JPH03198181A - Post-processing method for character recognition - Google Patents

Abstract

PURPOSE:To correct an error of a large character and a small character, etc., by a simple processing by collating a character list rule, and a result of character recognition by matching with a character dictionary. CONSTITUTION:A result of character recognition is corrected by collating a character list rule 18 in which a specific character, and a character which can be pre-positioned therein or a character which cannot be pre-positioned are described, and the result of character recognition by matching with a charac ter dictionary 15. For instance, 'ya', 'yu', 'yo', etc., of small characters of a KANA (Japanese syllabary) never appear in the head of a sentence, and always appear together with a large character in front thereof. As for its large character, any character is not always allowed, and a rule as a character-string by which only a limited character is allowed exists. Accordingly, with regard to a specific KANA small character, etc., to which a pre-positioned character is limited, the rule in which the character which can be pre-positioned or cannot be pre-positioned is contained is prepared, and it is collated with a result of character recognition. In such a way, an erroneous recognition of a large charac ter and a small character, etc., which become a rule violation can be corrected easily and surely.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a Japanese character recognition device, and particularly to post-processing of character recognition results by matching input characters with a dictionary.

[Conventional technology]

Generally, when performing character recognition, character images are normalized to a certain size and then matched with a dictionary. Uppercase letters such as ``ya'' and lowercase letters ``ya'' cannot be distinguished.

To solve this problem, methods have been devised to identify uppercase letters, lowercase letters, punctuation marks, etc. using the position and size of characters.

For example, a method of determining a character based on the relative positional relationship with the adjacent character on the left, as described in Japanese Patent Publication No. 1676/1980, and a method of writing a form as described in Japanese Patent Publication No. 60-9314. A method of determining a character based on the position of the character within a frame, a method of determining a character based on the arrangement, position, and size of N surrounding characters as described in Japanese Patent Laid-Open No. 63-78287 are known. There is.

[Problem to be solved by the invention]

However, since the font size and printing position differ depending on the Fendt, the problem cannot be said to have been solved when a variety of documents are to be recognized.

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide a post-processing method for correcting errors in character recognition results due to matching with a character dictionary, particularly errors in uppercase and lowercase letters that are likely to be misrecognized, by simple processing.

[Means to solve the problem]

The present invention obtains a single character recognition result by comparing a character sequence rule that describes a specific character and characters that can or cannot be prefixed to it with a character recognition result by matching with a character dictionary. It is characterized by correcting.

[For production]

For example, the lowercase kana letters ``ya'', ``yu'', and ``yo'' never appear at the beginning of a sentence, and they always appear with some kind of uppercase letter before them. Not all capital letters are allowed; only a limited number of characters are allowed. In other words, there are rules as strings.

Therefore, for specific kana lowercase letters etc. that are limited to the characters that can be prefixed,
By preparing rules that describe characters and comparing them with character recognition results, misrecognition of uppercase and lowercase letters that violate the rules can be easily and reliably corrected.

〔Example〕

FIG. 1 is a schematic block diagram of Kanji OCR according to an embodiment of the present invention, and FIG. 2 is a schematic flowchart of processing.

In FIG. 1, JMM image data to be processed is input from the image input unit 10 and stored in the one-image memory 11 (processing step ■).
Line cutting and character cutting are performed by the character cutting unit 12, and the cut out character image data is stored in the character image memory 1.
3 (processing step ■), the character recognition unit 14 performs preprocessing such as size normalization and feature extraction, and stores the extracted feature vectors and the character dictionary memory 15. By matching with the stored character dictionary, character candidates from the smallest distance to the highest N are selected (processing step ■).

The post-processing unit 16 is a part that performs uppercase/lowercase letter correction processing on the character recognition result, and the uppercase/lowercase letter correction processing is performed from processing step ① to processing step 0 shown in FIG.

The character arrangement rules used for this uppercase/lowercase letter correction processing are stored in the single character arrangement rule memory 18.

This character arrangement rule describes the characters that are allowed to come before characters that can only come before them (lowercase letters in Hiragana and Katakana), and the -
Examples are shown in Table 1.

Next, uppercase/lowercase letter correction processing will be described using an example of recognizing the five-character string ``Yogugu ni Aruru''. In this case, the recognition results shown in Table 2 are obtained by the 5-character recognition unit 14.As a result of not considering the size of the 6 characters, "O"
and "o", "yo" and "yo", "u" and "u" are incorrect Egl
! has been done.

First, the first candidate of the recognition result is compared with the character arrangement rule in order, and the process is continued until a target character that requires correction processing appears.
When the target character is found, the first candidate for the character immediately before the target character is compared with the character arrangement rules to see if it is a character that can be prefixed. As a result, if the previous character does not exist (the target character is the beginning of a sentence) or the previous character is not a character that can be prefixed, the second candidate for the target character position is replaced with the first candidate. If the second candidate is also a lowercase letter, compare the previous character with the rule in the same way for that character, and if it can be prefixed, replace it with the first candidate, otherwise replace it with the next character. The same process is performed for the candidates.

In this example, the first candidate "o" with character number 1 is the target character. However, since this is the beginning of the sentence, "o" is determined to be a longer character string, and is replaced with the second candidate "o". This "o" is not a target character, so character number 1
processing ends.

Since the first candidate character "yo" for the next character number 2 is the target character, the previous first candidate character "o" is compared with the content of the character arrangement rule. However, "o" is not in the rules, so
"Yo" is replaced with the second candidate "Yo J." Since "Yo" is not the target character, the processing for character number 2 ends.

When similar processing is repeated up to character number 8, misidentified characters are corrected in the recognition results as shown in Table 3.

Table 3 (Results of correction) Note: Underlining has been corrected. Note that the character arrangement rule can be such that a character that cannot be prefixed to the target character is written, but in the case of Japanese kana, The scale of the rule can be made smaller by writing characters that can be prefixed as in the example.

Recognition errors other than errors in uppercase/lowercase letters can also be corrected by creating similar rules.

In addition to the uppercase/lowercase letter correction process, the post-processing unit 17 may also perform a correction process based on the validity of the word or sentence.

The recognition results corrected by the post-processing unit 17 as described above are! ! ! The results are obtained in the identification result memory 16, which is outputted by the output section 19.

〔Effect of the invention〕

As described above, according to the present invention, errors in uppercase and lowercase letters in character recognition results obtained by matching with a character dictionary can be reliably corrected through simple processing.

[Brief explanation of drawings]

FIG. 1 is a schematic block diagram of Kanji OCR according to an embodiment of the present invention, and FIG. 2 is a schematic flowchart of processing. 10... Image human resources department, 11... Image memory, 12
... line/character cutting section, 13... character image memory, 14... character recognition section, 15... character dictionary memory, 16... recognition result memory, 17... post-processing section,
18...Character arrangement rule memory, 19...Output section.

Claims (1)

[Claims] (1) Character recognition results are obtained by comparing character sequence rules that describe specific characters and characters that can or cannot be prefixed to them with the character recognition results obtained by matching with a character dictionary. A post-processing method for character recognition in a character recognition device, characterized in that the character recognition is corrected.