JPH0793467A - Address reading system - Google Patents

Abstract

PURPOSE:To attain high accuracy and high speed in the automatic reading processing of address notation. CONSTITUTION:All the address notation characters are character-recognized at first. With respect to the first combination of some characters in the recognized address notation characters, the pertient name of the metropolis and districts is retrieved by narrowing an object to the metropolis and districts (a first hierarchy). When the name of an actual metropolis and districts, Tokyo metropolis, e.g. is found as the result of the retrieving, the object is narrowed to the inside of Tokyo metropolis (a second hierarchy) and with respect to the next combination of some characters in the recognized address notation characters, a pertient community name (ward, city, etc.) is retrieved. When an actual community name, Adachi ward, e.g. is found as the result of the retrieving, the object is narrowed to the inside of Adachi ward (a third hierarchy) and with respect to the furthermore next combination of some characters in the recognized address notation characters, a place (the name of town, etc.) is retrieved. An address is retrieved in an order from a high-order hierarchy and the retrieving object is narrowed whenever the retrieving result is obtained like this so as to improve the rate of correct reading and to reduce a processing time.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an address reading system which electronically reads and recognizes an address notation printed or handwritten on a mail piece.

[0002]

2. Description of the Related Art Generally, in an address reading process, characters are first recognized, then they are arranged to recognize words, and at the same time (or subsequently) words are arranged to recognize an address. That is, the recognition process in the address reading system can be divided into two processes, "character recognition" and "address recognition".

Here, the character recognition processing unit collates the read character image (character pattern) with a dictionary of all possible character types, but in word recognition (address notation word recognition), the processing amount is reduced, so that the similarity is achieved. Several high-character types (for example, 10
Only the seeds) are selected and sent to the address recognition processing section. However, in this case, the correct answer is
If you do not enter 0), you will not be able to read the address.

In the case of not relying on the present invention, one of the following methods can be considered as a method for constructing an address reading system. (1) In character recognition, each character is collated / evaluated with all the character types used in the address recognition process, and a certain number of character type candidates are selected from the one with the highest evaluation value and passed to the address recognition process. (2) Address notation is processed in order from the upper hierarchy (for example, prefecture → city → ward → town), and character recognition and address recognition are repeated for each hierarchy.

[0005]

In the case of the above (1), when the prefecture is recognized by the address recognition, it is used only for a city, a ward, or a town as a high-ranked candidate of the character recognition result of the prefecture name. If many characters are included, the correct character will be 10
Sometimes you fall out of a candidate for a species. In this case, of course, the correct word or address cannot be recognized, which makes it unreadable.

For example, in the example of FIG. 6A, "east"
When character recognition is performed from the three-character image of "Kyo" and "Miyako", the correct characters (underlined) are included in the 10 candidates for the two characters of "East" and "Miyako". For "Kyo", the correct characters are not included in the 10 candidates ("Kyo"
Is in the 11th candidate). In addition, "foot""standing""ward"
When the character recognition is performed from the three-character image of, the correct answer character (underlined) is included in 10 kinds of candidates for 2 characters of "foot" and "ku", but 10 kinds of for "standing" No correct character is included in the candidate of "(right" is included in the 11th candidate). In this example, "Adachi Ward, Tokyo" cannot be read.

On the other hand, in the case of the above (2), since the character recognition can be performed only in the address hierarchy to be read, the defect (unreadable) in the case of (1) can be solved. For example, if "Tokyo" is found, only a few letters that represent names such as "ward" and "city" will appear in the subsequent letters. In this case, rather than selecting the top 10 characters from all the character types that cover the regional names of all over Japan, choosing from the limited character types that appear after "Tokyo" is the correct answer for the names of "ward" and "city". Is more likely to be included in the 10 candidate character types, and finally the address reading rate is improved.

However, in the case of the above (2), since the time-consuming character recognition processing is repeated for each address notation hierarchy, the processing efficiency is poor and the processing time becomes long. For example, if the address hierarchy becomes deeper, such as "Kanagawa Prefecture", "Kawasaki City", "Suki Ward", "Yanagicho", the same character will be recognized many times (four times in this case) before the address judgment result is obtained. And the efficiency is low. In addition, since the character recognition of the next layer (for example, Sachi-ku) cannot be performed until the address recognition of the upper layer (for example, Kawasaki City) is completed, play time of the hardware used for character recognition is generated, which is more inefficient. Become. An object of the present invention is to provide an address reading system which is highly accurate and has high speed in automatic reading processing of an address notation.

