JPH06274702A - Character recognizing device - Google Patents

Abstract

PURPOSE:To improve the character recognizing precision by storing the correcting frequency information on the registered words and deciding the order of candidate words according to the correcting frequency of those words registered by users. CONSTITUTION:An after-processing part 16 holds a knowledge dictionary, and a word register part 14 registers additionally the word information desired by users into the knowledge dictionary. The correcting frequency information contained in the knowledge dictionary are increased by a control part 24 every time the registered information on a word is corrected. Then the part 16 supplied the character recognizing result through a recognizing part 12 and calculates the evaluation value of a character string consisting of the combination of candidate characters equivalent to a single unit of knowledge processing. Thus a character string whose elavulation value is less than the prescribed threshold value is sent to a result editing part 18 as a candidate word. If plural character strings have the same evaluation value, the candidate order of words registered by users that satisfies X<S is defined as 1 (X: correcting frequency information value of registered words, S: prescribed threshold value).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognizing device which can accurately and quickly process a form or a document and has a good recognition performance.

[0002]

2. Description of the Related Art Conventionally, a knowledge dictionary has been used to improve the recognition rate of handwritten characters. The information held by the knowledge dictionary is the information of the recognition target word, the context information, and other information. The user of the character recognition device can add and register word information that is not prepared in advance in the knowledge dictionary (hereinafter, the word registered by the user in this way is referred to as a user registered word). Hereinafter, as a conventional character recognition technique using word information of a knowledge dictionary, for example, a document:
The explanation will be given based on the technology shown in pp. 5-326 of the Proceedings of the IEICE General Conference, 1982.

First, a predetermined area of a form or a document is optically scanned and an optical signal from the paper surface is photoelectrically converted to obtain image data of the form or the document. Then, a character pattern to be recognized is cut out from the image data. Characters to be recognized are recognized based on the cut-out character pattern, and one or more candidate characters are obtained as a recognition result.

Then, regarding the character for one unit of knowledge processing,
A character string is created by combining one or more candidate characters obtained for each character, and the character string and a word in the knowledge dictionary are used by using the candidate rank or similarity assigned to each candidate character of the character string. The evaluation value between is calculated. Further, one or a plurality of candidate words with candidate ranks assigned according to the evaluation value of this character string are displayed. However, when the evaluation value of the character string is out of the predetermined range, a candidate character string formed by combining the candidate characters having the candidate rank of 1 is displayed.

If there are a plurality of candidate words having the same evaluation value, the user-registered word is generally given the highest priority, and the candidate ranks of the other words are determined according to the order registered in the knowledge dictionary.

If the candidate word having the first candidate rank is wrong, the correct answer word is directly input from the keyboard, or the candidate words having the second or lower candidate rank are displayed and if there is a correct answer word among them. Correction is made by selecting it.

[0007]

In the above character recognition technique, since the user-registered word has the highest priority when there are a plurality of candidate words having the same evaluation value, the word which is used less frequently and has many similar words. However, if such a word is registered as a user-registered word, the number of errors in the candidate word having the first candidate rank increases due to the influence of the word, and there is a problem that the character recognition performance deteriorates.

For example, consider a case where a correct answer word is obtained by calculating an evaluation value from a candidate character obtained by recognizing a place where "Abe" is entered in the name column of a certain form. If "A" is recognized here, the first and second characters in the candidate rank are "A" and "Kawa", respectively, and when the "part" is recognized, the first and second characters are candidate ranks. Are “na” and “part” respectively, and “Abe” is prepared in the knowledge dictionary in advance, and “kawana” is a word additionally registered by the user (“Ana”). "Na" is not present in the knowledge dictionary). Then, the evaluation values of "Abe" and "Kana" are equal, and although "Abe" is a word that appears much more frequently than "Kana", the user-registered word "Kana" takes precedence. Will be done.

In order to solve this problem, the present invention provides a character recognition apparatus in which the candidate rank of candidate words is determined according to the frequency of correction of user-registered words to improve the recognition accuracy. To aim.

