JPH04111186A - Character recognition result correction method for address character string - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention is directed to an optical character reader (OCR: opii
The present invention relates to a character recognition method using a computer (cal character reader), etc., and particularly to a method for correcting character recognition results for address character strings to correct recognition errors in character recognition results for addresses written on sales slips, delivery slips, etc.

[Conventional technology] Conventionally, character recognition systems using OCR convert an optical image on paper into an electrical signal, perform shaping processing, extract features, and identify individual characters using structural analysis methods or pattern matching. Recognizes character patterns.

In addition, when character recognition is performed on sales slips, delivery slips, and other forms that contain address character strings, the character recognition results are checked against a place name word dictionary, and in the process of selecting likely words, candidate characters from the character recognition results are combined. A character string is created using a string, and a word matching this string is searched from a place name word dictionary.

Note that if there is no completely matching word, a method is used in which points are assigned to partially matching candidate characters according to their candidate rankings, and the one with the highest score is determined as the correct word.

For this type of method, see, for example, "Knowledge Processing for Free-Text Address Strings," Seino et al., IEICE Spring National Conference D-465, pp. 6-185.1989, OC
About reading address data in R, Information Processing Society of Japan 34th Gold Country Conference 4E-5, pp. 1841.1987
”, etc.

[Problem to be solved by the invention] In the above-mentioned conventional technology, the recognition result of the character recognition device is
It was possible to correct recognition errors by checking with a place name dictionary.

However, since recognition results with multiple candidate characters are combined for one character pattern, the number of combinations increases, and the number of matching characters for each character string is checked for ζ second generated character strings. The number of times that character-by-character matching is performed becomes extremely large, increasing processing time. Conversely, if the number of candidate characters is reduced to reduce the number of combinations in order to shorten processing time, correct candidate characters may be excluded from the processing target.

In addition, since the candidate ranking is used to determine the degree of similarity, all candidate characters in the same ranking are uniformly considered to have the same probability. That is, even if the distance lIi value, which is a value representing the certainty of the character recognition result, differs, this difference cannot be reflected in post-processing.

The purpose of the present invention is to improve such problems, shorten processing time, and effectively utilize information in the character recognition process that was not utilized in conventional methods.
An object of the present invention is to provide a method for correcting character recognition results for address character strings, which can improve processing accuracy.

[Means for Solving the Problems] In order to achieve the above object, the method for correcting character recognition results for address character strings of the present invention is provided in a character recognition system that determines recognition results of written address character strings and corrects errors. , has a place name word table that expresses the hierarchical relationship of place name words that make up an address string, and an index table that is sorted by character code at the character position of each word for each hierarchy, and is output from a character recognition device. Using the index table for the recognition result of the address string, search for matching characters in each column and find the record number in the place name word table of the place name word where that character exists. By speeding up the process of extracting similar place name words from the table, from the extracted place name word group, using the place name word table,
We select place name words that can be correctly connected to the place name words in the upper hierarchy determined earlier, and then select the word with the highest degree of similarity to the selected place name word based on the evaluation value calculated based on the distance value for each character. It is distinctive in determining place name words.

[Operation 3] In the present invention, for each candidate character in the character recognition result, the index table is searched in order from the first column, and the record number in the place name word table of the place name word in which that character exists is determined. At this time, the record numbers that appear many times among the obtained record numbers are record numbers representing similar place name words. In this process, record numbers are extracted by comparing only one character, and when finding the number of matching characters and selecting similar place name words, only record numbers need to be compared.

In addition, the place name word table stores the record numbers of place name words, which are the upper level of each place name word.
Using this, it is possible to determine whether it is possible to correctly connect to the previously determined higher level place name.

Furthermore, in the evaluation value calculation process after the connection determination with the upper level is completed, a distance value ( Reciprocal of degree of conformance of character pattern)
Use.

This makes it possible to shorten the processing time and increase the precision with which character recognition results are corrected.

f Example] An embodiment of the present invention will be described below with reference to the drawings.