[0009]

The address reading system of the present invention includes an address dictionary (50) storing correct address notation data (prefecture name, city name, other real name); Used for address display consisting of the above address notation hierarchy (1st hierarchy / prefecture name “Tokyo”; 2nd hierarchy “Adachi-ku”; 3rd hierarchy “Nishiaraicho”; 4th hierarchy ...) One or more address notation ("Tokyo", "Adachi-ku")
Recognizing each of the characters that make up "Nishiarai Sakae Town" ...
"Stand""ward""west""new""well""Sakae""town" ...
..) character recognition means (10, 30; ST1)
0, ST12); and one or more character type tables (40 including minimum character types necessary for recognition of each of the one or more address notation words (“Tokyo”, “Adachi Ward”, “Nishiarai Sakaemachi”)). ST16, ST20, ST24), and; from among the character recognition results (“East”, “Kyo”, “Miyako” ...) Recognized by the character recognition means (10, 30) The character type table (40; ST16) of the predetermined number of character types (10 types for each character of "East", "Kyo", and "Miyako") necessary for determining the address hierarchy ("Tokyo")
And an address recognition means (20) for detecting the corresponding address notation word (“Tokyo”) by subtracting the address dictionary (50) with each combination of the extracted characters.

[0010]

With respect to each character forming the address, the degree of similarity (or evaluation value) with all the character types used in the address notation of each address hierarchy is obtained in advance and is summarized in the character table of each address hierarchy. When reading the hierarchy of address constituent characters in order from the upper level (prefecture name), only the similar character type necessary for reading the next level (city / ward name, etc.) is configured based on the reading result of the upper level characters. By selecting and referring to a character table to be read, it is possible to improve the reading accuracy rate of address characters and shorten the processing time.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An address reading system according to an embodiment of the present invention will be described below with reference to the drawings. FIG. 5 shows an outline of a hardware configuration applied to this address reading system. The system of this embodiment includes a photoelectric conversion unit 102,
A character recognition unit 10 including an area detection unit 104, a line detection unit 106, a character detection unit 108, and a character recognition unit 110, and a character dictionary 30 that includes all basic characters referred to by the character recognition unit 110 for character recognition.
And an address recognition unit 20 that recognizes an address from a combination of read characters recognized by the character recognition unit 10 with reference to the character type table 40 and the address dictionary 50.
The character recognition unit 110 and the address recognition unit 20 include a microcomputer (CPU), and the functions of these units are realized by software running on the CPU. The character recognition unit 10 itself can be configured by a character recognition technique adopted in a conventional optical character reader (OCR) or the like.

First, the contents of the system of this embodiment will be briefly described. The photoelectric conversion unit 102 captures the surface image of the mail including the character to be read, binarizes the image, and outputs the original image data including the character image.

The area detection unit 104 extracts the area in which the address is described from the window signal and the fine scanning signal obtained by scanning the surface of the mail. That is, first, the image information (original image data) of the entire postal matter is two-dimensionally compressed, and the processed image is used to make a global image into blocks, thereby performing a global area detection that is not captured in details. Next, projection data for each block is generated, the complexity / direction of the line segment is determined, and the determination result is edited by previously given address area editing knowledge to determine the address area.

From the address area, a binarization threshold value candidate obtained from the image density histogram in the area, a writing direction candidate obtained from the detected address area position, and a character style candidate such as handwritten / printed characters are output. . In the address area editing knowledge, data based on a statistical survey regarding address description positions / areas extracted from a large amount of postal matter is incorporated as address area image knowledge.

On the surface of the postal matter, there are a sender's name, a sender's address, a correspondence, a stamp stamp such as a stamp, and various kinds of advertisements, as well as an address.
Area, complexity, print direction, etc. are used to build address area image knowledge.

