JPH09190507A - Address reader - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the processing speed of an address reader by detecting and segmenting the horizontally arrayed characters appearing in vertical writing way, adding the marks to the detected character patterns to show the horizontally arrayed characters and then assigning the different wild cards to the marks to carry out the collation of them. SOLUTION: A character segmentation/identification part 109 segments the character images of town areas out of the character rows for every character and identifies every character to convert it into a character code. As a result, the similarity corresponding to the character code serving as a candidate is acquired together with the information showing vertical or horizontal writing way. A town area collation part 110 extracts the town area information on addresses by means of a town area word dictionary 112. Then the part 109 segments the character images of address display numbers for every character out of the information on the character segmentation candidates and the head position of the address display number and identifies every character to convert it into a character code. The presence or absence is judged for the horizontally arrayed characters appearing in vertical writing way based on the vertical/horizontal writing judging information. If the presence of such characters are recognized, the segmentation candidates are generated for the horizontally arrayed characters and these characters are identified.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an address reading device for reading address information written on a mail piece in order to automatically sort the mail piece.

[0002]

2. Description of the Related Art As address information, for example, "11 Takada, Akigawa City" is used.
It is assumed that “1-2” is described. At this time, "Takada, Akikawa-shi" is defined as town area information, and the part described by the number "111-2" is defined as address display number information. There are various description forms having the same meaning in the address display number. For example, “111 No. 2”, “111 No. 2”, “11”
"1-2" and "2 of 111" mean the same address display number.

Conventionally, as a method of reading the address display number, a method using a wild card as disclosed in Japanese Patent Laid-Open No. 7-53946 is known. The outline of this method is as follows. The result of cutting out and identifying the character patterns representing the address display number is output as a candidate character lattice in which one character pattern is accompanied by a plurality of candidate characters. It is considered that the correct address display number can be extracted by matching the address display number with this candidate character lattice. However, the address display number has various notations as described above, and the number indicating the claw or number can be any number. Therefore, it is virtually impossible to hold the word of the address display number for all combinations of numbers and notations in terms of storage capacity and processing speed.

Therefore, the numbers in the candidate characters are not directly subjected to word matching, but a candidate character lattice is created by replacing the number candidates with a wild card representing an arbitrary number. For example, "n" for arithmetic numbers and "k" for Chinese numbers. On the other hand, in the word dictionary for matching, the word in which the notation pattern of the address display number is represented by a wild card is registered. The word corresponding to the notation pattern of "111 No. 2" is "nnn address n". Then, word matching is performed between the word expressed by the wild card and the candidate character lattice replaced by the wild card.

With this, it is possible to avoid the problem of the combination described above, and it is possible to extract the word representing the correct address display number notation pattern. And
As a result of word matching, if the numeral part of the wildcard word that has been selected as a higher-ranking candidate is restored to a numeral while referring to the original candidate character lattice, a correct candidate can be obtained.

[0006]

However, the above method has a serious problem in the case of a vertically written address. For example, when an arithmetic number is written in the vertical notation as shown in FIG. 3, the portion of the arithmetic number is usually written horizontally.
At this time, if the character pattern of that portion is cut out and recognized and the knowledge processing of Japanese Patent Laid-Open No. 7-53946 is applied, the following problem occurs. That is, since the arithmetic number "1" and the delimiter "|" following it are very similar in shape, the number "1" and the delimiter "|" It comes up as a candidate character. Therefore, even if the above-mentioned knowledge processing is applied, "111 | 2", "11 | 12", and "1 | 112" which are allowed as the notation pattern of the address display number are listed as candidates with similar confidence. To be

In order to prevent such an adverse effect, a character in vertical writing and a horizontally arranged character in vertical writing are detected, and it is assumed that the characters immediately before and after the horizontally arranged character string are delimiters by distinguishing between them. It is also possible to impose an additional penalty when a character type other than the delimiter comes. However, “111 | 1 | 2” as shown in FIG. 3 (b) is a delimiter of “|” immediately after the side-by-side character “111”, but since it is not a horizontal writing character, “111 | 11”
The harmful effect of allowing the "2" candidate will appear. In order to eliminate this adverse effect, it is possible to reduce unnecessary candidates by imposing various conditions such that the number for arithmetic with two or more digits can only have horizontal characters in vertical writing. However, the more we try to impose various constraints in order to increase the accuracy, the more complicated it is that each case requires detailed case classification, and if the notation changes, its maintenance Will also be difficult.

