JP2974145B2 - Correcting character recognition results - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a method for correcting a character recognition result, and particularly to a method for effectively applying character recognition to a document in a language with limited character types such as English. The present invention relates to a method for correcting a character recognition result to be obtained.

[Conventional technology]

In character recognition of multi-font and multi-size English sentences, it is difficult to distinguish uppercase and lowercase letters and symbols (comma and apostrophe, minus and underscore, etc.) by a method that matches the characteristics of character patterns. Misrecognition is likely to occur due to an error in

As a method for reducing such misrecognition, a paper "Character Classification Method Focusing on Character Size and Position in English Document Recognition Processing" published on pp. 1-191 of the IEICE Spring National Convention, pp. 1-191. There is a method described in "

This presupposes that the heights of the characters are the same for each line in the preprocessing stage of character recognition, and extracts the upper and lower reference lines from the histogram of the top and bottom positions of the characters for each line. The characters on the line are classified into a plurality of categories based on the positional relationship between the characters and the upper and lower ends of a circumscribed rectangle of each character image and the height of each character.

[Problems to be solved by the invention]

However, according to the method of performing such character classification in the preprocessing stage and performing recognition based on the character pattern feature only for the determined category, if the character classification is incorrect, correct characters are absolutely excluded from candidates, It becomes extremely difficult to correct the erroneously recognized character by post-processing (for example, word processing).

Further, since the processing is performed in units of lines, it cannot be applied to an English document or the like in which the character size changes in words (spaces and character strings separated by spaces).

Furthermore, since the reference line is extracted from the line-by-line histogram, it is necessary to perform a skew correction process on the recognition document at the time of extracting the reference line, which complicates the preprocessing.

On the other hand, in the post-processing of character recognition, there is a method of identifying a character by comparing a candidate character of a similar pattern that cannot be determined only with the feature parameter with a shape parameter prepared in advance based on past statistics or the like (for example, In this method, accuracy depends on the quality of shape parameters prepared in advance,
Depending on the document to be recognized, it may be necessary to reconstruct the shape parameters.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a method of correcting a character recognition result that can solve the above-described problems.

[Means and actions for solving the problem]

According to the present invention, in a post-processing of character recognition, a first candidate of a character recognition result of a target character string is examined to extract a key character belonging to a specific category (a character category having an absolute uppercase height or an absolute lowercase height). Then, based on the key character, the relationship between the vertical position or height of the key character and the image of another character in the character string and the first character of the other character
According to the category to which the candidate belongs, the recognition result of another character in the character string is corrected.

〔Example〕

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a hardware configuration diagram showing an example of an English character recognition device according to the present invention. A scanner 1 reads an English document to be processed as a binary image, and an image memory 2 stores a binary image of the English document.
Reference numeral 3 denotes a central processing unit (CPU) that executes processes such as character extraction, feature extraction, dictionary matching, and post-processing. 4 is an RO that stores programs and parameters for such processing.
M and 5 are R for storing data obtained by processing.
AM. Reference numeral 6 denotes a dictionary (ROM or RAM) storing templates for character recognition.

The overall flow of the process is as follows. Line extraction and character extraction are performed on the binary image of the English document in the image memory 2. Image features of the extracted individual characters are extracted, compared with the template of each character registered in the dictionary 6, and candidate characters having a small distance are selected, for example, up to the n-th candidate. Stored in RAM5.

In addition, data used in post-processing is also extracted,
Stored in RAM5. This data includes the height positions of the upper and lower edges of the circumscribed rectangle of each character image, or
The height positions of the upper and lower reference lines of the area dividing line of the character are extracted by the method disclosed in the specification and the drawings of -111766. This extraction is performed at the time of character extraction or feature extraction. Further, a flag of the probability of the first candidate of each character by the evaluation function is set as needed.

FIG. 4 shows an example of data obtained for the correct word “Propose”. Here, n = 3.

The evaluation function of the likelihood is given by, for example, equation (1).

E = (D2−D1) × 100 / D1 (1) D1 is the distance of the first candidate D2 is the distance of the second candidate The larger the E value is, the higher the probability of the first candidate is. The distance difference between the first candidate and the second candidate is large, and the distance becomes large as the distance between the first candidates is small.

 If E> threshold, flag = 0. If E ≦ threshold, flag = 1 is set.