[0010]

In order to solve the above-mentioned problems, a character recognition device according to the present invention includes a recognition unit which outputs a recognition result of a character pattern cut out from image data of a quantized form or document, A word registration unit for registering word information in a knowledge dictionary; a post-processing unit that holds the knowledge dictionary and outputs a knowledge processing result based on the recognition result; a result editing unit that edits the knowledge processing result; A character recognition device including a display unit for displaying a knowledge processing result and an editing result, an input unit for inputting a correct word, a result editing unit, a display unit, and a control unit for controlling operations of the input unit. In the above, the knowledge dictionary stores frequency information related to the correction frequency of registered words, the control unit updates the frequency information, and the post-processing unit selects candidates according to the frequency information in the subsequent processing. Characterized in that it defines the candidate ranking of the word.

[0011]

According to the present invention, the post-processing unit determines the candidate rank of the candidate words according to the correction frequency of the user-registered words, so that the registered words that are less frequently used are preferentially selected. The above problem is solved.

[0012]

Embodiments of the present invention will be described below with reference to the drawings. It should be noted that the drawings are only schematically shown to the extent that the present invention can be understood.
The arrangement position, dimensions, input / output signals, and connection relationship are not limited to the illustrated example.

FIG. 1 is a functional block diagram for explaining one embodiment of the present invention. Character recognition device 10 of this embodiment
Is a recognition unit 12 that holds a recognition dictionary, cuts out character patterns from quantized image data of a form or document, outputs candidate characters of the cut-out character patterns, and a word desired by a user of the character recognition device. A word registration unit 14 that registers information, a post-processing unit 16 that holds a knowledge dictionary and outputs a knowledge processing result of a recognition result, a result editing unit 18 that corrects and confirms the recognition result, a knowledge processing result and an editing result. And a display unit 20 for displaying a correct answer word, and an input unit 22 for inputting a correct word.
The control unit 24 controls the operation of (1) and updates the frequency information related to the correction frequency of the user registration word. Further, in FIG. 1, reference numeral 26 is a photoelectric conversion unit that outputs quantized image data of a form or document, and 28 is an image memory that stores the image data from the photoelectric conversion unit 26.

FIG. 2 shows the format of the stored contents in the knowledge dictionary for one word. 3 in the figure
Reference numeral 0 is word information, 32 is a flag indicating whether or not the word is a user-registered word (hereinafter referred to as user flag), and 34 is frequency information related to correction frequency (hereinafter referred to as correction frequency information).

FIG. 3 shows an example of a form. In FIG. 3, 36 is an example of a form in which an address is described.
Reference numeral 8 is an entry frame for designating a character writing area.

Hereinafter, this embodiment will be described in detail with reference to FIGS. 1, 2 and 3. The photoelectric conversion unit 26 optically scans a predetermined reading range on a form or a document, photoelectrically converts an optical signal L from the form or the document, outputs image data quantized into black and white binary, and outputs the image memory. 28 stores this image data.

The recognition unit 12 cuts out a character pattern from the image data of the image memory 28, and extracts various characteristics relating to the character to be recognized from the cut-out character pattern.
Then, the characteristics of the character pattern are compared with the standard character pattern, and the candidate character is output. For one character, one or more candidate characters are obtained as a recognition result, and the candidate character is 1
In the case of the number of candidate characters, the candidate rank 1 is attached to the candidate character and outputted, and in the case of a plurality of candidate characters, the candidate rank assigned to each candidate character according to the similarity is attached to the candidate character. Output.

The word registration unit 14 additionally registers the word information desired by the user of the character recognition dictionary in the knowledge dictionary.

The post-processing unit 16 holds a knowledge dictionary, and the user flag 32 in the knowledge dictionary is set to 0 when the word information 30 is prepared in advance, and is set to 1 when the word information 30 is additionally registered by the user. Has been done. The correction frequency information 34 has an initial value of 0, and each time the word of the word information having the candidate rank of 1 is corrected by the result editing unit 18, the control unit 24 increments by 1 each time.

Further, the post-processing section 16 carries out the following knowledge processing based on the recognition result from the recognition section 12. Knowledge processing 1
When a recognition result of characters for a unit (for example, a recognition result of the area 38 for describing a prefecture name in the form 36 shown in FIG. 3) is input, a character string formed by combining candidate characters of each character for knowledge processing is used as a word. It is checked whether or not a word corresponding to a character string composed of candidate characters exists in the word information by collating with the information word. Then, when the character string A that matches the word of the word information is detected from the character strings formed by combining,
The evaluation value J of the character string A is calculated. Here, if S is the sum of the candidate ranks given to the candidate characters of the character string and N is the total number of characters that make up the character string, the evaluation value J is expressed as, for example, J = S ÷ N. be able to.