An object of the present invention is to solve the above problems without sacrificing the processing speed.

[0009]

In order to solve the above problems, the present invention provides three means.

The first means is to detect a horizontally arranged character appearing in vertical writing, cut it out, and add a mark to the detected character pattern to show that it is a horizontally arranged character.

According to the mark given by the first means, the second means separates the vertically written mathematical numbers and the horizontally arranged mathematical numbers which appear vertically even if they are the same. Matching is performed by assigning a wild card. First, we describe the candidate character lattice of the character identification result for word matching. In the case of FIG. 3 (a), the side-by-side portion "1"
The candidate character of the numeral for 11 ”is replaced with a wild card of“ N ”indicating that it is a horizontally arranged numeral. "2" is a wild card "n" for normal arithmetic digits
Replace with The other candidate characters are replaced as in the conventional case. In this way, a candidate character lattice is generated by replacing the candidate character lattice of the identification result with a wild card containing two types of arithmetic digits.

Next, the dictionary words to be matched will be described.
In the notation pattern dictionary of address display numbers, notation patterns in which partial horizontal alignment is mixed in vertical writing are used by using two types of arithmetic numeral wildcards such as "NNN" and "n". Store the expressed word. In this way, two types of wildcards, one for vertical writing and one for horizontal arrangement, are prepared for the arithmetic numerals, and the candidate character lattice using these wildcards and the words of the notation pattern of the address display number are collated. As a result, a correct writing pattern can be obtained even if there is a partial horizontal alignment in vertical writing.

The third means is to provide an index for high-speed word matching. In order to support various notation patterns of address display numbers, it is necessary to register many words including the notation patterns that appear horizontally in vertical writing, and it is a huge processing time to collate all these word groups. Need. Therefore, an index is provided so that unnecessary words are not matched. Two indexes are shown below.

The first index is to use the first and second candidate characters of the entered character as an index. It is highly likely that the first and second candidate characters that include a word character are correct words, and those that leak to that are unlikely. Therefore, by matching only the words including the candidate characters corresponding to the first and second characters, the correct solution can be searched at high speed without matching all the words.

The second index is a flag for vertical and horizontal writing. The address display number notation pattern can be divided into three types: one that appears only vertically, one that appears only horizontally, and one that appears in both.
For example, the above-described writing pattern "NNN-n" in which vertical writing is mixed with partial horizontal alignment appears only in vertical writing. Therefore, when performing vertical character cutting and identification, a flag relating to vertical and horizontal writing is output. On the other hand, the wildcard words are also provided with flags regarding vertical and horizontal writing, and by comparing these flags, unnecessary word matching can be avoided.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to FIGS.

FIG. 1 is a block diagram of the entire apparatus of this embodiment. The postal matter 100 is sequentially sent to the supply unit 101. In the supply unit 101, the mail piece passes through a predetermined position, and an image of the surface of the mail piece is captured by the image input unit 102 during the passage. The image input unit 102 digitizes an image signal from a photoelectric conversion element such as a line sensor and sends it to the address recognition unit 106. The address recognition unit 106 reads the address written on the surface of the postal matter and generates classification information. on the other hand,
The postal matter of which the surface image is captured is sent to the delay conveyance path 103. In the delay conveyance path 103, the mail is sent for the predetermined time for generating the sorting information from the surface image,
Move 03. In the classification unit 104, the address recognition unit 106
After sorting the mail items according to the sorting information from, the items are stored in the sorting shelf 105.

The address recognition unit 106 includes a control unit 107, an image processing unit 108, a character cutout identification unit 109, and a town area collation unit 11.
0, consisting of the address display number collation unit 111, the control unit 107
Controls the image processing unit 108, the character cutout identifying unit 109, the town area collating unit 110, and the address display number collating unit 111. The town area matching unit 110 uses the town area word dictionary 112, and the address display number matching unit 111 uses the house display number notation pattern dictionary 113.
Is accessed to automatically correct an error in the recognition result of the character cutout identifying unit 109.

The house display number notation pattern dictionary 113 is
It comprises a notation pattern word index dictionary 114 and a notation pattern word dictionary 115. Notation pattern word dictionary 11
Reference numeral 5 is a dictionary that stores the notation patterns of the blocks, and the notation pattern word index dictionary 114 is a dictionary that stores indexes for selectively retrieving necessary words from the notation pattern word dictionary 115.

FIG. 2 shows the flow of the entire processing of the address recognition unit 106 in FIG. The operation of the address recognition unit 106 will be described with reference to FIG.