The line detection unit 106 receives the output from the area detection unit 104, removes noise components such as a frame line surrounding a line of an address character or a ruled line on the address description, and separates and extracts each character line. Further, a print density histogram is obtained for each detected character line, and the optimum binarization threshold value for each detected character line is determined.

The character line unit detection algorithm here is basically the same as the address area detection algorithm.
However, the character line unit detection does not take a global view, and the underline area, the blank area, etc. are excluded by a more detailed image analysis than the case of the address area detection, and the image to be read (line image) is limitedly extracted.

The character detection unit 108 separates the images character by character according to the row image / binarization threshold value extracted / determined by the row detection unit 106. That is, it is assumed that the print type address is equal to or less than a threshold value determined in advance by the character line width data, and the others are handwritten addresses. Then, character separation in the row direction is detected from the projection information. The detection logic is a space detection caused by a space between characters, but regarding the contact between characters and the separation within characters, the detected character circumscribing frame is evaluated by fluctuation prediction based on a square and determined. Allow multiple detection candidates if the evaluation decision is ambiguous.

The character recognition unit 110 performs the same character recognition processing as a known optical character reader (OCR).
The character types to be identified prepared in advance in this character recognition process are all characters of Arabic numerals / kanji and katakana / hiragana and about 200 kanji used for addressing.
These characters (400 characters or less in total) are prepared for print type characters and handwritten characters (the maximum number of prepared characters is about 800 characters).

Since each of the above-mentioned units 102-110 can be constructed by conventional techniques, a more detailed description will be avoided. It is not realistic to attempt 100% correct character identification for an input character image containing ambiguity. Therefore, in the present embodiment, assuming that there is post-processing that uses address knowledge (address dictionary 50) later, 10 candidates are output in order from the highest similarity in the form of identification candidates in input image units. I have to. By doing so, it is possible to obtain a high identification rate that cannot be obtained by using a single identification candidate.

Further, regarding the division of the writing direction (vertical writing / horizontal writing) left as the ambiguity in the writing, the input character image is rotated by 90 ° and 180 ° for identification processing, and a series of rotation units are set. Output the character candidate matrix of.

The address recognition unit 20 determines the address from the character identification candidate matrix using a knowledge database (part of the dictionary 50) of the ward name, town name, large recipient name, etc. given to the delivery office. For example, taking the Adachi post office in Adachi-ku, Tokyo as an example, the standard notation town name prepared as an address is 31.
There are 10 types of seeds, such as Adachi Ward Office for large recipients.
There is a seed. However, since the notations of these names include various variations, learning is advanced according to the frequency of variations and the address knowledge data is increased unless it conflicts with other town names / large recipient names.

For example, taking the town name "Odai" in Adachi Ward as an example, the standard knowledge is "Odai", but this town name is "Kodai", "Otai", "Kotaimachi", "Odaicho", etc. There is a variation of. All of these modified names are stored in the address knowledge database (address dictionary 50) as pointing to "small stand" in Adachi Ward.

From the knowledge based on the address / major recipient's name including the above-described modifications, a word made up of a group of two or more characters concatenated is created, and the word is used to evaluate the character identification candidate matrix. For example, if the created word is “Tokyo” and the corresponding character identification candidate matrix is as shown in FIG. 6B, “East East group”, “East East capital”,
"Tokyo group", "Tokyo", "Kyoto group", "Kyoto capital",
From the group of words such as "Kyoto group" and "Kyoto group", "Tokyo"
The place name registered in the address dictionary 50 (as addressing knowledge) is finally selected.

After that, when all word matching is completed, knowledge processing is performed again for some word candidates. Similar to word matching, it is a process of evaluating whether or not a word sequence as an address or a large recipient name can be obtained while maintaining the adjacency relationship between words. For example, whether the word "Tokyo" is followed by the word "Adachi-ku" followed by the town name, or the word "adachi-ku" is followed by large recipient address words such as ward offices is evaluated. In this evaluation, a predetermined score is given for each word (the closer to the actual address and place name notation, the higher the score), and the address having the highest score can be output by accumulating the scores.

In order to identify the streets after the town name and the block (number / number etc.) except the mail addressed to the large recipient, the position of the detected town name on the character image and the address description format code (vertical writing / (Information such as horizontal writing / rotation) is passed to the next process as dependent information.