Whether or not the word and the character string A match is determined by, for example, whether or not all the character codes of the characters at the corresponding positions of the word and the character string A match. Then, when all the character strings formed for one unit of knowledge processing have been checked against the word information, the character string A whose evaluation value J is within a predetermined threshold is sent to the result editing unit 18 as a candidate word.
For each candidate word, the candidate rank is determined in order from the smallest evaluation value. If there are multiple character strings A with the same evaluation value J,
And if the user registration word is not included in it,
Candidate ranks are determined in the order of registration in the knowledge dictionary. Further, when there are a plurality of character strings A having the same evaluation value J and the user registration word is included therein, in the present embodiment, the candidate order is determined as follows according to the correction frequency information. To do.

Value X of correction frequency information of the user registered word
And the predetermined threshold value S is X <S, the candidate rank of the user registration word is set to 1, and the candidate ranks of the candidate words having the same evaluation value are determined in the order of registration in the knowledge dictionary.
If X ≧ S, the user-registered word is not prioritized, and the candidate ranks of candidate words having the same evaluation value are all determined in the order of registration in the recognition dictionary.

If only one character string A is detected when all the character strings for one unit of knowledge processing are matched with the words in the word information, the candidate editing order of the character string A is set to 1 and the result editing section Send to 18. Also, if no character string A is detected when all the character strings for one unit of knowledge processing have been matched with the words of the word information, the candidate rank of each character for one unit of knowledge processing is 1 A character string formed by combining characters is sent to the result editing unit 18 as the knowledge processing result.

The result editing unit 18 corrects the candidate word or candidate character whose candidate rank is 1 output by the post-processing unit 16 when it is incorrect.

The display unit 20 displays candidate words, candidate characters, and edited results.

The input unit 22 inputs a correct word when a candidate word having a candidate rank of 1 is incorrect. If there is one candidate word, or if there are multiple candidate words and their candidate rank is 2
If the correct word is not included in the following candidate words, enter the correct word using a keyboard or the like. When there are a plurality of candidate words and the candidate word having the candidate rank of 2 or less includes the correct answer, the correct answer word is selected with the mouse or the like.

The control unit 24 includes a result editing unit 18 and a display unit 2.
0, and controls the operation of the input unit 22, and also the result editing unit 1
Whenever the user-registered word whose candidate rank is 1 is corrected in 8, the correction frequency information in the knowledge dictionary is incremented by 1 each time.

The present invention is not limited to the above-mentioned embodiment, and therefore, the configuration, operation, processing content, input / output signal and numerical condition of each component may be arbitrarily changed. For example, when there are a plurality of candidate words having the same evaluation value and the user-registered words are included in the candidate words, two thresholds such that S1 <S2 instead of the threshold S in the determination of the candidate rank. S1 and S2 may be provided stepwise.
At this time, when the correction frequency information X is X <S1, the user registration word is given the highest priority, and when S1 ≦ X <S2, the user registration word is set to an intermediate rank of the candidate word group. Further, when X ≧ S2, the candidate rank is determined such that the user registration word has the lowest rank. Further, three or more threshold values S may be set stepwise.

In addition, when there are a plurality of candidate words having the same evaluation value and the user-registered words are included in the candidate words, the user sets the threshold S in determining the candidate rank at the time of registering the words. You may

Further, when there are a plurality of candidate words having the same evaluation value and the user-registered words are included in the candidate words, the threshold value S in the determination of the candidate rank is set to another value which is output at the same time in the knowledge processing. It may be changed automatically according to the number of candidate words, correction frequency information, context information, and the like.

In the above-described embodiment, the user has additionally registered the word information in the knowledge dictionary in which the word information is stored in advance in the word registration unit 14. However, another user-specific knowledge dictionary is prepared and stored therein. You may make it register necessary word information.