In step 120, the image input unit 102 captures an image of the surface of the mail addressing, and inputs the image to the address recognition unit 106.

In step 121, the image processing unit 108 extracts the address area from the image of the entire surface with the address sent from the image input unit 102, and further extracts the image of each character line from the image of the address area. . For example, assume that the address character line in FIG. 3A is obtained.

In step 122, the character cutout identifying section 10
9 cuts out the character image of the town area for each character from the image of the character line obtained in step 121, and then recognizes the character for each cut out character and converts it into a character code. here,
As a result of character identification, similarities corresponding to a plurality of character codes as candidates are obtained. In addition, information on the direction of characters, that is, vertical writing or horizontal writing can be obtained.

Details of step 122 will be described with reference to FIGS. Fig. 4 shows the extraction candidate hypothesis for extracting a character pattern for each character from the address character line in Fig. 3 (a).
It is a figure expressed in the form of a network. FIG. 7 is a diagram showing a processing procedure of the town area character cutout identification.

In step 170, the character cut-out identifying section 10
Reference numeral 9 creates a cutout candidate of a character pattern and holds the cutout candidate as shown in FIG. 4 in the form of a network. At this time, the address display number "111" is also cut out in the vertical direction in the same manner as the town area above it.

In step 171, the character cutout identifying section 10
In step 9, character identification is performed on the character cutout candidates generated in step 170. As a result of the identification, a plurality of character codes as candidates and the corresponding similarity are obtained. Further, it is determined whether the vertical writing is horizontal writing by identifying the character patterns vertically and horizontally and comparing the similarities of the respective identification results.

In step 172, the certainty factor of each cut-out candidate is calculated based on the similarity of the character identification result of each cut-out candidate and the gap between the character patterns. That is, for example, a character cutout candidate that does not actually exist, such as 127 in FIG.
Even if character recognition is performed, a high degree of similarity cannot be obtained. Therefore,
It is possible to calculate the certainty factor of the cut-out candidate mainly based on the similarity of identification.

In step 173, based on the certainty factor of each character segmentation candidate calculated in step 172, the segmentation candidate is selected so that the certainty factor of the segmentation of the entire character becomes high, and the candidate character lattice corresponding thereto is selected. Is generated.
For example, a cutout candidate as indicated by a thick line in FIG. 4 is selected, and the identification result for the cutout candidate is shown in FIG.
Can be output as the candidate character lattice of. FIG. 10A is an example of the candidate character table of the candidate character lattice.
0 (b) is an example of the corresponding cost table. Here, costs are assigned in descending order of similarity when characters are identified. The candidate character table in FIG. 10 (a) is
From 1st place to mth place (m is arbitrary) for each written character number,
Character codes of candidate characters are arranged. In the cost table of FIG. 10B, the cost for the candidate character is stored at a position corresponding to the position where the character code of the candidate character table of FIG. 10A is stored. For example, the cost of the first-ranked candidate character “autumn” with the written character number 1 is “0”.

The above steps 170 to 173
By the above process, the character cutout identification process of the town area in step 122 of FIG. 2 is performed.

In step 123, the town area collating unit 110
However, the town area information of the address is extracted from the candidate character lattice obtained in step 122 using the town area word dictionary 112.
As a method of extracting town area information, a method is known in which an automaton is generated from a candidate character lattice and word matching is performed. The details are described in "Japanese Patent Application No. 1-262881". As a result of the processing in step 123, the town area character string “Takada, Akigawa-shi” is used as town area information, and the start position of the address display number that the end of the town area is the fifth character, that is, the address display number starts from the sixth character. Information is obtained.

In step 124, the character cut-out identification section 10
9 again cuts out the character image of the address display number for each character from the character cutout candidates obtained in step 122 and the information of the start position of the address display number obtained in step 123, and then cuts out each character. Recognize characters and convert them to character codes. Here, in the case of vertical writing, it is determined whether or not there is a horizontally arranged character in vertical writing by using the determination information of vertical writing and horizontal writing obtained in step 122. If it is determined that there is a horizontally arranged character, a character cutout candidate is generated again for the horizontally arranged character to perform character identification.

Details of step 124 will be described with reference to FIGS. 5, 6, 8 and 10. FIG. 6 shows an example in which horizontal alignment characters are detected from the extraction candidates of FIG. 4 and the extraction candidates for the horizontal alignment characters are added. FIG. 8 is a diagram showing a processing procedure for identifying character cutouts of an address display number.

