JPH02173886A - Word recognizing system - Google Patents

Abstract

PURPOSE:To accurately perform recognition even when the meaning of an abbreviated name is overlapped by constituting a dictionary by registering a homonym as a reference type on a word, and executing the evaluation of the word by adding the homonym on the reference type. CONSTITUTION:Image information obtained at an input part 11 is supplied to a character detection recognition part 13, and information comprising a character is extracted from the image information, and a recognition processing is executed on individual character. Word recognition is executed on the above result at a word recognition part 15 by using a first data base. After the execution of the word recognition, address recognition is performed at an address recognition part 21 by using a second data base 19, and a recognition result is supplied to an output device 23, then, it is displayed. And the dictionary is comprised by registering the homonym as the reference type on the word, and the evaluation of the word is executed by adding the homonym on the reference type. In such a way, it is possible to accurately execute the word recognition.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a word recognition method, and for example, to a word recognition method that is the basis for searching for an address.

(Prior Art) There are various recognition methods, but in the case of address recognition, the following methods are adopted.

First, to perform address recognition, characters are recognized, followed by words. Based on this recognition result, addresses are searched by combining words. During word recognition, if synonyms were found, abbreviations were suggested as candidates. For example, rHiGHWAYJ rI(tWAYJ rH
When the description WYj was read and recognized, its abbreviation rHWYJ was obtained from the dictionary as a recognition candidate. This is rHiGHWAYJ rHiWAYJ rHW
It was very effective when both YJ had the same meaning, that is, expressway.

However, after repeated tests and studies by the inventor of the present application, it was found that the meanings of the abbreviations overlapped in some cases, making it impossible to perform correct address searches.

(Problems to be Solved by the Invention) As described above, in the conventional technology, word candidates are expressed by a single abbreviation, so when the meanings of the abbreviations overlap, correct address retrieval cannot be performed.

Therefore, it is an object of the present invention to provide a word recognition method that allows more accurate recognition even when abbreviations have overlapping meanings.

[Structure of the Invention] (Means for Solving the Problems) In order to solve the above problems, the present invention provides a word recognition method in which words are recognized using a dictionary. The method is characterized in that a synonym is registered and configured, a synonym is added to the standard form, and the word is evaluated.

(Operation) According to the present invention, all possible meanings are assigned to an abbreviation and a plurality of candidates are used as recognition candidates, so that word recognition is performed more accurately.

(Example) Next, an example of the present invention will be described in detail using the drawings.

This embodiment relates to a mail address recognition system in the United States.

This address recognition device has an input section 11 as shown in FIG.
including. This input unit 11 is composed of a photoelectric conversion device such as a CcD, and converts an image on a mail item into rOJ rlJ image information. The image information obtained by this input section 11 is supplied to a character detection recognition section 13. The character detection and recognition unit 13 extracts information constituting characters from image information and performs recognition processing on each character.

The word recognition unit 15 executes word recognition on this result using the first database.

After word recognition, the address recognition unit 21 uses the second database 19 to recognize the address. This recognition result is supplied to the output device 23 and displayed.

Next, details of the operation processing in the above configuration will be explained.

Character detection by the character detection recognition unit 13 is achieved using a well-known technique. First, statistics on the distribution of image information in two axes (horizontal and vertical) are taken. That is, projections are taken in each of the two axis directions and the number of "1"s is counted. Then, since the number of "1"s, that is, the number of pixels is larger in the area where the character is present than in other areas, the detection of the character can be easily realized. However, at this time, the distribution is based on pixels with a coarser density than those used in normal recognition processing. For example, the reading ability in COD is 8 lines/
mm, and if normal recognition processing is performed at a rate of 8 lines/mm, the above detection process is performed with an accuracy of 1 line/mm. Processing using coarse pixels is preferable because it prevents errors in character detection.

In this character detection, projections are taken in the two-axis directions, and there is a break (area with no pixels) in the projection in the - direction, and the interval between the breaks is greater than a certain value. This direction is the first
is called the direction of Projections in other directions are either seamless or closely spaced.

Here, it is considered that "rows" are constructed along the direction of the cut. Therefore, along the above cut,
Cut out the “row”.

Once a row is cut out, a projection is taken along a second direction perpendicular to the first direction with respect to the image of the area constituting this row. As a result, pixels are counted in areas with characters, and pixels are not counted in areas without characters. In this process, the pixel density to be handled may be made more precise than that at the time of line extraction. Similarly to the above, the presence or absence of a character appears as a projection in the second direction. Therefore, characters can be cut out from a line, and character inspection can be achieved.

Following character verification, character recognition is performed. In this embodiment, character recognition is performed using, for example, a composite similarity method.

The processing up to this point is the processing in the character detection and recognition unit 13. From the character detection and recognition unit 13, rHJ riJ rWJ
Information about characters such as rYJ can be obtained.

After this "character recognition," the word recognition unit 15 executes word recognition, which is a combination of characters.In addition, regarding recognition, rather than simply obtaining a recognition result, multiple recognition candidates are Words are obtained in descending order of likelihood, and in normal cases, a plurality of words are output as recognition candidates by reading them on a piece of mail.

For the sake of explanation, please note that rxxx is included in the address on the target mail.
x HiWAYJ, and as a result of character recognition, rHiW
Suppose that AYJ is extracted as a word. This word rHi
WAYJ and the contents of the first database 17 are compared and verified.

The relevant portion of the first database 17 is configured as shown in Table 1 below.

In Table 1, ``Letter type'' in the left column of Table 1 includes abbreviations.

Therefore, since the word read from the mail is "HiWAY," this word is compared with the "letter type" in the first database 17 shown in Table 1. In this example, since the second line of Table 1 matches, the standard form is first extracted as a recognition candidate. Furthermore, rHiWAYJ written on the mail is treated as a synonym and one of the recognition candidates. This is because rHiWAYJ [Expressway (
This is because in addition to the case where it is Hi G HWA Y) J, it may also be a place name.

In this way, two recognition candidates rHiWAYJrHiGHW
AYJ is output and sent to the address recognition section 21. The address recognition unit 21 uses one second database to narrow down the recognition candidates to one (.Specifically, it includes the words before and after the read word and decides on one recognition candidate. So, assuming it is a place name, rxxxxx HiGHWAY
J is selected.

This result is displayed, etc. by the output unit 23.

Also, assume that the mail has the words "RORORORO HWY XXXX" written on it and is read as such.

In this case, in recognizing the word rHWYJ,
[The word WYj and its standard form “HiGHWAYJ” are extracted as recognition candidates.

Then, these two candidates are finally narrowed down to one by the address recognition unit 21.

[Effects of the Invention] As described above, according to the present invention, since the standard form of a word and a synonym are used as recognition candidates, word recognition can be performed more accurately.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram showing the configuration of an address recognition device according to an embodiment of the present invention. 11...Human power department 13...Character detection recognition unit 15...Word recognition unit 17...First database 19...Second database 21...Address recognition unit Representative Patent Attorney Norichika Keisuke 23... Output part January 2nd - Kyo 1
times

Claims (1)

[Claims] (1) In a word recognition method that recognizes words using a dictionary, the dictionary is configured by registering standard forms and synonyms for the words, and the synonyms are added to the standard forms to evaluate the words. A word recognition method characterized by: