JPH0262682A - Method for recognizing character - Google Patents

Abstract

PURPOSE:To high speedily and accurately output a recognized result even for a character having similar data by estimating the character after classifying candidate characters according to the position and the size of a character picture to be obtained when the character is extracted. CONSTITUTION:The cross directional marginal distribution of a character picture 2 rearranged into a horizontal state is counted, a standard character height (2) and the position of a character row reference line (1) can be grasped from the shape of the cross directional marginal distribution 3, and simultaneously with the extraction of the standard character height (2) and the position of the character row reference line (1), the position of a character row is extracted. When an object character is an alphanumeric character, since every alphanumeric character except lowercase letters (i) and (j) is composed of one linked component, the character can be extracted by obtaining the outline of the character picture. The candidate characters are classified by comparing extracted character height and character reference line with the standard character height and the character row reference line extracted in a block B. It is recognized what the extracted character is by comparing the dictionary data of the limited candidate characters with the data of the extracted character.

Description

Detailed Description of the Invention [Industrial Application Field] The present invention is a character recognition device that inputs characters written on paper as an image, searches for a character area from the character image, and converts it into a code number. Regarding recognition methods.

[Traditional technique! ] In recent years, with rapid advances in character recognition devices, it has become possible to automatically extract character areas from various combined images, cut out and recognize individual characters, and automatically create document files. Various methods have been devised for character recognition.

For example, as one of the character recognition methods, mesh features (Research and Practical Application Report Vol. 34, No. 1, pp. 47-57)
There is. This method expresses the rough shape distribution of the entire character. The feature extraction method involves dividing the circumscribed rectangle of a character to obtain small regions of 1 x 1. The area of the character part included in each of the small areas is counted and used as a mesh feature. The mesh feature has a LuxLuco bootara for each character, and the character can be estimated by comparing the ratio of the area of the character part in a certain area of LuxLuco with the data owned as a dictionary. becomes possible.

In addition, as another method, there is a 7-eral characteristic (Research and Practical Application Report Vol. 34 No. 1 p, p, 47-57). This method focuses on peripheral information of characters. The method for extracting features is to first find a circumscribed rectangle of a character pattern, and then divide each side of the circumscribed rectangle into squares. Next, it scans toward the characters from the divided side, and counts the area until it encounters the first character, and then the area until it encounters the next character. By performing the same processing on each divided side, it is possible to obtain a peripheral characteristic having WX4X2 data, and by comparing the data of WX4X2 with the data held as a dictionary. This makes it possible to estimate characters.

[Problems to be Solved by the Invention] However, it is very difficult to narrow down the candidate characters to one using only the mesh features and peripheral features of the character recognition method.

For example, taking alphanumeric characters as an example, the metsy feature has a lowercase g and the number 8, the lowercase q and the number 9, and the peripheral feature has a ■ and a lowercase 71, a lowercase r and a lowercase letter. The data of characters such as t1, uppercase G and lowercase e, are very close to each other. Therefore, when characters are recognized using the method described above, there is a problem in that characters having similar data are very easily confused.

SUMMARY OF THE INVENTION The present invention is intended to solve the above-mentioned problems, and its purpose is to provide a method for quickly and accurately producing recognition results even for characters having similar data.

[Means for Solving the Problems] The character recognition method of the present invention reads a character image written on paper etc. using an optical image input means, recognizes the characters in the input image data, and replaces them with a code number. 1) Characters are estimated after candidate characters are classified according to the position and size of character images obtained during character extraction.

(2) Candidate characters are classified by comparing the character reference line obtained in the character extraction stage, which is a stage before character recognition, and the position of the character image.

(3) Candidate characters are classified by comparing the standard character height obtained in the character line extraction stage, which is a stage before character recognition, and the size of the character image.

[Example] Hereinafter, the present invention will be explained in detail based on examples.

The character recognition method of the present invention is generally used in hardware as shown in the block diagram of FIG. The hardware includes an image input device 342 for inputting character images to be recognized, a RAM 52 for storing image information input by the image input device 34 and calculation results, and a CI for executing calculations.
It is composed of a PU 31, a ROM 33 that stores dictionary data and calculation programs for character recognition, and a moisture blur 35 that displays the recognized results.

Hereinafter, the character recognition method of the present invention will be explained in detail according to the flowchart shown in FIG.

In block A, a character image 1 to be recognized is input. At this time, the character image 1 inputted varies depending on the reading width of the image input device 34, but normally a character image 1 of multiple lines is input, but the character image 1 shown in FIG. 1 only has one line. Although not shown, the character recognition method of the present invention does not change at all even when multiple lines are input. If the image input by the image input device 34 is an oblique image as shown in FIG. 1 (-), recognition will be prevented from block B onward. Arrange them horizontally as shown in Figure 1<b>.

In block B, character image 2 rearranged horizontally
Count the lateral marginal distribution 4 of . Generally, the peripheral distribution in the lateral direction has a shape as shown in 5 in FIG. When the input image has multiple lines, a horizontal peripheral distribution similar to that shown in FIG. 1(C) is created for the same number of lines. Looking at the shape of the horizontal peripheral distribution 3, it is possible to know the standard character height ■ (the height of the lowercase letter X) and the position of the character line reference line ■. In block B, the position of the character line is extracted at the same time as the standard character height (■) and the character line reference line (■).

In block 0, word regions are extracted from the character lines extracted in block B. Word regions can be extracted by counting the peripheral distribution in the direction perpendicular to the character line, checking the intervals between clusters of the peripheral distribution, and determining whether it is a character interval or a word interval.