In step 180, step 12 in FIG.
Using the address display number start position information obtained in step 3, it is determined for each cut-out candidate whether there is a horizontally aligned character from the position 128 in FIG. Here, the determination of whether or not there is a side-by-side character exists is that the character identification result of the cut-out candidate has a low degree of similarity, or "|"
It is performed depending on whether or not there is a delimiter such as “Ding”, “eye”, “ban”, “ground”, “go”. FIG. 5 shows an example in which side-by-side characters are detected.

In step 181, when it is determined that there is a horizontally arranged character based on the result of the horizontally arranged character determination in step 180, the character cutout recognition unit 109 creates a character pattern candidate, and the character pattern candidate is generated as shown in FIG. Add cutout candidates in the form of a network.

In step 182, the character cutout recognition unit 10
In step 9, a character identification is performed on the character cutout candidate additionally generated in step 181. As a result of the identification, a plurality of character codes as candidates and the corresponding similarity are obtained.

In step 183, step 17 in FIG.
Similar to 2, the certainty factor of each cutout candidate is calculated based on the similarity of the character identification result of each cutout candidate and the gap between the character patterns.

In step 184, based on the certainty factors of the respective character segmentation candidates calculated in step 183, the segmentation candidate is selected so that the certainty factor of the segmentation of the entire character becomes high, and the candidate character lattice corresponding thereto is selected. Is generated.
For example, a cutout candidate as indicated by a bold line in FIG. 6 can be selected, and the identification result for the cutout candidate can be output as the candidate character lattice in FIG. 11. FIG.
1 (a) is an example of a candidate character table in the candidate character lattice, and FIG. 11 (b) is an example of a corresponding cost table. At the entry character numbers "6-1", "6-2", and "6-3", a horizontal alignment flag indicating that the character pattern of the entry character number "6" in FIG. 10 is identified as a horizontal alignment character is set. There is.

The above steps 180 to 184
By the processing of step 2, the character cutout identification of the address display number in step 124 of FIG. 2 is performed.

In step 125, the address display number collation unit 111 uses the residence display number notation pattern dictionary 113 to extract the address display number from the candidate character lattice obtained in step 124. Details of step 125 are shown in FIG.
This will be described in detail with reference to FIGS. 9 and 11 to 17. In FIG. 9, step 191 is the feature of the present invention.

In step 190, step 12 in FIG.
The candidate character lattice generated in 4 is input to the address display number matching unit 111 in FIG.

In step 191, a wildcard lattice is generated from the candidate character lattice obtained in step 190. Here, the wildcard lattice is a lattice in which the numbers in the candidate character lattice are replaced with wildcards representing arbitrary numbers in order to match the address display numbers.

Details of step 191 are shown in FIGS.
This will be described with reference to FIG.

FIG. 12 shows a conversion table used to generate a wild card lattice, which will be described in detail below.

The table of "numerals" in the classification is a table relating to the numbers representing the chome, block, and house display number, and if there are arbitrary numbers and Chinese numbers in the vertically-written candidate character lattice. , Are converted into wildcards “n” and “k”, respectively. However, in order to distinguish horizontally aligned numbers in vertical writing from vertical writing "n",
Converted to a wildcard "N". Similarly, in the "alphabet" table, vertically written alphabetic characters are converted into "a" and horizontally arranged alphabetic characters in vertical writing are converted into "A".

The "delimiter" table is a table relating to the characters "chome" and "-" used to separate the numbers of the chome and the block when there is "3 chome 8-1" in the address, for example. Is. Here, the vertical "Ding",
The "eye", "ban", "ground", "no", and "no" are not converted and the characters are used as they are. On the other hand, vertically written "|", ".", And "/" are converted into wildcards called "|", but "-", which appears in horizontal alignment during vertical writing,
"-" Is converted to "-".

The "other" table is a table relating to characters other than the "numeric" table and the "delimiter" table, and all the characters other than those described above are converted into a wildcard "e". That is, "other" refers to all but the characters related to the chome, block, and house display number.

FIG. 13 is an example in which a wild card lattice is generated from the candidate character lattice of the address display number shown in FIG. 11 using the conversion table of FIG. The wild card lattice is composed of two parts, a wild card table shown in FIG. 13 (a) and a cost table shown in FIG. 13 (b).

