JPH01171080A - Recognizing device for error automatically correcting character - Google Patents

Abstract

PURPOSE:To omit a collating time with a word spelling by automatically checking and correcting erroneous recognition in specified feature quantity according to the information of the position, size and external form of a character and the constituting rule of the character itself with obtaining the capital letter and small letter of English, a figure and a special symbol as a recognizing object character. CONSTITUTION:The title device equips a character size deciding means 22 to compute the height of the character by the three-step constitution (upper, medium and lower) feature of the character and to decide whether the character is large or small, a character external form deciding means 23 to decide whether the character has a wide width or a long length, a character collating means 40 to successively compares the prescribed feature quantity to be extracted from the input character and the feature quantity of each standard character pattern and to determine the character whose quantity of distance is also small and a character constituting rule correcting means 50, etc. Thus, the auxiliary information of the input character such as the position of the character string of the input character, the size and external form of the character are compared with the auxiliary information of a standard template 45 and the erroneous recognition in the main feature quantity of the character such as stroke extraction or a peripheral feature is detected. Then, the correction of a similar character is executed and spelling collation is eliminated.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a character recognition device that automatically detects and corrects erroneously recognized characters based on character construction rules.

BACKGROUND OF THE INVENTION FIGS. 2 and 3 are block diagrams showing conventional character recognition devices. In Figures 2 and 3, to scans a document with an optical scanning device such as an image scanner (in+age 5canner), and marks white pixels on the document as "0" and black pixels as "1". Binary data (b
This is an image scanning means that converts the image into initial data and then stores it in a buffer. 20 examines the value data stored in the buffer and determines the horizontal temporal distribution of the image (in+age histogra+w) and the threshold value (
This is a character string or character cutting means that separates a character string according to the vertical characteristic distribution of the image (thresllold value) and separates the characters from the separated character string according to the vertical characteristic distribution of the image. 30 is a predetermined feature extraction method, for example, stroke extraction
on ) and peripheral features
The character feature extraction means analyzes the characters cut out by the cutout means 20, calculates the feature amount of the input character, and extracts the features for identification. 45 is the character feature extraction means 30 stored in memory in advance.
This is a standard character template for matching each character to be identified extracted in . Reference numeral 40 compares the feature amount of the input character extracted by the character feature extraction means 30 with the standard character template 45 for comparison, calculates the difference between them, and selects candidates for identification results in descending order of accuracy according to the difference value. It is a character matching means that outputs . 51 is a means (FIG. 2) for manually inspecting and correcting the results identified by the character matching means 40;
2 is a means (FIG. 3) for checking and correcting the spelling of words in sentences of a manuscript. Reference numeral 60 denotes a recognition result display means for displaying the identification result on a terminal or printer.

Misrecognition by recognition systems that use characters as the unit of recognition processing, as described above, can be divided into three types as described below.

(1) Mixture of uppercase and lowercase letters: As shown in Figures 4 and 5, C/c, 010,
S/s%U/u, V/v, W/w, X/x
, Z/z, P/p, etc.

(2) Mixing of similar characters: 01010.1/l/I, 5/S, 9/g.

(3) Others: l → ], ゛ (ellipsis) →, (comma), etc.

Conventional character recognition systems have only two types of processing ζj for misrecognition. One method is to manually detect errors in the artificial inspection shown in FIG. 2, as in the correction means 50, and correct them using keyboard input. The other method is to make corrections 2 using a spelling check like the correcting means 52 in the spelling check shown in FIG.

In the system shown in FIG. 2, the character recognition results are manually matched with the original, and if there are any misrecognitions, they are corrected by inputting them from the keyboard. Such inspection and keyboard input correction schemes are time consuming and labor intensive and cannot be automated.