After recognition up to the town name, it becomes clear that the character image in the street name represents the chome / block. The address recognition unit 20 identifies this part of the chome / block, determines a final address, converts it into an address classification code, and outputs it.

In summary, the embodiment of FIG. 5 functions as follows. That is, the character recognition unit 10 refers to the character dictionary 30 to read / recognize handwritten characters or printed characters. The address recognition unit 20 refers to the character table 40 and creates one or more combinations of one or more read characters recognized by the character recognition unit 10. Next, the address dictionary 50 is looked up by the created combination of read characters, and the correct address display word corresponding to the read character combination is read from this dictionary 50. The character code of one or more address display words read from the dictionary 50 and the code of the address display numerical values (chome, number, number, etc.) following these address display words are collected, and the address read result (address classification code) is collected. ) Is output.

Next, referring to FIG. 1, the process from the address notation character after character recognition until the address is read will be described in several stages. (1) First, the address character written on the paper surface of the mail is the photoelectric conversion unit 102 and the area detection unit 10 shown in FIG.
4, line detection unit 106, and character detection unit 1
After 08, it is divided into each character as shown in the character image of FIG.

Next, each character image is collated / evaluated with all the character types registered in the character dictionary 30 in the character recognition unit 110 of FIG. 5 (step ST10, step ST12 of FIG. 1), FIG. 6 (a). "Character image" at the top of
The character recognition result as shown in the line is obtained. (Here, the character image is exemplified by the correct character "Adachi-ku, Tokyo ...", but these may include incorrect characters such as "East-higashi Ashito-ku ..." In order to significantly reduce the processing amount of the address recognition processing, the characters handled in the address recognition processing are limited to a certain number of candidates, for example, 10 candidates from the one having the highest evaluation value as shown in FIG. In this case, when the evaluation value of the correct answer of each character is 11th or less, the address reading becomes impossible and the reading rate of the correct answer decreases.

That is, if this candidate selection processing is performed before the character recognition unit 110 outputs the character recognition result,
Even when reading a character used for a prefecture name such as "Kyo" in FIG. 6A, a city that does not appear in the prefecture name,
Characters such as ward and town name go up to the top, and correct answer "Kyo" is 1
Frequently, there are cases where the candidate does not remain 0.

(2) Therefore, in the present embodiment, in order to prevent the correct characters from remaining in the candidates, the recognition result in the character recognition unit 110 (for example, 1-3-2 Nishiaraieicho, Adachi-ku, Tokyo) is used. All are sent to the address recognition unit 20. Then, in the address recognition unit 20, the address is read step by step from the upper hierarchy (prefecture name) of the address notation.

First, the prefecture is designated as the reading target (step ST14). Next, from the character type table 40, a list of character types required to read only prefecture names (1 prefecture, 2 prefectures, 43 prefectures) (character tables dedicated to prefectures)
Is read from the character dictionary / basic character candidate 30 (step ST16).

Then, for each address notation character (“East”, “Kyo”, “Miyako”) as a result of character recognition, the top 10 candidates in the character type list are selected. As a result, as shown in FIG. 6B, the corresponding character candidates designated by the prefecture are obtained.

That is, as in the example of FIG. 6 (a), when the top 10 candidates of the character recognition result are simply selected, the correct character of "Kyo" which is not included in the 10 candidates is selected by the prefecture. You will be in the 10 candidates. Combination word of these character candidates (east, Kyoto, county, capital, etc.) and address dictionary 5
The registered content of 0 is checked and the correct address notation (Tokyo) is read.

(3) In the example of FIG. 6B, "Tokyo" has been read, so the next read target is the city / ward / county / island name of the suburbs. Next, the Tokyo metropolitan area (name of other prefecture, for example, the method is the same in Kanagawa prefecture) is designated as the reading target (step ST18).

Next, from the character type table 40, the character type list (ward / city / county / island-specific character table; that is, area name character table) necessary for reading the local name (23 wards, 27 cities, etc.) in the suburbs is the basic character. It is read from the candidate 30 (step ST20).

