JPH05298489A - System for recognizing character - Google Patents

Abstract

PURPOSE:To improve recognition precision by compensating a collation value at every character category and outputting the character category candidate of a character pattern which is read from a recognition object range. CONSTITUTION:Character category frequency memory parts 7-1 and 7-2 are respectively provided in accordance with the respective recognition object ranges when plural recognition object ranges exist in a slip. Then, a memory selecting part 8 selects one character category frequency memory part corresponding to the present recognition object range. Moreover, a collation value compensating part 9 compensates the collation value at every character category, which is obtained by an identification collating part 3, based on frequency information at every character category. The collation value compensated by the collation value compensating part 9 is inputted to a candidate character category sorting part 6. When the collation value by the classification of the character category is inputted, the collation value compensating part 9 controls the memory selecting part, permits the memory selecting part to select the character category frequency memory part 7-1 corresponding to the present recognition object range and reads frequency information by the classification of the character category.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method suitable for application to OCR and the like, and in particular, character recognition for improving the misrecognition probability by narrowing the target character category to a small number in the recognition process. Regarding the scheme.

[0002]

2. Description of the Related Art FIG. 4 is a block diagram showing a configuration of a main part of a character recognition apparatus using a conventional character recognition method. In the figure, reference numeral 1 is a character pattern which is obtained by binary-quantizing a character string described in a recognition target range on a form in a character unit by means of a scanner or a character cutout, and identifies a feature vector for identification. A feature extraction unit 2 to be created is a dictionary memory unit that stores a standard feature vector representing a standard feature of a large number of character patterns to be recognized as collation information for each character category, and 3 is stored in the dictionary memory unit 2. The identification collating units 4-1 and 4-2 identify the collation information sequentially read in character category units, collate with the feature vector, and output a collation value representing the similarity of the feature vector to the collation information in character category units. Regarding the matching information used by the matching unit 3 in the matching process, a character category limit in which information of availability (possible = 1, not = 0) is stored for each character category. A memory unit, the character category limited memory unit 4-1 and 4-2 when the recognition target range of a plurality of locations existing on the form, provided corresponding to each recognition target range. However, the recognition target range of the common character category is also used. In the figure, only those corresponding to the two recognition target ranges are shown, and the character category limited memory units in the other recognition target ranges are omitted. FIG. 5 shows an example of the stored contents when the character category is 3000 types.

Reference numeral 5 is a character category limited memory section 4-1, 4
-2, a memory unit corresponding to the current recognition target range is selected, and the usability information sequentially read in character category units from the selected memory unit is input to the identification collation unit 3 and the identification collation unit 3 In the memory selection unit for performing the matching operation only in the character category whose availability information is “1”, 6 sorts the matching values of all the character categories obtained by the matching operation in the identification matching unit 3, The candidate character category sorting unit outputs the character category candidates of the character pattern read from the current recognition target range. The character category candidates output from the candidate character category sorting unit 6 are a word matching unit and a recognition result display unit (not shown). Etc., and is used for realizing selection of a candidate character category string that matches a word character string that may exist and correction by the operator of the character reading device.

In such a configuration, when recognizing a character string described in a recognition target range on a form, the character string is binary-quantized in character units by means such as a scanner or a character cutout and characterized as a character pattern. Input to the extraction unit 1.

Then, the feature extraction unit 1 analyzes the features of the input character pattern and creates a feature vector for identification. This feature vector is input to the identification and collation unit 3.

When the feature vector of the character pattern is input, the identification collation unit 3 sequentially reads out the collation information from the dictionary memory unit 2 for each character category. In addition, the memory selection unit 5
Of the information on availability of use stored in the character category limited memory unit 4-1 by controlling the memory selection unit 5 to select the character category limited memory unit (for example, 4-1) according to the current recognition target range. Information about availability of use corresponding to the character category of the collation information read from the dictionary memory unit 2 is synchronously read. Then, the feature vector input from the feature extraction unit 1 is collated with the collation information read from the dictionary memory unit 2 only in the character category for which the usability information is “1”, and the feature for that collation information is collated. A matching value indicating the similarity of the vector is output.

[0007] For example, when the availability information is "1" only for hiragana, only the collation information of hiragana is collated, and the collation value for each character of hiragana is output.

This matching value is used as the candidate character category sorting unit 6
Entered in. The candidate character category sorting unit 6 sorts the matching values of all the character categories obtained by the matching operation in the identification matching unit 3, and outputs the character category candidates of the character pattern read from the current recognition target range.