The wild card lattice shown in FIG. 13 is generated from the candidate character table having the highest rank in the candidate character table shown in FIG. First, the first-ranked candidate character is converted into a wildcard and written in the first-ranked place of the wildcard table shown in FIG. At the same time, the cost corresponding to the converted candidate character is written in the corresponding place in the cost table of FIG. 13 (b).

Next, the second-ranked candidate character is examined, and if it belongs to the same wildcard as the first-ranked candidate character, it is duplicated and therefore omitted. If they belong to different wildcards, the costs corresponding to the wildcards are calculated in the wildcard table of FIG.
It writes in the cost table of FIG.13 (b). Hereinafter, the same thing is repeated for all the candidate characters.

For example, consider conversion of the column of the entry character number "6-1" in the candidate character table of FIG. 11 (a). "6-
Since a flag indicating that "1" is a side-by-side character is set, first, the first-ranked candidate character "1" is converted into a wildcard "N" and entered in the wildcard table of FIG. 13 (a). Write in the first place of the character number "6-1". At the same time, the cost "0" corresponding to the candidate character "1" is written in the corresponding place in the cost table shown in FIG. 13 (b).

Next, looking at the second place, the candidate character "no" belongs to a different wild card "no", so it is written in the second place of the cost table of FIG. 13B. Further, looking at the third place, the candidate character “|” does not exist as a side-by-side character, so it is omitted. Similar processing is performed for the fourth place, and when looking at the fifth place, the candidate character "7" is the first place "1".
Since it belongs to the same wildcard "N" as in the above, it is omitted and nothing is written in the wildcard table of FIG. 13 (b).
Hereinafter, the same processing is performed on the candidate characters of all ranks.

Through the above processing of step 191, a wildcard lattice is generated from the candidate character lattice obtained in step 190.

At step 192, an automaton for word matching is generated from the wild card lattice generated at step 191. Details of step 192 will be described with reference to FIG.

FIG. 14 shows a process of extracting a word from a dictionary and performing word matching of a notation pattern of an address display number using an automaton. First, the automaton 2
41 will be described. To match the dictionary word that represents the address display number pattern with the wildcard lattice,
A finite state automaton 241 is generated from the wild card lattice shown in FIG. Then, the automaton 241 sequentially inputs the dictionary words represented as a character string, and calculates what the average cost of the words is.

In the automaton 241, a circle indicates a state, and a number written therein indicates a state number. In addition, between states corresponds to each written character position of the word,
Lines between states indicate transition paths. The character on the left side of the transition path indicates that when the character is input to the automaton in a certain state, the transition path is followed to transition to the next state. Further, "other" on the transition path represents all characters other than those explicitly specified as the characters corresponding to the transition path.
The number in brackets [] on the transition route is the cost of having a transition by following the route.

As an example, using the automaton 241,
Consider the cost calculation when the word "NNN-n" 240 is entered. First, the cost [0] of "N" is added when transitioning from state 1 to state 2, the cost [2] of "N" is added when transitioning from state 2 to state 3, and so on. Goes on. Then, after all the transitions of the number of characters of the word are completed, the integrated cost is divided by the number of characters of the word to obtain the average cost 242 of the word.

Next, the generation of the finite state automaton 241 from the wild card lattice of FIG. 13 will be described. First,
A transition path between the state 1 and the state 2 of the automaton 241 of FIG. 14 is generated, and the candidate character of the entry character number “6-1” in the wildcard table of FIG. At the same time, the costs in the cost table shown in FIG. 13B corresponding to the candidate characters are similarly assigned to the transition paths between the states 1 and 2, respectively. Next, "other" is generated as a transition path that represents all characters other than the specified one, and the cost is set to 15. Similarly, states and transition paths are generated one after another from the wild card lattice. Here, the cost of the transition route does not necessarily have to be the above value, and may be an arbitrary number.

By the above processing of step 192, the automaton 241 is generated from the wild card lattice.

In step 193, the automaton 241 generated in step 192 and the notation pattern word index dictionary 114 and the notation pattern word dictionary 115 shown in FIG. Details of step 193 are shown in FIGS.
5, it demonstrates using FIG. FIG. 15 is a PAD showing the flow of the automaton matching process. FIG. 16 is a diagram showing the configurations of the address display number word dictionary and the address display number word index dictionary.