In block D, a character region is extracted from the word region extracted in block O. To extract the character area,
It is also possible to extract using the results of the marginal distribution counted in block C, but if the target characters are alphanumeric characters, the lowercase letters 1 and j are omitted and all are made up of one connected component. Therefore, characters can be extracted by taking the outline of the character image. In the present invention, characters are extracted using character outlines. Furthermore, the height of the character and the reference line of the character are determined at the same time as character outline extraction.

In block (2), the characters extracted in block D are recognized. In block E, first, the character height and character reference line extracted in block D are compared with the standard character height and character line reference line extracted in block B.

Classify candidate characters. In general, the relationship between the height of a character and the reference line of the character can be divided into three groups as shown in FIG. The first group is as shown in Figure 2-11. The group has the character height■
-1 is considerably higher than the standard character height■, and the character reference line■
-1 is located almost at the same position as the character line reference line ■. The characters belonging to this character group are the 0HARA2 and 0HkRk5 groups in FIG. The second group is
In this group, which is a group as shown in FIG. 2-12, the character height -2 is almost equal to the standard character height, and the character reference line -2 is also almost equal to the character line reference line -2. The characters belonging to this group of characters are 0HARA in Figure 4.
There are 4 groups. The third group is shown in Figure 2-1.
This is a group as shown in 3. In this group, the character height ■-3 is considerably higher than the standard characters, and the character reference line ■-5 is located considerably lower than the character line reference line ■. The characters belonging to the group of characters are 0 in Figure 4.
This is the HARA1 group.

As described above, candidate characters can be narrowed down to three groups based on the information on the character height and character reference line. Therefore, the number of candidate characters was initially 52 characters, but now it is 34 characters, at least 5 characters, and on average, the time used for recognition is halved.In block E, the number of candidate characters narrowed down in this way is The dictionary data and the extracted character data are compared to recognize the extracted character.

As described above, when recognition of extracted characters is completed in block E, block D and block E are repeated until recognition of words extracted in block C is completed, and when recognition of words is completed, character lines extracted by block B are Block C, Block D, Block E until the recognition of
Repeat the series of actions. Furthermore, block B, block 0 . A series of operations in block D and block E is repeated to coat the character image. In this way, the character image 2 input by the image input device 34 is converted into one document file. Thereafter, in block F, the recognized result is displayed on the display 65 and the process ends.

As described above, the character recognition method of the present invention counts the peripheral distribution in the horizontal direction after rearranging the input images horizontally. Therefore, by examining the peripheral distribution, the position of a character line can be determined. At the same time, it becomes possible to know the character line reference line and standard character height. By comparing the character line reference line and standard character height obtained in this way with the character reference line and character height obtained at the time of character extraction, candidate characters can be separated into three groups. By classifying candidate characters into three groups, what used to be 52 candidate characters becomes a maximum of 34 candidate characters and a minimum of 5 candidate characters. As a result, the time required for recognition can be reduced to less than half of the conventional time on average. In addition, by using such position information and size information, r and t1 lowercase q and number 9 (not shown in Figure 3), lowercase letter g and number 8 (not shown in Figure 3) It becomes possible to clearly classify characters that have very similar data such as "no characters", and it becomes possible to recognize them quickly and accurately.

[Effects of the Invention] As described above, according to the present invention, the horizontal peripheral distribution of an input image is counted, and when determining the position of a character line, the character line reference line and standard character height are determined. Thereafter, candidate characters can be narrowed down by comparing the character line reference line and standard character height with the character reference line and character height obtained during character extraction. As a result, characters with similar data can be classified, and since the number of candidate characters is small, recognition is possible at high speed and with high probability. As a result, OOR, which is currently widely used, can be made easier to use and more reliable.

[Brief explanation of the drawing]

FIGS. 1(α) to (C) are diagrams showing how to obtain a character line reference line and standard character height in the character recognition method of the present invention. FIGS. 2(9) to 2(C) are diagrams showing a classification method using a character line reference line and standard character height in the character recognition method of the present invention. FIGS. 3(α) to Cd) are diagrams showing classification according to character line reference lines and standard character heights in the character recognition method of the present invention. FIG. 4 is a block diagram of a character recognition device in which the character recognition method of the present invention is used. FIG. 5 shows a flowchart of the character recognition means of the present invention. 1... Input character image 2... Corrected character image 5... Horizontal peripheral distribution ■... Character line reference line■-112t3...Character reference line■...
... Standard character height ■-1, 2, 5 ... Character height 11.12, 13 ... Extracted character 21 ...
・Group 0HAEA122...Group 0H
ARA223...Group O-HA RA 5
24...Group 0HARA431...
... 0PU 2 ... RAM 3 ... ROM 4 ... Image input device 5 ... Display ... Block A ... ...Block B ...Block 0 ...Block D ...Block E ...Block F and above

Claims (3)

[Claims] (1) In a character recognition method that reads a character image written on paper etc. using an optical image input means, recognizes the characters in the input image data, and replaces them with a code number, A character recognition method characterized by estimating a character after classifying candidate characters according to size. (2) The first method is characterized in that candidate characters are classified by comparing the character line reference line obtained in the character line extraction stage, which is a stage before character recognition, and the position of the character image.
Character recognition method described in section. (3) The character recognition method according to item 1, characterized in that candidate characters are classified by comparing the size of the character image with the standard character height obtained in the character line extraction stage, which is a stage before character recognition. .