With such a configuration, when the character category described in the recognition target range is limited to, for example, only hiragana, if the information of availability is set to "1" only in hiragana, When recognizing a character string in the recognition target range, it is possible to perform collation excluding character categories other than hiragana, and the misreading probability can be reduced.

[0010]

However, in the above-mentioned conventional technique, it is possible to specify whether or not to use character categories, but there are many character categories to be recognized, such as when recognizing Chinese characters. In this case, since the degree of narrowing down the range of the character category is small, there is a problem in that the effect of reducing the misreading probability does not increase.

The present invention has been made to solve such a problem. Even when there are many character categories to be recognized, the range of character categories is narrowed down to improve the effective recognition accuracy. The object of the present invention is to provide a character recognition method capable of performing the character recognition.

[0012]

In order to achieve the above object, the present invention analyzes a feature of a character pattern read from a recognition target range and creates a feature vector for identification, and a feature extraction unit of the recognition target. A dictionary memory means storing collation information representing standard characteristics of a large number of character patterns for each character category, and collation information stored in this dictionary memory means is sequentially read out in character category units and collated with the characteristic vector to collate. Identification matching means that outputs a matching value indicating the similarity of the feature vector to information in character category units, and the matching values obtained in all character categories are sorted, and character category candidates of the character pattern read from the recognition target range are output. In the character recognition method including means for sorting candidate character categories, Character category frequency memory means for storing frequency information regarding appearance probability for each gori, and matching value correction means for correcting the matching value for each character category obtained by the identification and matching means based on the frequency information for each character category are provided. The matching value corrected by the matching value correcting means is input to the candidate character category sorting means, and the character category candidates of the character pattern read from the recognition target range are output.

[0013]

According to the above means, the matching value for each character category is corrected based on the frequency information regarding the appearance probability of each character category that appears in the recognition target range, and the corrected matching value is input to the candidate character category sorting means. Then, since the character category candidates of the character pattern read from the recognition target range are output, the character category candidates having a high appearance probability are prioritized,
The degree of narrowing can be increased. As a result, effective recognition accuracy can be improved.

[0014]

EXAMPLES The present invention will be described in detail below with reference to illustrated examples.

FIG. 1 is a block diagram showing an embodiment of the configuration of the main part of a character recognition device using the character recognition system of the present invention. In the figure, the same parts as those in FIG. 4 are indicated by the same symbols, and the description thereof will be omitted. In FIG. 1, 7-1, 7-
Reference numeral 2 denotes a character category frequency memory unit that stores frequency information regarding the appearance probability of each character category that appears in the recognition target range. The character category frequency memory units 7-1 and 7-
2 is when there are multiple recognition target areas on the form,
It is provided corresponding to each recognition target range. However,
The recognition target range of the common character category is shared. In the figure, only those corresponding to the two recognition target ranges are shown, and the character category frequency memory units of the other recognition target ranges are omitted.

Reference numeral 8 is a memory selection unit for selecting one character category frequency memory unit according to the current recognition target range. Reference numeral 9 is a matching value for each character category obtained by the identification and matching unit 3 and frequency information for each character category. Is a collation value correction unit that corrects the collation value based on the collation value.

In such a structure, the form 10 in which the character string to be recognized is described has a structure as shown in FIG.
It is assumed that the form code column 11, the place name column 12, the address column 13, and the family name column 14 in the form 10 are read and recognized.

When recognizing "2" of the character string "66 address 2" in the address column 13 among these, the identification matching unit 3 obtains a matching value as shown in FIG. 3 (a). That is, although there are characters similar to "2" such as "Z" and "Otsu", the standard feature vector of the number "2", the alphabet "Z", and the Chinese character "Otsu" can be read from the address column 13. The matching value representing the degree of similarity of the numeral "2" is, for example, "2" = 2.
It is obtained with numerical values such as 0, “Z” = 15, and “B” = 10. Here, it is shown that the smaller this numerical value is, the closer it is to the standard feature vector.

The collation value expressed as such a numerical value is input to the collation value correction unit 9. When the collation value for each character category is input, the collation value correction unit 9 controls the memory selection unit 5 so that the memory selection unit 5 stores the character category frequency memory unit according to the address column 13 which is the current recognition target range. (For example, 7-
1) is selected, and the character category frequency memory unit 7-1
The frequency information for each character category stored in is read.