First, the structure of the dictionary shown in FIG. 16 will be described. The notation pattern word dictionary 115 is a word that represents notation patterns of chome, block, and house display number using wild cards,
And a dictionary that stores the search information and attributes thereof. For example, “NNN-n”, “NNN's n”, etc. are stored as the words representing the notation patterns of chome, block, and house display number. Notation pattern word index dictionary 11
Reference numeral 4 is a dictionary that stores an index for selectively retrieving words necessary for matching from the writing pattern word dictionary 115. The index is the first of the dictionary words
The first and second characters are used.

The notation pattern word dictionary 115 includes a word 278 representing an address display number notation pattern, a relative address 279 between words having the same second character as the second character, and a flag 280 indicating vertical and horizontal writing.

The flag 280 indicating vertical / horizontal writing stores a number of 1 if the word representing the block notation belongs to horizontal writing, 2 if it belongs to vertical writing, and 3 if it belongs to both vertical and horizontal writing.

Notation pattern word index dictionary 114
Is composed of two tables, a first character index table 270 and a second character index table 274.
The first character index table 270 includes a table 271 in which the first character number of a dictionary word is stored, the number of words 272 having the same first character 272, and a pointer table 273 to the notation pattern word dictionary 115. The second character index table 274 similarly stores a table 275 storing the second character number of the dictionary word, the number of words 276 having the same second character 276, the notation pattern word dictionary 11
5 pointer table 277.

Next, the flow of processing when searching for words in the notation pattern word dictionary 115 using the notation pattern word index dictionary 114 will be described. The arrow indicated by a solid line indicates the flow of search when searching for a word in which the first character is the same character using the first character index table 270. The arrow indicated by the dotted line shows the flow of search when searching for a word in which the second character is the same character using the second character index table 274.

For example, when retrieving a word in which the first character of the word is "N" from the dictionary, the number of words in which the first character of the first character index table 270 is the same character
The pointer P1 (1) of m1 and "N" is referred to. Since the pointer P1 (1) stores the first pointer of the word whose first character starts with "N" in the address display number word dictionary, that word is referred to. In the word dictionary, words with the same first character are arranged consecutively, so if you find the first word starting with "N", you can search words one after another by simply incrementing the pointer. it can. Then, when the search is performed m1 times, the end of the word starting with "N" is reached, and the word search is terminated at that point.

The same applies when searching for the second character of a word as an index. For example, when searching for a word in which the second character of the word is "Ding" from the dictionary, the number of words q1 in which the first character of the second character index table 224 is the same character, and a pointer of "Ding" See P2 (1). The pointer P2 (1) stores the pointer of the word whose second character is "Ding" in the writing pattern word dictionary, so that the word is referred to. After that, by referring to the relative address 279 between words having the same second character of the word and shifting the pointer, it is possible to search for the word "Ding" having the same second character. Then, when the search is performed q1 times, the second character ends at the end of the word with the same "ding", so the word search ends there.

Here, when it is known whether the address is written vertically or horizontally in the character segmentation / character recognition in step 122 of FIG. 2, by referring to the flag 280 indicating vertical / horizontal writing, among the searched words You can narrow down and extract the corresponding words from.

Next, the specific processing contents of step 193 will be described using the processing flow of FIG. 15 and FIGS. 13, 14 and 16.

In step 250, a flag representing vertical and horizontal writing of the address obtained in step 122 of FIG. 2 is set for the address to be collated.

At step 252, the first entry of the entry character number "6-1" in the wild card table of FIG. 13 (a).
The candidate character of rank is taken out, and the number of dictionary words 272 and the pointer 273 of the word dictionary are acquired by referring to the first character index table 270 of FIG.

In step 254, the word pointed to by the pointer is searched, and it is checked by referring to the corresponding table 280 in the dictionary whether this word is consistent with the vertical / horizontal writing flag set in step 250. If there is no contradiction, go to step 255.

In step 255, the retrieved word is searched for in FIG.
4 is input to the automaton 241 and the average cost is calculated while transitioning the states.

In step 253, the processes of steps 254 to 255 are repeated for the number of dictionary words obtained in step 252.

At step 251, the candidate character of the next rank of the entered character address "6-1" in FIG. 13 (a) is taken out,
The processing from step 252 to step 255 is repeated several times as many as the candidate characters of the entered character address "6-1".

In step 257, the first candidate character of the entry character number "6-2" in the wild card table of FIG. 13 (a) is taken out and the dictionary is referred to by the second character index table 274 of FIG. The number of words 276 and the pointer 277 of the word dictionary are acquired.

In step 259, the word pointed to by the pointer is searched, and it is checked by referring to the corresponding table in the dictionary whether this word is consistent with the vertical / horizontal writing flag set in step 250. If there is no contradiction, go to step 260.