Then, for each address notation character (“foot”, “standing”, “ward”) as a result of character recognition, the top 10 candidates in the character type list are selected. As a result, as shown in FIG. 6C, the corresponding character candidate under the designation of Tokyo is obtained.

That is, when the top 10 candidates of the character recognition result are simply selected (the example of FIG. 6A), the correct character of "standing" which is not included in the 10 candidates is selected by the designated Tokyo. You will be in the 10 candidates. The combination word of these character candidates (Adachi, Kita, Ward, Ara, Tachi, Nation, etc.) is collated with the registered contents of the address dictionary 50 to read the correct address notation (Adachi Ward).

(4) Next, Adachi Ward (name of another suburban area, for example, Fuchu City) is designated as a reading target (step ST22). Subsequently, a character type list (town name / place name-dedicated character table) necessary for reading the area name (town name, etc.) in Adachi-ku is read from the character type table 40 from the basic character candidate 30 (step ST24).

Then, for each address notation character (“West”, “New”, “I”, “Sakae”, “Machi”) as a result of character recognition, the top 10 candidates in the character type list are selected. As a result, a corresponding character candidate designated by Adachi Ward is obtained.

That is, the address notation characters (“west”, “new”)
Correct characters will be included in the top 10 candidates for each of "I", "Sakae", and "Machi". The combination word of each of these 10-character candidates (50 characters in total) is collated with the registered contents of the address dictionary 50, and correct address notation (Nishiaraeicho)
Is read.

If the read address is at the final stage, the process is terminated. If there is a lower layer, the read target is updated, another character type list (dedicated character table) is read from the basic character candidate 30, and the corresponding The address notation is read.

(5) Thereafter, the character codes of the prefecture name and the ward municipality name (Nishiaraimachi, Adachi-ku, Tokyo) that have been read are output (step ST26), and the block identification code (numerical code of chome / address / number) is output to this. ) Is added (step ST28). Then, a final address classification code is obtained from these character / numerical codes, and the address recognition unit 20 outputs the address classification code (step ST3).
0).

FIG. 2 shows a procedure of reading the address display (prefecture) of the first hierarchy among the address notation characters after character recognition in the address reading system according to the embodiment of the present invention. First, a character string starting with the prefecture name is input to the character recognition unit 10 in FIG. 5 (step ST100). Then, the CPU of the character recognition unit 110 refers to the character dictionary 30 to perform the recognition process of the input character string (step ST102). The storage capacity of the character dictionary 30 is normally 800.
Although it may be about characters, it may be larger (for example, 3000 characters) depending on the situation.

After the character recognition processing is completed, the unit 110 outputs a character string starting with the prefecture name (step ST
104). Here, assuming that the character recognition is correctly performed, from the unit 110, for example, "East" and "Kyo"
"Capital""foot""standing""ward""west""new""well""Sakae"
A set of character codes indicating "town" ... Is output. When this character recognition includes an error, a set of character codes such as "East", "East", "Group", "Feet", "North", "Ku" ... Is output.

The set of character codes output from the character recognition unit 110 is the upper layer of the address notation (prefecture name).
They are arranged in order from the lower hierarchy (city / ward / town name). A set (character string) of such character codes is passed to the address recognition unit 20.

In the address recognition unit 20, the reading target is set to "prefecture" in order to start reading the prefecture name (step ST106). Then, the CPU of the address recognition unit 20 extracts a prefecture character table composed of characters used only for the prefecture name from the basic character candidate 30 (step ST108). The storage capacity of the prefecture character table is enough for 100 characters.

The CPU of the unit 20 extracts three characters (for example, "East", "East", and "Group") from the beginning of the character string after character recognition, and for each character, from the prefecture character table, in descending order of similarity. Extract 10 characters (step S
T110). That is, a total of 30 characters are listed as candidates for reading the prefecture name.

Of the prefecture names, Kanagawa prefecture, Wakayama prefecture, and Kagoshima prefecture have four characters, but these prefecture names can be identified by the first three characters (Kanagawa, Wakayama, Kagoshima).
On the other hand, if only two characters are read from the beginning of the character string after character recognition, if the prefecture name is written in hiragana (for example, "Yamagata prefecture", "Yamamanashi prefecture", "Yamaguchi prefecture"), the prefecture The first name cannot be identified. If three characters are read, it is possible to identify the prefecture name even if the hiragana entry is included.