The spelling check and correction means 52 shown in FIG. 3 is a method for detecting and correcting errors by spelling check. For example, in March 1987, when did the FIE Transmission System begin? Narcissism? IEEE TRA
NSACTIONS ON PATTERN
ANALYSIS AND MACllNE IN
On the Recognition of
Pr1nted Cbaracters of A
The system in the paper adopted this method. In the spelling matching and correction method, the checking is performed word by word, so words for spelling matching must be stored in advance. Therefore, a word template that stores such words requires a large amount of memory. Furthermore, as the number of word templates for matching increases, the processing time for testing and checking also increases. In addition, derivative words (D), e.g.
``'OCR'S'') cannot be inspected or corrected because the target is limited to the original reference word template.

Problems to be Solved by the Invention As described above, according to the conventional example, (1) Character recognition systems cannot automatically check and correct misrecognitions, and the number of times that misrecognitions have to be checked manually is greatly increased. There will be more.

(2) If spelling verification is used, the memory storage area and verification time will be large.

(3) There is no function to inspect and correct derived words.

There was a problem.゛Means for Solving the Problems In order to solve the above problems, the present invention provides, for example, a three-level structure of characters (top, top,
(middle, bottom) character position determining means for determining whether the corresponding input character is in the upper half or the lower half of the character string based on the characteristics;
Character size determining means for calculating the height of a character based on, for example, the positional coordinates of the corresponding input character and determining whether the character is a large character or a small character; and a character outer shape determination means for determining whether this character is a wide character or a long character based on the aspect ratio of this character, and a character string position of each recognition target character,
A character composition flag dictionary that stores flags composed of character sizes and character outlines, and a character pattern with the smallest distance by comparing a predetermined feature amount extracted from the corresponding input character and the feature amount of each standard character pattern one by one. If necessary, compare the information obtained by the character matching means that determines the character as an initially recognized character, the character position determining means, the character size determining means, and the character outer shape determining means of the input character with the initially recognized character composition flag. The present invention is also an automatic error-correcting character recognition device that includes a character structure rule correction means that determines whether or not it is necessary to correct similar characters by making a judgment based on rules regarding the context of characters.

Effect The present invention has the above-mentioned configuration, and the position of the input character string;
Compare the auxiliary information of input characters such as character size and character outline with the auxiliary information of the same character in the standard template such as position, character size, character outline, etc., and extract main features of characters such as stroke extraction and peripheral features. Detects misrecognition of characters and corrects similar characters.

Example FIG. 1 is a block diagram showing one embodiment of the present invention.

In FIG. 1, 10 is an image scanning means, and 20 is a character string or character cutting means, which is the same as the structure shown in FIG. 2. Reference numeral 21 denotes a character position determining means for determining the position of a character separated by the character string or the character cutting means 20 from the positional coordinates in the character string. This character position determining means 21 operates as follows. First, the position reference line of the character string is determined by the upper and lower limits of the separated characters.

Next, as shown in Figure 4, the uppercase letters and 0"
Since the height ratio of lowercase letters such as "a" and "a" is 3:2, the reference line coordinate ythr = 2/3 (upper limit of character string, lower limit of character string) is determined. Here, the upper limit coordinates of the character are yto
As p, if ytop≧ythr, it means that the character is located in the upper half of the string.

Otherwise, the character is placed in the bottom half of the string. As a result, the character position flag (CRF) is set to 1 or 0. , this character position characteristic data is useful for mixing similar characters by the character construction rule error correction means 50.

For example, as shown in FIG. 4, the characters "9", "," and ', which have similar shapes but different positions, can be automatically checked.

Reference numeral 22 denotes a character size determining means for determining whether the character is a large character or a small character based on the coordinates of the character separated by the character cutting means 20. The judgment method is as follows: H(1) = ytoo-ybattolI, where yt
op and ybottos are the coordinates of the upper and lower limits of the character, respectively. Let H(i) be the height of the character, H^LI
E is the average height of all letters. If H(i)≦H^
In the case of LIE, as shown in FIG. 5, characters distributed in the middle of a general text manuscript are determined to be small characters.