In step 260, the retrieved word is searched for in FIG.
4 is input to the automaton 241 and the average cost is calculated while transitioning the states.

In step 258, the processes of steps 259 to 260 are repeated by the number of dictionary words obtained in step 257.

At step 256, the candidate character of the next rank of the entered character address "6-2" of FIG. 13A is taken out, and the processing from step 257 to step 260 is performed.
Repeat “2” candidate character several times.

In step 261, the words obtained by the processing of steps 250 to 260 and the costs thereof are rearranged in the order of increasing and decreasing.

In step 262, the upper L (L> 1) words with the low cost rearranged in step 261 are selected.

The above steps 250 to 262
By the processing described above, the automaton character string collation of step 193 in FIG. 9 is performed, and the upper L words having a small average cost and the costs thereof are obtained.

At step 194, the wild cards representing the numbers "N", "n", and "k" of the words represented by the top L wild cards obtained at step 193 are restored to the original numbers. To generate candidates. Here L
FIG. 17 shows a result of restoring the word “NNN-n” 240 of FIG. 14 into a numeral when = 1. In the process of digit restoration, the candidate character table and the word “NNN-” shown in FIG.
n ”240 is aligned. After that, the candidate numeral of the place corresponding to the numeral "N" is taken out from the candidate character table of FIG. 11 (a), and the actual chome and block are generated. Also,
The costs corresponding to the numbers are extracted from the cost table of FIG. 11B, and the costs for the restored notation pattern words are added up.

In step 195, P (P> 1) candidates having a small cost are selected from the candidates restored to numbers in step 194 as the address display number collation result. If P = 2, “111 | 2” and “171 | 2” are output as the address display number collation result.

Steps 190 to 195 above
By the processing up to step 125, the address display number collation in FIG. 2 is performed, and the candidate of the address display number is obtained as the collation result.

At step 126, the town area candidate “Takada, Akikawa-shi” obtained by the town area collation at step 123 and the address display number candidates “111 | 2”, “at the address display number collation at step 125. 171 | 2 ”is connected to generate an address candidate. In this example, “111 | 2, Takada, Akigawa-shi”,
"171 | 2 Takada, Akigawa-shi" is obtained. Further, using this address information, control information for controlling the sorting unit 104 in FIG. 1 is generated.

[0087]

The present invention has the following three effects. First, it is possible to correctly recognize the address display numbers of various writing patterns by effectively using the information that the characters are horizontally arranged characters in vertical writing. According to the present invention, for a candidate character lattice that is the result of extracting and identifying a character pattern, even if the same arithmetic number is used, vertical arithmetic numbers and horizontal arithmetic numbers that appear vertically are different wildcards. Is assigned for verification. Therefore, it is possible to extract from the candidate character lattice a notation pattern that includes the horizontally aligned characters in vertical writing. As a result, it is possible to extend the conventional framework for word matching without incorporating complicated rules such as side-by-side characters that can be only arithmetic digits, or only delimiters such as "|" immediately before and after the side-by-side characters. It is possible to recognize the address display number with high accuracy.

The second is that the notation pattern of the address display number can be examined in detail, but it is processed at high speed. First, when searching a word from a dictionary,
Since the candidate character group corresponding to each character pattern is searched as an index, the number of words to be matched can be reduced. Furthermore, since each word has an attribute corresponding to vertical and horizontal writing, if it is known in advance whether the address display number to be recognized is vertical or horizontal writing, it is possible to prevent unnecessary word collation. This enables high speed processing.

Thirdly, since a dictionary collation method is used instead of a rule incorporation method by a program, when a new notation pattern occurs, it can be easily registered in the dictionary and maintenance is easy.

[Brief description of the drawings]

FIG. 1 is an overall configuration diagram of a device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing the flow of the entire processing of the apparatus of the first embodiment of the present invention.

FIG. 3 is a diagram showing an example of an input image of a vertically written address.

FIG. 4 is a diagram in which character cutout candidates are expressed in the form of a network for the input image of FIG.

5 is a diagram showing an example in which horizontally aligned characters are detected in the network of FIG.

FIG. 6 is a diagram showing an example in which cutout candidates are added to the horizontally arranged characters detected in FIG. 5;

FIG. 7 is a PA showing a flow of processing for character cutout identification of a town area.
It is D.

FIG. 8 is a diagram showing a flow of processing for identifying character cut-out of an address display number including horizontally arranged characters in vertical writing.