In the above-mentioned embodiment, the evaluation value J is
A value obtained by dividing the sum S of the candidate ranks given to each candidate of the character string by the total number N of the characters forming the character string was used, but instead of the sum S of the candidate ranks, the score corresponding to each candidate rank ( For example, 100 points for the first candidate ranking and 90 for the second candidate ranking
It is also possible to use a sum of (corresponding points that become lower as the candidate rank goes down, such as associating points). Alternatively, instead of the sum S of the candidate ranks, the sum of the appearance frequencies of the candidate characters in the character string (in this case, the appearance frequency is held in advance in the recognition dictionary) may be used. Alternatively, instead of the sum S of the candidate ranks, the distance between the dictionary matrix of the candidate characters and the feature amount of the character pattern corresponding to the candidate character may be obtained and the sum of the distances of the respective candidates of the character string may be used. . Alternatively, instead of the sum S of the candidate ranks, the sum of the appearance frequencies of the candidate characters in the character string and the sum of the candidate ranks may be used.

In addition to the above, the knowledge processing of the post-processing section may be performed as follows. In order to check whether or not a word corresponding to a character string consisting of candidate characters exists in word information, a character string for one unit of knowledge processing is collated with a word of word information in a knowledge dictionary, and the character string and the word The degree of similarity (or degree of disagreement) between the two is calculated. As a word corresponding to the character string, for example, a word whose similarity to the character string exceeds a predetermined threshold value (or a word whose dissimilarity to the character string does not exceed a predetermined threshold value) is detected. Then, (1) When a character string whose similarity exceeds a predetermined threshold (or a character string whose degree of disagreement does not exceed a predetermined threshold) is detected, the maximum similarity (or minimum) of the detected character strings is detected. Of the word information corresponding to the character string of the maximum similarity (or the degree of dissimilarity), and the maximum similarity (or the degree of disagreement) as the evaluation value of the knowledge processing. Output as. (2) Even if all the character strings for one unit of knowledge processing are completely matched with the words of the word information, one character string whose similarity exceeds a predetermined threshold value (or a character string whose dissimilarity degree does not exceed a predetermined threshold value) is If none of them is detected, the character string of the combination of the candidate characters having the first rank of the candidate is output as the knowledge processing result, and the lower limit value of the predetermined similarity degree (or the predetermined upper limit value of the dissimilarity degree). ) Is output as an evaluation value. The lower limit value of the similarity (or the upper limit value of the dissimilarity) indicates that the word corresponding to the character string including the candidate character does not exist in the word information.

Further, in the above-described embodiment, an example of knowledge processing using word information has been described, but the present invention may be applied to a character recognition device that performs knowledge processing using context information and other knowledge information.

[0035]

As described above, according to the present invention, the candidate rank is determined according to the correction frequency of the user-registered word, so that no matter what word is additionally registered, this device can be continuously used. This makes it possible to prevent the recognition performance from deteriorating. As a result, even non-specialists can easily register words. Therefore, it is possible to provide a character recognition device having improved recognition performance so that a form or a document can be processed accurately and quickly.

[Brief description of drawings]

FIG. 1 is a functional block diagram showing a configuration of an exemplary embodiment of the present invention.

FIG. 2 is a diagram showing a format of stored contents in a knowledge dictionary.

FIG. 3 is a diagram showing an example of a form.

[Explanation of symbols]

 10 character recognition device 12 recognition unit 14 word registration unit 16 post-processing unit 18 result editing unit 20 display unit 22 input unit 24 control unit 26 photoelectric conversion unit 28 image memory

Claims (1)

[Claims] 1. A recognition unit for outputting a recognition result of a character pattern cut out from image data of a quantized form or document, a word registration unit for registering word information in a knowledge dictionary, and holding the knowledge dictionary. To input a correct word, a post-processing unit that outputs a knowledge processing result based on the recognition result, a result editing unit that edits the knowledge processing result, a display unit that displays the knowledge processing result and the edited result, In a character recognition device comprising an input unit of, a result editing unit, a display unit, and a control unit that controls the operation of the input unit, the knowledge dictionary stores frequency information related to the correction frequency of registered words. The character recognition device, wherein the control unit updates the frequency information, and the post-processing unit determines a candidate rank of candidate words according to the frequency information in subsequent processing.