Next, the CPU of the unit 20 takes out one character from each of the extracted three character sets and each of the ten character group groups and sequentially creates three character combination words, and these three character words and the address dictionary 50 (knowledge database). A pair of address notation names registered in (including) are successively compared (step ST
112).

As a result of this pairwise comparison, if a matching word (registered word "Tokyo" for the character string "Tokyo") is found,
Or if you can find the corresponding word (registered word "Tokyo" for the character string "East and East") using the knowledge database,
The process moves to the next process (step ST114, Yes). If the matching word or the corresponding word is not found, it is considered unreadable and the process is terminated (step ST114, No), and the process proceeds to another mail address reading process.

When a matching word (registered word "Tokyo" for the character string "Tokyo") is found in the above pair comparison, it is checked whether the current address hierarchy is the final hierarchy (step ST116). If it is the final layer (step ST116, Yes), the CPU of the unit 20 displays the block code (chome / number /
No.) and the corresponding address classification code is obtained from the numerical code, and the address read determination result is output (step ST118).

However, at this point (step ST106-
Since the prefecture was read in step ST114), it is not the final layer (step ST116, No).

FIG. 3 is a continuation of the procedure of FIG. 2 and shows a procedure of reading the address display (ward / city, etc.) of the second hierarchy among the address notation characters after character recognition. Address recognition unit 2
0 means to start reading the next ward / city / county / island based on the comparison result in step ST112,
The reading target is set to "Tokyo" (step ST12).
0). Then, the CPU of the address recognition unit 20 extracts a Tokyo character table composed of characters used only for the place name under Tokyo from the basic character candidate 30 (step S).
T122). The memory capacity of the Tokyo character table is 100.
Just enough letters.

The CPU of the unit 20 uses the five characters (for example, "foot" and "north") of the fourth and subsequent characters of the character string after character recognition.
"Ward", "west", "new") is extracted, and for each character, 10 characters are extracted from the Tokyo character table in descending order of similarity (step ST124). That is, a total of 50 characters can be cited as a candidate for reading the place name of a ward / city / county / island in the suburbs.

Among the place names in the suburbs, the longest characters are 5 letters of Higashi Kurume City and Musashi Murayama City, and the shortest letters are 2 letters such as Kita Ward and Minato Ward. Most of these place names are It consists of three characters (Adachi Ward, Fuchu City, etc.). Long place names in Tokyo with 5 characters including Hiragana notation (such as "Higashi Kurume City""HigashimurayamaCity""HigashiyamatoCity")
Can be distinguished.

Next, the CPU of the unit 20 takes out one character from each of the extracted five character sets and each of the ten character set groups and sequentially creates a combination word of two to five characters and these two to five character words and the address. A pair of address notation names registered in the dictionary 50 (including the knowledge database) are successively compared (step ST126).

As a result of this pairwise comparison, if a matching word (registered word "Adachi-ku" for the character string "Adachi-ku") is found,
Or if a corresponding word (registered word "Adachi-ku" for the character string "Ashikita-ku") is found using the knowledge database,
The process moves to the next process (step ST128, Yes). If no matching word or corresponding word is found, it is considered unreadable and the process is terminated (step ST128, No), and another mail address reading process is started.

When a matching word (registered word "Adachi-ku" for the character string "Adachi-ku") is found in the above paired comparison, it is checked whether or not the current address hierarchy is the final hierarchy (step ST130). If it is the final layer (step ST130, Yes), the CPU of the unit 20 displays the block code (chome / number /
No.) and the corresponding address classification code is obtained from the numerical code, and the address read determination result is output (step ST132).

However, at this point (step ST120-
In step ST128), the place name (upper to middle hierarchy in the address notation) of a ward / city, etc. under Tokyo is read, and it is not the final hierarchy (step ST130, No).

