of a recognition target character pattern, a dictionary memory that stores a standard vector for each character category in advance, and the feature vector and a standard vector stored in the dictionary memory. An identification unit that outputs, as a recognition result, category information for a character category group that is a higher-ranked candidate in ascending order of candidate distance values obtained by collating with and distance value information that is a sequence of candidate distance values corresponding to individual character categories. In a character recognition device including, a coefficient memory storing a discrimination coefficient value for each category for distinguishing a correct reading tendency and a misreading tendency of each character category, and a difference between the adjacent candidate distance values from the distance value information. Difference value data creating means for creating the value data, and the created character string of the difference value data and the distance value information in the corresponding character category in the coefficient memory. Character recognition apparatus comprising: the first candidate reliability determining means for determining a candidate confidence for first candidate category of the recognition result by determining coefficients and dot product of Li, the. 4. The character recognition device according to claim 3, wherein the first-rank candidate probability is multiplied by a ratio of the first-rank candidate distance value and the i (i> 1) -rank candidate distance value to obtain the i-th candidate of the recognition result. A character recognition device further comprising a second candidate probability determination means for determining a candidate probability for a character.