On the other hand, if )l(i)>HAuE, the characters shown in FIGS. 7 and 8 are determined to be large characters. Big,
Set the character case flag (C3F) to 1 or 0 depending on the lowercase letter.

FIG. 6 is a diagram showing a three-level distribution of uppercase and lowercase letters.

23 calculates the width of the character from the coordinates of the character separated by the character cutting means 20, and determines the outer shape of the character using the information obtained by the character size determining means 22 as shown in the following equation.

W(j)=X right
)=W(i)/H(i) where, Wright
-X 1ef are the coordinates of the right and left limits of the character, respectively, and W
(i) is the width of the character, H(i) is the height of the character, F(
i) is the ratio of the width and height of a character, ie refers to the aspect ratio of the character. If F(i)≧threshold (thre
shold value), the character is a wide character. For example, characters such as "M" and "W" are applicable. Otherwise, make the character a long character.

For example, characters such as 'b', 'f', and '1' correspond to this. Similarly, depending on whether the character is wide or long, l
Alternatively, add 0 to determine the character outline flag (CCF).

Character position flags (CPF) for the three types of characters that are likely to be misrecognized as described above11. Table 1 shows the character size flag (C9F) and character outline flag (CCF).

(The following is a margin) The first table 55 is a composition flag dictionary that stores character composition flags consisting of three character position flags, character size flags, and character outline flags for each character to be identified.

In FIG. 1 showing an embodiment of the present invention, reference numeral 40 compares the composite feature amount of the input character stroke extraction and peripheral features obtained by the character feature extraction means 30 with the feature amount of the character to be identified, and calculates the distance between them. This is a character matching means that calculates and selects the recognized character chain complement in order from the character pattern with the smallest distance. The character composition rule error correction means 50 uses the flags obtained by the character position determination means 21, the character size determination means 22, and the character outer shape determination means 23 to correct the character of the initially recognized character stored in the character composition flag dictionary 55. Check against configuration flags. If there is no match as a result of the comparison, it means that similar characters with different flags have been erroneously recognized, as shown in Table 1. In this case, processing such as correcting uppercase and lowercase letters, correcting numbers and characters, and correcting symbols is performed. If the comparison results match, it is because a correct recognition result has been obtained, or because similar characters with the same flags have been recognized as shown in Table 1. If it is a similar character with the same flag, it is checked according to the rules of the context of the character structure. If it matches the rules, it is considered a correct recognition result, but if it doesn't, the identification characters need to be corrected again.

The character composition rule error correction means 50 corrects the character composition according to the rules for the preceding and following characters as described below.

(1) The first letter after the punctuation mark ``,'' must be capitalized.

(2) If the character immediately before the abbreviation "2" (hereinafter referred to as the previous character) is a lowercase letter, the character immediately after the abbreviation (hereinafter referred to as the later character) is always a lowercase letter. For example, the letter "S" as shown in FIG. 9 corresponds to this.

(3) If the first character is a lowercase letter and the last character is also a lowercase letter, the characters between the two characters must be lowercase.

(4) If the first character is a space, and (A) the second character is a capital letter, then the character before the second character must be a capital letter. FIG. 10 corresponds to this.

(B) If the trailing character is a lowercase letter, then a. If the trailing character is a large character, and the position of the trailing character is higher than that character, then that character is always a lowercase letter. For example, the character "C" shown in FIG. 11 corresponds to this.

b. If the second character is a small character and the position of the second character is lower than that character, that character is always uppercase. for example,
This corresponds to the letter "S" shown in FIG.

(5) If the preceding letter is an uppercase letter, and (A) the position of that letter is lower than the preceding letter, then that letter is always a lowercase letter. For example, as shown in FIG.
C” character corresponds.

(B) If the position of that character is equal to the previous character, the 10th
As shown in the figure, the letters are always uppercase or large letters.