FIG. 2 is an explanatory diagram of a place name word table in one embodiment of the present invention, and FIG. 3 is an explanatory diagram of an index table for the first and second columns at the prefecture level in one embodiment of the present invention.

The character recognition system of this embodiment includes a character reading device such as ○CR, a CPU that controls the entire system, a memory for storing various programs and necessary data, a display device such as a CRT, an input device such as a keyboard, and an index. It consists of an external storage device for storing tables, place name word tables, etc.

When a sales slip, delivery slip, or other form containing an address string is inserted, this character reading device recognizes each character, converts it into an electrical signal, and sends it to CPIJ.

In addition, the CPU searches the index table in memory based on the recognition results, extracts place name words from the place name word table in memory using the record number obtained thereby, and corrects the address string recognition results. do.

Further, the place name word table expresses the hierarchical relationship of place name words that make up the address character string.

Specifically, it is shown as shown in FIG. In the prefecture table shown in FIG. 2, the number on the left side of the table is the record number of the record itself, and the record number stored on the right side of the place name is the record number indicating the upper record. For example, ``Kyotofu'' and ``Maizurufu'' are included in ``Kyotofu'' (located in the lower level of ``Kyotofu''), and the record number indicates the record in which these higher-level place name revolution words are stored. is also "10".

In addition, the index table is sorted by the character code at the character position of each word for each layer of place name words that have a hierarchical relationship.

Specifically, it is shown in Figure 3, and for each character position,
It consists of candidate characters and record numbers indicating the relationship between candidate characters and place name words.

Next, a method for correcting address character string recognition results in this embodiment will be described.

FIG. 1 is an operation flowchart of a character recognition system according to an embodiment of the present invention, and FIG. 4 is an explanatory diagram showing candidate characters and distance values of character recognition results in an embodiment of the present invention.
FIG. 5 is an explanatory diagram showing record numbers extracted in one embodiment of the present invention.

In this example, the form is "Katsukagoyacho, Shimogyo Ward, Kyoto City, Kyoto Prefecture".
Let's talk about the case where this is entered.

As shown in Figure 1, each character of the address "Katsukagoya-cho J, Shimogyo-ku, Kyoto City, Kyoto Prefecture" written on the form entered into the character reading device is recognized, and the recognition result shown in Figure 4 is obtained. 101
), make a fix for this.

First, based on the character recognition result of the address character string, matching characters are searched for each column by referring to the index table shown in FIG. 3 (102).

In other words, the index table sorted by the first character of place name words at the prefecture level is run through a binary search algorithm, using "East", which is the first candidate in the recognition results for the first column in Figure 4, as a key. Search using ``Pato'' as the first character to obtain the record number of a place name word whose first character is ``Pato''.Next, search for the second place candidate ゛ky''', a place name word whose first character is ``kyo'' Obtain the record number. From here on, this process is performed until the end of the candidate ranking, and the process for the first column is completed. The above process is also performed for the second column 100 and beyond, resulting in the maximum character string length at the prefecture level. Continue up to the fourth column. In Fig. 3, (a) shows the first column, and (b) shows the second column.

In this way, a record number group of place name words is obtained as shown in FIG.

Next, the obtained record numbers are sorted (103),
A record number with a high frequency of appearance is selected (104), and the connection with higher level words is checked by referring to the place name word table (l○5).

That is, merge-sort the record numbers in Figure 5,
Record numbers that appear more than half of the word length are judged to have similarity and are extracted.

At this time, the extracted record number is r
l OJ, r40J and 2 characters match 1. It is r31J.

Note that if there is a previously determined upper-level place name word in the address string being processed, the number of candidates is narrowed down using the hierarchical structure of the upper-level and lower-level place names in the address string. In other words, as shown in Figure 2, the connection with the upper level is determined by checking whether the record number of the upper level stored in the place name word table matches the record number of the previously determined place name word. I do. However, if there is no higher-level place name word, the connection determination process is omitted, so it is not performed in the case of the prefecture level (Kyoto Prefecture).