FIG. 9 is a diagram showing a flow of processing for performing address display number matching.

FIG. 10 is a diagram showing an example of a candidate character lattice for performing town area matching.

FIG. 11 is a diagram showing an example of a candidate character lattice of an address display number area.

FIG. 12 is a diagram showing a conversion table for generating a wildcard lattice for performing address display number matching from candidate character lattices.

13 is a diagram showing an example of a wildcard lattice generated from the candidate character lattice of FIG. 6 using the conversion table of FIG.

FIG. 14 is a diagram showing a process outline of automaton word matching.

FIG. 15 is a flowchart P showing a flow of processing of automaton word matching.
It is an AD diagram.

FIG. 16 is a diagram showing configurations of an address display number word index dictionary and an address display number word dictionary.

FIG. 17 is a diagram showing an example of address display number candidates in which a numerical portion is restored from the address display number word as a result of collation.

[Explanation of symbols]

100 ... Mail, 101 ... Mail supply section, 102 ... Image input section, 103 ... Delayed transport section, 104 ... Sorting section, 105
... Division shelf, 106 ... Address recognition unit, 107 ... Control unit, 10
8 ... Image processing unit, 109 ... Character recognition unit, 110 ... Town area collation unit, 111 ... Address display number collation unit, 112 ... Town area word dictionary, 113 ... Address display number notation pattern dictionary, 114
... notation pattern word index dictionary, 115 ... notation pattern word dictionary.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Katsumi Marukawa, Katsumi Marukawa, 1-280, Higashi Koikeku, Kokubunji, Tokyo, Central Research Laboratory, Hitachi, Ltd. (72) Tatsuhiko Kagehiro, 1-280, Higashi Koikeku, Kokubunji, Tokyo Hitachi, Ltd. Inside the Central Research Laboratory (72) Inventor Hiroshi Sako 1-280 Higashi Koikekubo, Kokubunji City, Tokyo Inside Hitachi Central Research Center (72) Inventor Yoshihiro Shima 1-280 Higashi Koikeku, Tokyo Kokubunji City Central Research, Hitachi, Ltd. In-house (72) Inventor Shigeru Watanabe 1 Ikegami, Haruoka-cho, Owariasahi-shi, Aichi Hitachi Ltd. Office Systems Division

Claims (2)

[Claims] 1. An image input by converting an image on a mail into an electric signal in an address reading device that detects character information on the mail and reads address information including town area information and address display number information. Input means and character cut-out identification means for extracting and identifying character information from the input image and outputting an identification candidate character group for each cut-out character pattern. A character cut-out identifying means for detecting and re-cutting out character information for identification, outputting an identification candidate character group for each cut-out character pattern; and a recognition candidate character group output from the character cut-out identifying means. , A town area collating means for recognizing a town area by collating with a town area dictionary storing town area information, and an address display number area for detecting the head of the address display number area based on the recognition result of the town area collating means Inspection Means and each character pattern after the head of the address display number area determined based on the output from the address display number area detecting means, as the output result of the character cut-out identifying means, horizontal arrangement from the vertically written address information is performed. The numbers in the candidate character group of the identification result for each character pattern detected as a character are divided into vertical writing and horizontally aligned characters in vertical writing, and each is replaced with a wild card that represents an arbitrary number. A wildcard conversion means for converting, an address display number notation pattern dictionary that holds various notation patterns of the address display number expressed by the above wildcard as words, an output result of the wildcard conversion means, and an address display number notation pattern dictionary The address display number matching means for recognizing the notation pattern of the address display number by matching the word with The output result of the display number collating means is compared with the candidate character group output from the character recognizing means, the numbers in the candidate character group replaced by the wildcard are restored, and the number restoration for outputting the address display number candidates is performed. An address reading device comprising means. 2. The address reading device according to claim 1, wherein the character direction detecting means for detecting whether the character group representing the address information on the mail is vertically written or horizontally written, and each notation pattern. For the address display number notation pattern dictionary in which the word has the attribute regarding whether the notation pattern appears in vertical writing or the writing pattern that appears in horizontal writing, and the candidate character group of each character pattern in any position replaced with a wildcard, Using the character code of the candidate character group as an index, word writing means for searching words from the house display number notation pattern dictionary and vertical writing from the words searched by the word searching means based on the output of the character direction detecting means An address reading device, comprising: a word selecting unit that reads out only words having the same horizontal writing attribute.