FIG. 4 is a continuation of the procedure of FIG. 3 and shows a procedure of reading the address display (town / village or other place name) of the third hierarchy among the address notation characters after character recognition. The address recognition unit 20 sets the reading target to "Adachi-ku" in order to start reading the place name in Adachi-ku next based on the comparison result in step ST126 (step ST).
134). Then, the CPU of the address recognition unit 20 extracts an Adachi-ku character table composed of characters used only for place names in Adachi-ku from the basic character candidates 30 (step ST136). Adachi Ward character table also has a storage capacity of 10
0 characters is enough.

The CPU of the unit 20 has seven characters after the seventh character of the character string after character recognition (for example, "west" and "new").
"I", "Sakae", "town", "1", "Ding") are taken out, and for each character, 10 characters are extracted from the Tokyo character table in descending order of similarity (step ST138). That is, a total of 70 characters are listed as candidates for reading the place name in Adachi City.

Among the place names in Tokyo, the longest characters are 5 characters such as Nishiarai Sakaemachi and the shortest characters are 2 characters such as Aoi. With 7 characters including Hiragana notation, it is possible to distinguish long place names in Adachi City (such as "Nishiaraisakae Town" and "Nishiaraihon Town").

Next, the CPU of the unit 20 takes out one character from each of the extracted 7 sets and 10 types of character set groups, and sequentially creates a combination word of 2 to 7 characters and these 2 to 7 character words and the address. A pair of address notation names registered in the dictionary 50 (including the knowledge database) are successively compared (step ST140).

As a result of this pairwise comparison, if a matching word (registered word "Nishiarai Sakae town" for the character string "Nishiarai Sakae town") is found, or a corresponding word (for example, for the character string "Rooster Shindon Eitada") is utilized using the knowledge database. If the registered word "Nishiarai-cho" is found, move to the next process (step S).
T142, yes). If the matching word or the corresponding word is not found, it is considered unreadable and the process is terminated (step ST142, No), and another mail address reading process is performed.

When a matching word (registered word "Nishiarai Sakae" for the character string "Nishiarai Sakae") is found in the above paired comparison,
It is checked whether or not there is a character string that has not undergone pair comparison (that is, read determination has not been performed) among the input character strings (step ST144). If there is an uncompared character string (step ST144, Yes), unit 2
The CPU of 0 refers to the Adachi Ward character table for each character other than the block (things with numbers such as chome / ban / go), and compares the 10 character candidates used as place names in Adachi Ward with the uncompared character string. A selection is made for each character (step ST146). Then, all the combination words of each character candidate and the Adachi-ku place name in the address dictionary 50 are pair-compared (step ST148).

If no matching word or corresponding word is found as a result of this pair-wise comparison, it is considered unreadable and the process is terminated (step ST142, No), and another mail address reading process is started. If a matching word or a corresponding word is found (step ST142, Yes) and there is no uncompared character string after that (step ST144, No), the CPU of the unit 20 judges that time is the last hierarchy of the address notation and is read. The address code (Nishiarai-cho, Adachi-ku, Tokyo) is attached to the numerical code of the block display (chome / ban / go etc.), the corresponding address classification code is obtained from it, and the address read determination result is output ( Step ST15
0).

[0069]

According to the present invention, when an address is recognized in order from the upper hierarchy, only the character types required for each hierarchy are referred to, so that no unnecessary character candidates appear.
Since it is possible to select the optimum character candidate including the correct answer character, it is possible to improve the address reading correct answer rate.

Further, since the address recognition is started after finishing all the character recognition, the character recognition is performed during the address recognition processing (step ST14 to step ST30 in FIG. 1) when a large number of address readings are continuously performed. The hardware (10) can start the next character recognition of the address notation (step ST10 to step ST12 in FIG. 1). Then, since the character recognition and the address recognition can be performed in parallel, the processing efficiency is improved.

[Brief description of drawings]

FIG. 1 is a diagram illustrating an outline of how a correct address is read from an address notation character after character recognition in an address reading system according to an embodiment of the present invention.

FIG. 2 is a flowchart for explaining a procedure of reading the address display (prefecture) of the first layer among the address notation characters after character recognition in the address reading system according to the embodiment of the present invention.

FIG. 3 is a flowchart that is a continuation of the procedure of FIG. 2 and illustrates a procedure of reading the address display (ward / city, etc.) of the second layer among the address notation characters after character recognition.

