JPH0527157B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention relates to a character recognition post-processing method, particularly a character recognition device, which reads a text containing a mixture of non-kanji including kanji, hiragana, katakana, etc., and performs character recognition. This invention relates to a character recognition post-processing method for correcting recognition errors that occur after character recognition is performed.

[Prior art and problems]

For example, optical character recognition devices have already been used to recognize Japanese texts including kanji and non-kanji characters, but there are no methods for correcting errors that occur after recognition, that is, post-processing measures for character recognition. We have not yet reached a stage where we are fully satisfied.

Conventionally, the method considered as a post-processing measure for character recognition is to divide the text into word parts and perform separate writing etc. in advance when handling complex Japanese sentences containing the above characters. There are two methods: one in which the position of words used in a sentence is kept in mind when filling in the text, and another in which the specific position of the word in error is notified when an error is corrected when an error occurs as a result of character recognition. .

However, in any of the above methods, extra steps must be taken by human intervention when filling out or editing information. In the above example, it is necessary to separate notes when writing, or to be aware of the position of specific words in advance, and it is necessary to perform processing like a dummy step for the final purpose. Ta.

[Purpose of the invention]

The present invention solves this problem in view of the above-mentioned problems.In the present invention, in order to distinguish between kanji and non-kanji in a text, separate identification methods are used to improve the accuracy of kanji and non-kanji identification. At the same time, a word dictionary is prepared according to the character strings that form words of a specific length for the identified kanji string, and a character recognition post-processing method that can efficiently take measures against and correct errors after recognition is developed. is intended to provide.

[Structure of the invention]

In order to achieve this object, the character recognition post-processing method of the present invention uses a character recognition device capable of classifying non-kanji such as hiragana characters and kanji, and a kanji/non-kanji determination means for determining whether the character is a kanji or a non-kanji. and,
A non-kanji identification means for identifying non-kanji characters, a kanji identification means for identifying kanji characters, a kanji character string extraction means for extracting a kanji character string, and when the kanji character string is longer than a specified number of characters, it is separated into a specified number of characters. The present invention is characterized in that it includes word separation means and performs word recognition by comparing the Kanji character string recognized by the Kanji identification means with a word dictionary that matches the length of the Kanji characters.

[Embodiments of the invention]

Before explaining the present invention in detail based on one embodiment, its outline will be explained with reference to FIG.

When the sentence shown in Figure 2a is read using OCR, for example, and the kanji and non-kanji characters are recognized one by one, and the recognition result shown in Figure 2b is obtained, the operator first corrects the error in the non-kanji part by, for example, key input. Then, a modified result as shown in c. Then, post-processing is performed by comparing the kanji part with a word dictionary.In this case, for parts where there are many consecutive kanji characters, most words have combinations of two characters, so for example, they are separated into two-character parts. Then perform post-processing.
As a result, "fingers and feet", "human power", "re-fitting", "one eye"...
etc. can be corrected to correct values such as "designation", "input", "device", "same", etc.

An embodiment of the present invention will be described based on FIGS. 1 to 5.

FIG. 1 is a configuration diagram of an embodiment of the present invention, FIG. 2 is an explanatory diagram of the recognition error correction process in the present invention,
Figure 3 shows a method for identifying kanji and non-kanji characters in a sentence.
The figure shows an example of analyzing and identifying kanji using the number of loops, the number of connected components, and the average number of black runs shown in Figure 3. Figure 5 shows an example of the contour line segment series method in Figure 3 (Japanese Patent Application Laid-Open No.
58-225849).

In Figure 1, 1 is a document input section containing kanji and non-kanji (hiragana, katakana, etc.) to be read, 2 is a kanji/non-kanji determination circuit, 3 is a non-kanji identification circuit, and 4 is kanji identification. circuit,
5 is a kanji character string extraction circuit; 6 is a display device; K is a key thereof; 7 is a word separation circuit; 8 is a word dictionary section in which groups of words having different meanings are stored in kanji strings of a predetermined length; 9 indicates a word processing circuit.