Next, post-processing for erroneous recognition correction is executed, which will be described for each embodiment. Although the unit of post-processing is a word that is a character string sandwiched between spaces, a processing range may be a key character search range, a range of a fixed number of characters before, after, or before or after a key character. It is possible.

 Here, the categories of characters are as follows.

A) Absolute uppercase letters (characters with uppercase level height) ABDEFGHIJKLMNQRT bdfhiklt 123456789 b) Absolute lowercase letters (characters with lowercase level height) aemnr c) Similar characters (uppercase and lowercase characters (Similar characters) COSUVWXZ cosuvwxz d) Descender (characters below the baseline: see Fig. 5) PY gjpqy e) Symbol Comma, apostrophe, minus, underscore, etc. Note that the capital letters are absolute capital letters. Character and absolute lowercase height character are selected.

Example 1 The processing flow is as shown in FIG.

From the first character of the word (word), an absolute uppercase character or an absolute lowercase character is searched for the first candidate (process). In the example of FIG. 4, since the first candidate r of the second character is a character having an absolute lowercase height, the second character is used as a key character.

The information is compared with the key characters in order from the first character of the word, and the recognition result is corrected by a method corresponding to the category to which the character (first candidate) belongs (processing,). In the example of the word in FIG. 4, the processing contents are as follows.

(1) First character (correct answer "P") Since the first candidate is the descender p, the difference between the lower end height position of the key character (second character) and the lower end height position of the first character is Compare thresholds.

If the difference is smaller than or equal to the threshold, there is a high possibility that no character appears below the baseline, so the second candidate P is output as a result (the recognition result is corrected).

If the difference is greater than the threshold value, there is a high possibility that a character appears below the baseline, that is, a descender.
The first candidate is output as it is as a recognition result (not corrected).

 In the example, it is the former.

(2) Second character (correct "r") Since this is a key character, the first candidate r is output as it is (no correction).

(3) Third character (correct "o") Since the first candidate is a similar character O (upper case), the character height of the key character (difference between the upper end position and the lower end position) and the third character Is compared with the threshold value.

If the difference is smaller than or equal to the threshold value, a lowercase letter o is forcibly output (corrected).

 If difference> threshold, output uppercase O (no correction).

 In the example, it is the former.

(4) Fourth Character (Correct "p") Since the first candidate is descender p, it is the same as the first character. In this example, the first candidate small letter p is output.

(5) Fifth character (correct answer "o") Since the first candidate is a similar character o, it is the same as the third character. In this example, a lowercase o is output.

(6) Sixth Character (Correct "s") Since the first candidate is a similar character S, it is the same as the third character. In this example, a small letter s is output.

(7) Seventh Character (Correct “e”) Since the first candidate is the absolute lowercase height character e, the difference between the character height of the key character and the character height of the seventh character is compared with a threshold.

If the difference is smaller than or equal to the threshold, the absolute lowercase height character in the candidate is output. Since this example is this case, e is output. In addition,
If there are two or more candidates for absolute lower case height characters, the candidate with the highest ranking is selected.

If difference> threshold, output the absolute capital height character in the candidate. If there are two or more absolute capital height character candidates, the candidate with the highest ranking is selected.

To summarize, the key character information (absolute capital letter height,
If the first candidate is an absolute uppercase letter, an absolute lowercase letter or a descender, the first candidate is an absolute uppercase letter, an absolute lowercase letter, or a descender. According to the result of the judgment corresponding to each category, the correct character is selected from the candidates. However, if the first candidate is a similar character, the first candidate is forcibly converted to uppercase or lowercase according to the result of the specific judgment. Correct characters.

The symbols are as follows. By comparing the height position of the upper or lower end of the character with the height position of the upper or lower end of the key character, a comma and an apostrophe are discriminated, and a minus and an underscore are output, and a more likely one is output.

The above processing is repeated until a space is detected, and when a space is detected, the processing of one word ends.

Although not shown in FIG. 2, if a key character is not found in a word, a flag or the like is added to the word. The correction of such a word is performed in the word processing (correction processing by collation with the word dictionary) in the subsequent stage.

Example 2 The processing flow is as shown in FIG.