For example, as shown in FIG. 3B, for the character categories of the numeral "2", the alphabet "Z", and the Chinese character "Otsu", "2" = 1.0, "Z" = 0. 2, "Oto" = 0.1
The frequency information expressed numerically is read out. Here, it is indicated that the larger the numerical value, the higher the appearance probability.

When such frequency information is read, the matching value A is corrected by the frequency information B using a correction arithmetic expression, for example, "matching value A / frequency information B = corrected matching value C".

As a result, the corrected collation value C as shown in FIG. 3C is obtained. That is, according to the collation value output from the identification and collation unit 3, the number “2” read from the address column 13 is similar to the Chinese character “Otsu”, but the address column 13 has the Chinese character “ Since the probability that "B" appears will be small, it is corrected by the frequency information B representing the appearance probability and the candidate rank is lowered, and instead, the candidate rank of the number "2" is advanced.

Therefore, in this example, character category candidates are output in the order of the number "2", the alphabet "Z", and the Chinese character "Otsu".

Therefore, in the form 10 as shown in FIG. 2, if the frequency information corresponding to the place name column 12 is smaller in frequency information regarding numbers and is larger in frequency information regarding kanji, the category range of candidate characters is reduced. can do. Similarly, in the family name field 14, "prefecture", "city",
By reducing the frequency information about Kanji such as "town" and increasing the frequency information about Kanji such as "husband" and "child" that are often used for surnames, the category range of candidate characters is reduced. be able to.

[0025]

As described above, according to the present invention, the matching value for each character category is corrected based on the frequency information regarding the appearance probability of each character category that appears in the recognition target range, and the corrected matching value is calculated. Since the candidate character category is input to the sorting means, even if the candidate rank is lower than other character categories due to similar glyphs, this rank is corrected according to the appearance probability, and the priority rank is set. It can be output as a character category candidate having a high character. As a result, the degree of narrowing down the character category candidates will increase,
There is an effect that the effective recognition accuracy can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a configuration of a main part of a character recognition device using a character recognition system of the present invention.

FIG. 2 is a diagram showing an example of a form in which a character string to be recognized is described.

FIG. 3 is an explanatory diagram showing examples of matching values, frequency information, and corrected matching values.

FIG. 4 is a block diagram showing a configuration of a main part of a character recognition device using a conventional character recognition method.

FIG. 5 is an explanatory diagram showing an example of usability information for each character category in a conventional character recognition device.

[Explanation of symbols]

1 ... Feature extraction unit, 2 ... Dictionary memory unit, 3 ... Identification and collation unit,
6 ... Candidate character category sorting section, 7-1, 7-2 ... Character category frequency memory section, 8 ... Memory selecting section, 9 ... Collation value correcting section.

Claims (3)

[Claims] 1. A feature extraction means for analyzing a feature of a character pattern read from a recognition target range to create a feature vector for identification, and collation information representing a standard feature of a large number of recognition target character patterns. The dictionary memory means stored for each category and the collation information stored in the dictionary memory means are sequentially read in character category units, collated with the feature vector, and a collation value representing the similarity of the feature vector to the collation information is presented in character category units. In the character recognition method provided with the identification and collation means for outputting to, the collation values obtained in all the character categories, and the candidate character category sorting means for outputting the character category candidates of the character pattern read from the recognition target range, A character character that stores frequency information regarding the appearance probability of each character category that appears in the recognition target range. A category frequency memory means and a matching value correcting means for correcting the matching value for each character category obtained by the identification matching means based on the frequency information for each character category are provided, and the matching value corrected by this matching value correcting means is provided. A character recognition method characterized by inputting to the candidate character category sorting means and outputting a character category candidate of a character pattern read from a recognition target range. 2. The character category frequency memory means is provided corresponding to each of a plurality of recognition target ranges, and the frequency information stored in one character category frequency memory means corresponding to the current recognition target range is read out. 2. The character recognition method according to claim 1, wherein the matching value for each character category obtained by the identification and matching means is corrected. 3. The character category frequency memory means is provided for each recognition target range in which a different character category appears among a plurality of recognition target ranges, and is stored in one character category frequency memory means corresponding to the current recognition target among them. The character recognition method according to claim 1, wherein the frequency information is read and the matching value for each character category obtained by the identification and matching means is corrected.