In the character recognition processing method according to the present invention, for a sentence containing a mixture of kanji, it is first accurately determined whether each character string is a kanji or a non-kanji, and then it is determined whether it is a kanji. Since a word is a series of several kanji characters, post-processing for words is automatically performed only on kanji strings.

Next, the operation of the embodiment shown in FIG. 1 will be explained.

For example, the characters in the document input section 1 obtained by reading the document with OCR, that is, the kanji in the sentence,
Non-kanji characters are sequentially determined through a kanji/non-kanji determination circuit 2. As a result, the identification circuit 3 determines what kind of characters (hiragana,
Katakana, etc.) is identified. On the other hand, the output determined to be a kanji is sent to a kanji identification circuit 4, and the type of kanji is identified for each character.

Since the non-kanji output output from the non-kanji identification circuit 3 and the kanji output output from the kanji identification circuit 4 are displayed on the display device 6, recognized non-kanji characters, such as erroneous hiragana, are displayed on the display device 6. Modifications are made by the operator using key K on the display device 6. As a result, ``wa'', ``ho'', ``ka'', etc. in Figure 2b are changed to ``re'', ``ha'', ``ga''(``ka'') as shown in Figure 2c.
is corrected as (due to misjudgment of kanji/non-kanji).

On the other hand, the kanji output from the identification circuit 4 is sent to a kanji character string extraction circuit 5, which extracts kanji character strings sandwiched between non-kanji characters. The extracted kanji string is then sent to a word separation circuit, where if the kanji string is long, such as 4, 5, or 6 words, it is separated into predetermined units. For example, the character string "character recognition device" is divided into "character,""recognition," and "device." In this way, Kanji characters separated by a certain length, for example two words, are sent to the word post-processing circuit 9. The word post-processing circuit 9 is connected to the word dictionary section 8 and receives kanji output from there. Further, the word dictionary section 8 stores in advance a large number of word groups having different meanings and having a certain length. Therefore, in the word post-processing circuit 9, the words from the word separation circuit 7 and
Words from word dictionary part 8 (both have equal length)
are compared, and it is determined whether the former recognized word is correct. At this time, since a plurality of candidates are extracted and sent to the identification circuit 4, these candidates are also used for comparison and comparison.

That is, if there is an error in a non-kanji character, such as a hiragana character, the character with the error has already been corrected on the display device 6 after recognition, so if there is an error in the recognized kanji character, the kanji will be included. Based on the meanings before and after the characters, the error kanji from the word separation circuit 7 is compared with the kanji output from the word dictionary section 8 in the word post-processing circuit 9, and since it is determined that an error has occurred, the error kanji is sent from the dictionary section 8. will be automatically replaced with the correct word in the output.

For example, if a sentence like the one shown in Figure 2 a is written in document 1, after reading it, the kanji,
Assume that there is an error as shown in b in the sentences recognized by the non-kanji identification circuits 4 and 3. Since the recognition error sentence is displayed as it is on the display device 6, the operator looks at it and corrects the hiragana in which the error exists, such as c, by pressing a key. That is, in the illustrated example, ``wa'' is corrected to ``re'', ``ho'' to ``ha'', and ``KA'', which has been recognized as a kanji character, to ``GA''. In this hiragana correction process, the operator does not have to judge whether or not the kanji word is correct or correct it at all.

Errors in the recognition of kanji words are automatically corrected as shown in d of FIG. That is, if there is an error in a kanji word as shown in b in FIG.
Error words from ``finger and foot'', ``rearrangement'', ``first glance'', ``manpower''
As a result of sequentially extracting and comparing the input "convenient" and each candidate obtained during these conversions from the word dictionary section 8, the correct words "designation", "input", "device", and "same" were found. is automatically replaced as shown in d.

Next, we will outline a specific method for distinguishing between kanji and non-kanji. Regarding this, an application has already been filed in Japanese Patent Application No. 169510/1983 by the same applicant.

