JP2977244B2 - Character recognition method and character recognition device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to identification of characters having similar shapes, and in particular, to kanji.
The present invention relates to identification of similar characters in a character recognition device such as OCR and correction of a recognition result candidate including similar characters.

[Prior Art] Generally, a character recognition device such as a kanji OCR compresses or decompresses (normalizes) an input character image to a certain size, extracts features, and extracts features of standard characters registered in a recognition dictionary. By collation, a character with a small feature distance is obtained as a recognition result candidate, for example, up to the tenth highest.

However, in the normalization process, the height and width are independently normalized to the same size regardless of the size and shape of the character, so that characters with similar shapes cannot be correctly recognized in many cases. Particularly, in a character recognition device that processes Japanese language, since there are extremely many such similar characters, a number of similar characters become recognition result candidates, and it is an important technical problem to determine which of the similar characters should be answered correctly.

In order to solve this problem, the applicant of the present application
[0091] The invention of "Similar character identification method using in-line white space ratio" has been proposed by the patent application No. 0091,
Has proposed an invention of a "similar character identification method using a character area ratio".

[Problems to be solved by the invention]

The above-mentioned “similar character identification method using in-line upper space ratio” uses the ratio of the space between the upper reference line of the line and the upper end of the character to the standard width of the line (in-line upper space ratio) to obtain the recognition result. This is a method of determining which of the candidate similar characters should be the correct answer (high-ranking candidate), and the shapes such as “tsu” and “tsu”, “.” And “o” and “O” are similar. Character identification is possible.

However, this method uses similar characters that are close in height and differ only in width, such as ":" and "two", Etc. cannot be identified. In addition, it is necessary to extract reference lines above and below the line, which are global information on the entire line, and the processing load for that is not light.

In addition, the above-mentioned “similar character identification method using character area ratio”
Is a method of evaluating a similar character as a recognition result candidate by using a ratio (character area ratio) between a standard character area of a line and an area of each character rectangle, and it is possible to identify similar similar characters. is there. However, characters that are vertically or horizontally long and have a similar character rectangle area, such as “1” and “one”, Etc. cannot be identified. In addition, the load of processing for extracting global information such as a standard character area of a line is not light.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character recognition method and a character recognition device that can identify a similar character that is difficult to identify by the above method with high accuracy and that does not require the extraction of global information with a large processing load. is there.

[Means for solving the problem]

According to the present invention, storage means (herein, referred to as an aspect ratio table memory) is provided in which a standard aspect ratio is stored in advance for specific characters that are difficult to identify by feature matching, and a plurality of specific characters can be obtained as recognition result candidates. In this case, the aspect ratio before normalization of the input character is compared with the standard aspect ratio of the specific character in the aspect ratio table memory, and the recognition result candidate having the smallest difference is regarded as the correct character candidate. It is.

(Operation)

For example, the normalized images of “2”, “:” and “=” are very similar images as shown in FIG. 3A, and are almost indistinguishable by feature matching. Similarly, “U”, “1” and “]” and “I”, “1” and “•” are almost indistinguishable after normalization as shown in FIGS. 3 (b) and (c). Absent. These have been difficult to identify sufficiently using conventional similar character identification methods.

However, when examining the aspect ratio (the ratio of the height and width of the character circumscribed rectangle) before normalization to the character in the above example, there is a considerable difference. Therefore, such similar characters can be easily identified by preparing their standard aspect ratio in advance, and among similar characters that have become recognition result candidates, characters having a high possibility of being correct. The recognition result candidate can be corrected so as to be a candidate of higher rank.

Since the aspect ratio, which is information on individual characters, is used to identify similar characters, there is no need to extract global information or the like on the entire line.

〔Example〕

FIG. 1 is a schematic block diagram of a character recognition device according to one embodiment of the present invention.

In this character recognition device, an image input unit 10 reads an image of a document (to be written horizontally) by a scanner or the like,
The binary image data is input and stored in the image memory 11. The line / character extraction unit 12 performs line extraction and character extraction for the input image on the image memory 11, stores the extracted character image data in the character image memory 13, and stores the character position, height, width, length and width. The ratio is stored in the cutout information memory 14 as cutout information. The character aspect ratio is used for identifying or evaluating similar characters according to the present invention.

The character recognition unit 15 reads a character image from the character image memory 13, normalizes the character image, performs feature extraction, and registers the extracted feature in a character dictionary (recognition dictionary) stored in the character dictionary memory 16. Compare the standard character features
Recognition result candidates with small feature distances are obtained up to the N-th order in ascending order of distance and stored in the recognition result memory 17.

The aspect ratio table memory 18 and the similar character determination unit 19 are all added according to the present invention. The aspect ratio table memory 17 stores a table in which a standard aspect ratio of a specific character which is difficult to be identified by the feature matching in the character recognition unit 15 is registered. Similar character determination unit 19
Uses the aspect ratio of the character to identify or evaluate similar characters and correct the recognition result candidates in the recognition result candidates in the recognition result memory 17.

The result output unit 20 outputs the corrected recognition result data obtained in the recognition result memory 17 to an output device such as a display or a printer.

Next, using the character string “New York City” as an example, the details of processing for identifying or evaluating a similar character and correcting a recognition result candidate will be specifically described.