In the first embodiment, the upper and lower positions of the circumscribed rectangle of the character are used as the height position of the character in the processing, but in the second embodiment, the reference position determination method of Japanese Patent Application No. 62-111766 is used. Reference positions above and below the region dividing line are used. Other than this, it is the same as the first embodiment.

Embodiment 3 The processing flow is as shown in FIG.

In the first and second embodiments, there is a concern that an incorrect key character may be extracted in the case of a poor quality document. For example, when B and 8 are obtained as recognition candidates for crushed a, B or 8
Is extracted as a key character. The erroneous first candidate of this character is output as it is, but even though the absolute capital height character is selected as a key character, the actual height position information of the character (crushed a) is displayed. Is equivalent to an absolute lowercase character, so other characters cannot be modified normally.

In the third embodiment, key characters are extracted while excluding characters whose likelihood flag is 1 by the evaluation function in the processing, so that erroneous extraction of key characters in the case of a poor quality document and correction mistakes are thereby prevented. be able to.

 Except for this, it is the same as the first and second embodiments.

Embodiment 4 The processing flow is as shown in FIG.

The process performs key character extraction similar to the process in FIG. Here, when a key character whose absolute uppercase character or absolute lowercase character is the first candidate is extracted, the same processing as in the processing in FIG. Make similar corrections.

If the key character is not extracted, the word is corrected in accordance with the uppercase (P, Y) and lowercase (gjpqy) information of the descender in the word and the general rules of the sentence (processing ~).

The following are examples of cases where key characters cannot be extracted.

(1) In the case of a word such as "scoop" that does not include absolute capital height characters or absolute lowercase height characters.

(2) When the flag is 1 even if an absolute uppercase character or an absolute lowercase character is included.

The correction process is performed as follows, for example. In the case of "scoop" in the above (1), if the flag of p is 0, it is assumed that there is a lowercase letter in the word, and the similar characters in the word are corrected to lowercase letters. In the case of the above (2), a similar sentence in a sentence is corrected to a lower case by using a rule that an English sentence generally starts with an upper case letter, ends with a lower case letter and ends with a period.

〔The invention's effect〕

As described above in detail, according to the present invention, in the post-processing of character recognition, the first character recognition result of the target character string is obtained.
By examining the candidates, key characters belonging to a specific category (character category of absolute uppercase height or absolute lowercase height) are extracted, and the relationship between the vertical position or height of the image of the key character and other characters and other In order to correct the character recognition result of another character in the character string according to the category to which the first candidate of the character belongs, the following effects can be obtained.

(1) The processing target is not limited to a document in which the character height in a line is constant, and it is also possible to correct the character recognition result of an English document whose character size (height) changes for each word.

(2) Accurate characters are absolutely excluded as in the method of character classification in the preprocessing, and correction is not impossible or extremely difficult. Further, information directly related to the image is information for each individual character. Therefore, troublesome skew correction is not required.

(3) On the basis of the characters (key characters) appearing in the target character string, the other characters in the target character string are compared. Accuracy does not depend on the quality of the shape parameters prepared in advance based on the above, and there is no need for maintenance such as preparation, updating, and correction of parameters.

[Brief description of the drawings]

FIG. 1 is a hardware configuration diagram showing an example of an English character recognition device according to the present invention, FIGS. 2 and 3 are schematic flowcharts of correction processing in an embodiment of the present invention, respectively, and FIG. FIG. 5 is a diagram showing an example of data used in the post-processing, and FIG. 5 is a diagram showing an absolute uppercase character, an absolute lowercase character, and descender characters. 1 ... Scanner, 2 ... Image memory, 3 ... Central processing unit, 4 ... ROM, 5 ... RAM, 6 ... Dictionary.

Claims (3)

(57) [Claims] In a post-processing of character recognition, a first candidate of a character recognition result of a target character string is examined to extract a key character belonging to a character category of an absolute uppercase height or an absolute lowercase height. Correcting the recognition result of another character in the character string according to the relationship between the vertical position or height of the image of the character and another character in the character string and the category to which the first candidate of the other character belongs. How to correct the character recognition result. 2. The method according to claim 1, wherein key characters are extracted and corrected in units of words sandwiched between spaces. 3. The method for correcting a character recognition result according to claim 1, wherein the likelihood of the character recognition result is evaluated, and a character having a low probability is removed from the key character.