As shown in Figure 3, the discrimination between complex kanji with a large number of strokes and non-kanji with a small number of strokes (e.g. hiragana, katakana) is determined by analyzing the number of loops and the number of connected components described below in the first step. If the determination cannot be made, the second step is to check the average number of black runs described below and determine the number of black runs. ",""river," etc.), the contour line segment series is examined in the third step to make a final distinction between the two.

FIG. 4 shows an example in which the identification method up to the first and second stages of FIG. 3 is applied to the kanji character "kan".
In the figure, the number of loops is the part forming the ring on the right side of "Kan", and in this case, the number of loops is two. The number of connected components is the number of separate and independent parts of each stroke, and in the illustrated example, there are three in the crested part, and one in the part similar to the crested crest on the right.
, and one of the blocks containing its bottom loop
That's a total of 5 pieces.

The average number of black runs is divided into the number of black runs in columns and the number of black runs in rows. In the example shown, the number of black runs in a column is the number of black dots (with information) that exist when a kanji is scanned vertically. This means that the number is 3 for the sandy part and 6 for the right side part. Expressing this in a general formula, the average number of black runs in the column n _y is The average number of black runs in a row n _x is It turns out that.

As described above, by using the number of loops, the number of connected components, and the average number of black runs, characters can be separated into characters with a large number of strokes and characters with a small number of strokes. Next, inside this small number of strokes characters,
Contour segment features are used to classify non-kanji characters and kanji characters with a small number of strokes. FIG. 5 shows an example of contour line segment feature extraction. Contour line segments are determined by whether each line segment is open (○) or closed (●) at the edge of the character.
4 types of line segments (○―○ ○―● ●―○ ●―●)
I can do it. The appearance series of these four types of line segments are very stable when the structure of the original pattern is simple. Therefore, by checking the appearance series (order of appearance) of outline segments for a character with a small number of strokes, it is possible to know the category to which the character belongs.

In this way, in the present invention, kanji and non-kanji are finally identified with considerable accuracy by going through the three identification steps shown in FIG. is corrected by the operator on the display device 6, and recognition errors occurring in the kanji characters can be automatically corrected into correct words.

〔Effect of the invention〕

As described above, in the present invention, when creating a document containing a mixture of kanji and non-kanji characters, it is possible to separate the sentences, or to be aware of specific character positions in the sentences, or to write specific words. Since identification errors for Chinese characters can be automatically corrected without being scheduled as recognition points, efficient character recognition post-processing can be performed without requiring extra manpower. For example, an optical character recognition device is obtained.

[Brief explanation of drawings]

Figure 1 shows the configuration of an embodiment of the character recognition post-processing method of the present invention, Figure 2 shows the process of correcting recognition errors after character recognition processed by the present invention, and Figure 3 shows how to correct kanji and non-kanji characters in a sentence. Identification method, Figure 4 shows an example of kanji analysis and identification using the first and second stage identification methods in Figure 3, and Figure 5 shows an example of kanji identification using the third stage identification method in Figure 3. show. In the figure, 1 is a document input section, 2 is a kanji/non-kanji discrimination circuit, 3 is a non-kanji discrimination circuit, 4 is a kanji discrimination circuit, 5 is a kanji character string extraction circuit, 6 is a display device, K is a key, 7 is a word separation circuit, 8 is a word dictionary section, and 9 is a word post-processing circuit.

Claims (1)

[Scope of Claims] 1. In a character recognition device capable of classifying non-kanji such as hiragana characters and kanji, there is provided a kanji/non-kanji determination means for determining whether a kanji is a kanji or a non-kanji, and a non-kanji character for identifying a non-kanji. comprising an identification means, a kanji identification means for identifying kanji, a kanji character string extraction means for extracting a kanji character string, and a word separation means for separating a kanji character string into a predetermined number of characters when the kanji character string is longer than a predetermined number of characters; A character recognition post-processing method characterized by performing word recognition by comparing a kanji character string recognized by an identification means with a word dictionary that matches the kanji length.