It is assumed that recognition result candidates and cutout information as shown in Table 1 are obtained in the recognition result memory 17 and the cutout information memory 14 for this character string.

In the recognition result, the characters of character numbers 1, 2, 3, and 5 are erroneously recognized. These characters are similar characters whose shapes are almost the same by normalization as shown in FIG. 3, and it is difficult to identify the input characters by normalizing and then recognizing them.

The standard aspect ratio for such similar characters is stored in the aspect ratio table memory 18, but this value may be obtained in advance by a statistical method, or may be obtained dynamically from the target document currently being recognized. May be determined. The description method of the aspect ratio is arbitrary. For example, the value calculated by an expression such as aspect ratio = width of character circumscribed rectangle (character width) / height of character circumscribed rectangle (character height) is defined as an aspect ratio. What is necessary is just to have as a parameter to represent.

It is assumed that the standard aspect ratio shown in Table 2 is stored in the aspect ratio table memory 18, for example.

The similar character determination unit 19 compares the recognition result candidate for each input character with the character stored in the aspect ratio table memory 18 and determines the standard aspect ratio of the matching character and the aspect ratio of the input character in the aspect ratio table memory. 18 and extraction information memory
Read from 14, calculate the absolute value of the difference (comparison result),
Holds the calculated value.

In the case of the recognition result of the character number "2", the first candidate ":" is a character stored in the aspect ratio table memory 18, the standard aspect ratio is 0.2, and the aspect ratio of the input character is
Since it is 1.1, the difference is 0.9. "=" For the second candidate
Are also characters in the aspect ratio table memory 18, and the difference between the standard aspect ratio 6.0 and the aspect ratio 1.1 of the input character is 4.9.
The third candidate is also a character in the aspect ratio table memory 18,
The difference in aspect ratio is zero.

Next, sorting is performed using the aspect ratio difference as a key so that the candidate with the smaller aspect ratio difference is ranked higher, that is, as a correct character candidate. However, if there is no candidate or only one candidate in the aspect ratio table memory 17, this sorting is not performed.

By this sorting, the recognition result of the character number 1 is changed to “2” in the first place, “:” in the second place, and “=” in the third place,
The recognition character “:” is corrected to the correct character “2”. By the same processing, the recognition results shown in Table 1 are corrected as shown in Table 3.

From this table, it can be seen that all the characters of the character string "New York City" were correctly recognized.

FIG. 2 is a flowchart showing the processing contents of the similar character determination unit 19 described so far. Is a processing portion for calculating an aspect ratio difference for a similar character in a recognition result candidate for each input character. Is a processing portion for confirming that two or more similar characters are included in the recognition result candidates and performing sorting based on the aspect ratio difference.

The aspect ratio calculation and the sorting may be executed after first confirming that two or more similar characters exist in the candidate characters.

In addition, in the similar character determination unit 19, together with the above-described processing, similar character identification based on the “similar character identification method using the in-line space ratio” or the “similar character identification method using the character area ratio” is performed. It is also possible to execute processing for correcting the recognition result.

〔The invention's effect〕

As described above, according to the present invention, by simple processing that does not need to extract global information, “2”, “:” and “=”, and “Yu” and “1”, which have conventionally been difficult to identify. "When"]",
It is possible to identify similar characters of various shapes such as “one”, “1” and “•”, and appropriately correct the recognition result candidate for such various similar characters and make a mistake. It is possible to improve the recognition rate for input characters having many easy similar characters.

[Brief description of the drawings]

FIG. 1 is a block diagram of a character recognition device according to an embodiment of the present invention, FIG. 2 is a flowchart showing processing contents of a similar character determination unit, and FIG. 3 is an explanatory diagram of a change in character shape due to normalization. . 10 ... Image input unit, 11 ... Image memory, 12 ... Line / character cutout unit, 13 ... Character image memory, 14 ... Cutout information memory, 15 ... Character recognition unit, 16 ... Character dictionary memory, 17: recognition result memory, 18: aspect ratio table memory, 19: similar character determination unit, 20: result output unit.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G06K 9/68 G06K 9/46

Claims (2)

(57) [Claims] A character recognition method for extracting a feature by normalizing an input character and comparing the extracted feature with a feature of a standard character registered in a recognition dictionary to obtain a recognition result candidate. A standard aspect ratio is stored for a character, and when a plurality of specific characters are obtained as a recognition result candidate, the aspect ratio before normalization of the input character and the standard value of the specific character of the recognition result candidate are obtained. A character recognition method comprising comparing an aspect ratio with a recognition result candidate having the smallest difference as a correct character candidate. 2. A character recognition apparatus for extracting a feature by normalizing an input character and comparing the extracted feature with a feature of a standard character registered in a recognition dictionary to obtain a recognition result candidate. Storage means for storing the standard aspect ratio of the input character, means for detecting the aspect ratio of the input character before normalization, and, when a plurality of specific characters are obtained as recognition result candidates, before the input character is normalized. Comparing the aspect ratio of the character with the standard aspect ratio of the specific character stored in the storage unit, and determining a recognition result candidate having the smallest difference as a correct character candidate. Character recognition device.