FIG. 4 is a flowchart that is a continuation of the procedure of FIG. 3 and illustrates a procedure of reading the address display (town / village or other place name) of the third hierarchy among the address notation characters after character recognition.

FIG. 5 is a diagram illustrating an outline of a hardware configuration applied to an address reading system according to an embodiment of the present invention.

FIG. 6 is a diagram illustrating a specific example of how address notation characters are recognized.

[Explanation of symbols]

10 ... Character recognition unit, 102 ... Photoelectric conversion unit, 104
... Area detection unit, 106 ... Row detection unit, 108
… Character detection unit, 110… Character recognition unit (CP
U), 20 ... Address recognition unit (CPU), 30 ... Character dictionary (basic character candidate), 40 ... Character type table (prefecture name character table / region name character table), 50 ... Address dictionary.

Claims (5)

[Claims] 1. An address dictionary that stores correct address notation data and an address dictionary that is composed of one or more address notation layers 1
A character recognition means for recognizing each of the characters forming the above address notation word and outputting a character recognition result, and a minimum character type necessary for recognizing each of the at least one address notation word. Of the above character type table and the character recognition result recognized by the character recognition means, a predetermined number of character types necessary for determining the address hierarchy to be recognized from the character type table is extracted from the character type table, and each of the extracted characters is extracted. An address reading system comprising: an address recognition unit that detects the corresponding address notation word by drawing the address dictionary in combination. 2. The character type table includes a second character type table configured only of character types necessary for recognizing an area specified by the address notation detected by the address recognizing means, and the address recognizing means. The address reading system according to claim 1, wherein the address dictionary is configured to be searched by subtracting the address dictionary with each combination of characters extracted from the second character type table. . 3. An address dictionary storing correct address notation data, and recognizing each character constituting a plurality of address notation words used for address indication including at least first and second address notation hierarchies. A character recognition unit that outputs a character recognition result, a first character type table configured to include a minimum character type necessary for recognizing the address notation word of the first address notation hierarchy, and the character recognition unit. Of the recognized character recognition results, a predetermined number of character types necessary for the determination of the first address notation hierarchy are extracted from the first character type table, and the address dictionary is subtracted with each combination of the extracted characters. , A first address recognition means for detecting the corresponding first address notation word, and a minimum necessary for recognition of an area specified by the first address notation word detected by the first address recognition means. Including the character type From the second character type table configured and the character recognition result recognized by the character recognizing means, a predetermined number of character types necessary for determining the second address notation hierarchy are extracted from the second character type table. A second address recognition means for detecting the corresponding second address notation word by drawing the address dictionary with each combination of the extracted characters; and the second address recognition means for detecting the corresponding second address notation word. An address reading system comprising: means for outputting a corresponding address code from the first and second address notation words. 4. A character recognition means for recognizing a character corresponding to a plurality of character patterns and outputting an address notation character string, an address dictionary storing data of correct address notation, and a name of a prefecture. With reference to the prefecture character table composed of the characters used, the prefecture character table and the address dictionary, it corresponds to the combination of the first few characters of the address notation string recognized by the character recognition means. A prefecture name detecting means for detecting a prefecture name, a region name character table composed of characters used for a region name representing a place in the region of the prefecture name detected by the prefecture name detecting means, and By referring to the area name character table and the address dictionary, it corresponds to a combination of several characters appearing next to the prefecture name in the address notation character string recognized by the character recognition means. An area name detecting means for detecting an area name, and a means for outputting a corresponding address code from the prefecture name detected by the prefecture name detecting means and the area name detected by the area name detecting means are provided. An address reading device characterized by the above. 5. An address notation character string is output by recognizing characters corresponding to a plurality of character patterns, and a prefecture character table composed of characters used for the names of prefectures and correct address notation data are stored. The address dictionary that is being used to detect the prefecture name corresponding to the combination of the first few characters of the output address notation character string, and the region representing the place within the region of the detected prefecture name Detecting a region name corresponding to a combination of several characters appearing next to the prefecture name in the output address notation character string by referring to the region name character table composed of characters used in the name and the address dictionary Then, the address reading method is characterized in that a corresponding address code is output from the detected prefecture name and the detected area name.