(6) In words, letters must not be mixed with numbers, so letters can only appear before or after a string of numbers.

The checking method according to the character context rules described above can now be easily implemented in a programming language, such as the C language. 60 is a recognition result display means for displaying the recognition result that has been automatically corrected.

A hardware layout diagram of a character recognition device like this embodiment is shown in FIG. The core of this character recognition device is a CPU and a digital signal processor.
It consists of a recognition device including a signal processor. For the original to be recognized, binary data is obtained using an image scanning device, and then this data is inspected, each character is separated by a process of cutting out character strings or characters, and character boundary directions (Boundary
Direction) and background density (Backgro
and Density), each feature is matched with a pre-stored standard template for matching,
The feature amounts are compared, the difference values of the feature amounts are accumulated, and the order of identification candidate characters is determined based on the values, and the candidate character with the smallest difference is taken as the first recognition result. Then, inspection and correction processing is automatically performed using the character structure rules of the present invention. Then, the obtained result becomes the final recognition result.

This invention is not limited to the above-mentioned embodiments, and can be implemented with modifications as long as the gist thereof is not changed. For example, even systems with different feature extraction means or matching means can be automatically inspected and corrected using the present invention. Furthermore, the object of the character construction rules of the present invention is not limited to English characters, but can also be applied to various spelling systems, such as German and French.

Effects of the Invention According to the present invention, English uppercase and lowercase letters, numbers, and special symbols are recognized as characters to be recognized, and information about the position, size, outline, and composition rules of the characters themselves are used to automatically recognize them using predetermined features. Since misrecognition can be inspected and corrected, memory for storing a large number of standard word spellings for comparison is no longer required, and the time required for matching word spellings can be saved.
Its practical effects are great.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a character recognition device according to an embodiment of the present invention, FIGS. 2 and 3 are block diagrams showing the configuration of a conventional character recognition device, and FIG. Figures 5 and 7 are explanatory diagrams showing examples of uppercase and lowercase letters that are likely to be mixed, Figure 6 is an explanatory diagram showing the three-level distribution of uppercase and lowercase letters, and 8th UgJ is the average of the characters. FIGS. 9 to 12 are explanatory diagrams showing examples of character composition rules, and FIG. 13 is a hardware layout diagram in an embodiment of the present invention. IO...Image scanning means, 20...Character string or character cutting means, 21...Character position determining means, 22...
Character size determination means, 23...Character outer shape determination means, 30
...Character feature extraction means, 40...Character matching means, 4
5... Standard character template, 51... Character composition rule error correction means, 55... Character composition flag dictionary, 60
...Recognition result display means. Name of agent: Patent attorney Toshio Nakao Hakaichimei III
Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7

Claims (1)

[Claims] Character position judgment that cuts out characters for each character string from input image data, determines whether the position of the corresponding input character is in the upper half or lower half of the character string, and stores 1 or 0 in the position flag. Determine whether a character is large or small based on the method and the height of the corresponding input character, and set the size flag to 1.
Alternatively, a character size determining means for storing 0, and a character external shape determining means for determining whether a character is a wide character or a long character based on the aspect ratio of the input character and storing 1 or 0 in an external shape flag; A character composition flag dictionary that stores flags including the position flag, character size flag, character outline flag, etc. for each standard character within the recognition range, and predetermined feature amounts extracted from the corresponding input character and each standard. a character matching means for comparing character pattern feature quantities point by point with character pattern feature quantities and determining the feature quantity of the character pattern with the smallest distance as the initially recognized character; a set of position, size, and outline flags of the input character; and the character composition graph dictionary. Compare the configuration flags of the initially recognized characters and if the two sets of flags do not match, immediately modify the similar characters, and if they match, check whether similar characters need to be modified further according to the character structure rules before and after the characters. 1. An automatic error correction character recognition device comprising: a character configuration rule error correction means for determining whether or not the character structure rule is incorrect.