Next, among the place name words obtained in this way, an evaluation value is calculated to determine which one is the most likely, and when comparing the recognition results with the place name word table, it is determined that the The one with the lowest evaluation value in
It is determined that it is the most probable (106).

In this evaluation value calculation, a value (distance value) representing the probability of each candidate character is used to calculate the evaluation value as shown below.

(In-distance NIM = distance value of the nth candidate - distance value of the first candidate (b)) Evaluation value = sum of distance differences between matched characters. However, the distance value is the reciprocal of the degree of matching of the character pattern, The larger the value, the lower the certainty. That is, the candidate characters for character recognition are arranged in ascending order with respect to the distance value. Also, the distance difference of the first candidate is O.

In this embodiment, in step 104, the words that match three characters are "101 Kyoto Prefecture", "40: Nara-kei", and 2.
“31: Osaka Prefecture” is extracted as a word with character matching.

The calculation of each evaluation value is as follows.

Kyoto Prefecture (2994-1454) + (1632-1632) + (
3861-1495)・3906 Nara
Prefecture (6534-1454) + (7962-1632) + (
6305-1495) = 1622° Osaka Prefecture (6509-1454) + (386
1-1495) = 7421 Among these, if we rank the correct answer candidates in the order of ``those with the largest number of matching characters and the smallest evaluation value'', we will choose ``Kyoto Prefecture'', ``Nara Prefecture'' from the most likely. ``Kei'', then ``Osaka Prefecture''. Therefore,
At the prefecture level in this embodiment, "Kyoto Prefecture" is the first correct place name word. Note that if the required evaluation value exceeds a system-defined threshold, it is excluded from the candidate words. For example, if the evaluation values of "Kyoto Prefecture" and "Nara-kei" exceed the threshold value, and only the evaluation value of "Osaka Prefecture" is within the threshold value, "Osaka Prefecture" becomes the correct word. This concludes the prefecture-level verification. Also,
In this embodiment, since place names at the prefecture level are determined using three characters, the next comparison at the city, ward, and county level is performed starting from the fourth column of the recognition results.

Furthermore, by repeating the above method, a correct address string is obtained, and the correction process for the character recognition result of the address string is completed (107).

[Effects of the Invention] According to the present invention, as a method for correcting errors in character recognition results of address character strings, processing time can be shortened by performing character string matching using an index table for each character position. In addition, the recognition result of the address character string output from the character recognition device is judged using the connection relationship with higher level words using a hierarchically structured place name word dictionary, and further using the distance value, Since the probability of candidate words is calculated, it is possible to effectively utilize information from the character recognition process, which is not used in conventional methods, and improve processing accuracy.

[Brief explanation of the drawing]

Fig. 1 is an operation flowchart of a character recognition system in an embodiment of the present invention, Fig. 2 is an explanatory diagram of a place name word table in an embodiment of the present invention, and Fig. 3 is a prefecture-level diagram in an embodiment of the present invention. FIG. 4 is an explanatory diagram showing candidate characters and distance values of character recognition results in one embodiment of the present invention, and FIG. 5 is an explanatory diagram of the index table for the first and second columns of the present invention. It is an explanatory diagram showing record numbers extracted in an example. Diagram Diagram Diagram Diagram Zhuang Castle Vein Dragon Bandit Vein

Claims (1)

[Claims] (1) In a character recognition result correction method for a character recognition system that determines the recognition result of a written address character string and corrects errors, a place name word table that expresses the hierarchical relationship of place name words that make up the address character string; Each hierarchy has an index table sorted by character code at the character position of each word, and by referring to the index table, searching for matching characters in each column, and searching for the place name of the place name word where the character exists. Find the record number in the word table, refer to the place name word table, extract similar place name words based on the record number, and from the extracted place name word group, select the place name words in the upper hierarchy determined earlier. It is characterized by selecting place name words that can be correctly connected to place name words, and determining the place name word with the highest degree of similarity to the selected place name words based on an evaluation value calculated based on the distance value for each character. How to correct character recognition results for address strings.