JPH1185909A - Address recognizing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decrease the frequency of character recognition and to improve the character recognition precision by limiting the character category of standard patterns to be compared with respective input patterns to only character categories which can be connected by referring to an address tree when an address character string is recognized. SOLUTION: Character segmentation 401 cuts a character pattern candidate out of a character line image. Then character discrimination 402 discriminates the respective character pattern candidates and passes the results to a search processing 404. Then the search processing 404 dynamically generate a search tree for searching an address tree stored in an address dictionary 408 according to the character recognition result. After the generation of the search tree is completed, character string candidate selection 405 selects and outputs the path reaching a leaf from the starting node of the search tree, i.e., a recognized character string. Standard pattern selection 403 selects a standard pattern in a character category needed to expand a next node of the search tree is selected out of a standard pattern dictionary 406 and outputted to a character discrimination processing 402.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to character reading, and more particularly to address reading for reading an address written on a document such as a mail.

[0002]

2. Description of the Related Art The following procedure is generally known for reading an address written on a mail.

(1) The position of a destination character line is determined (character line extraction).

(2) Extract a character pattern (character extraction).

(3) A character category (character code) of each character pattern is identified (character identification).

(4) Interpret the character identification result as an address character string (character string collation).

Here, for (3), a pattern matching method for selecting the most similar to the input pattern from the standard patterns stored in advance is often used.

Now, as an example, consider a case where five character patterns as shown in FIG. 1B are cut out from a character line as shown in FIG. It is also assumed that the address to be recognized is as shown in FIG. 2 (virtual address).

In the conventional method, in this case, in the character identification processing, 19 characters of "Asahi Yasuiguchi City City Address Matsugami Shinnakamachi Tambaita Kitamotori" which is a character category to be read in each of the five character patterns. Compare with the standard pattern of the category (actually, about 400
In order to read zero characters and place names in one city, several hundred characters are read.) Then, the candidate character obtained as a result of the character identification processing is compared with the address character string, and the most similar character is selected.

[0010]

In the above-described processing procedure, it is necessary to compare the extracted character pattern with the standard patterns of all the character categories that can be present in the address. There was a problem that

A first problem to be solved by the present invention is as follows.
The purpose is to reduce the processing amount by narrowing down the character category of the standard pattern to be compared with each pattern by using knowledge about the arrangement of characters appearing in the address. In the above example, the city and town in the address are targeted.However, in this project, not only the city and town names, but also the names of street addresses, names of apartments and buildings, names, etc. were included. Target address.

When a multiple hypothesis testing method (character extraction method for automatically separating contacted handwritten characters) is used for character extraction (1558, National Institute of Electronics and Communication Engineers, 1558, National Convention, 1984), characters that need to be subjected to character identification need to be identified. There is a problem that the number of candidate patterns increases and the processing amount increases.

The second problem to be solved by the present invention is as follows.
An object of the present invention is to reduce the number of times of character identification required in the multiple hypothesis testing method and further improve the identification accuracy.

[0014]

According to the present invention, knowledge of a character string of an address is represented by a tree structure in which each character of the address character string corresponds to a node as shown in FIG. Hereinafter, such a tree structure is referred to as an “address tree”. Then, by using the character identification to search the address tree while selecting child nodes in the address tree (nodes connected to the terminal side of each tree structure via a branch), the address character string is obtained. recognize. At this time, the above problem can be solved by limiting the character category of the standard pattern to be compared with each input pattern to only the character categories that can be connected with reference to the address tree.

For example, when selecting the next node from the starting node, the first character pattern in FIG.
Matches with the standard pattern of the character type. If the character is similar to only one of the characters, a node corresponding to the character that is more similar is selected, and the search for a node is performed in the same manner. If the input character pattern is sufficiently similar to both characters at the next node, both nodes corresponding to both characters are expanded as candidates for the next node (s), and so on. Proceed. If the input character pattern is not similar to any character at the next node, the search for that branch is stopped. Thus, the leaves of the tree structure, or "town"
When a node corresponding to a character such as “Sato”, “Asahi”, “I”, “Matsu” or “Mouth” is reached, a character string from the starting node to the leaf is read. If a plurality of leaves are reached (this means that a plurality of address reading results remain as candidates), a more detailed judgment (for example, the certainty of the character identification result at the time of developing each node). Are selected based on the average value of.

Assuming that one standard pattern is used for each character category, the conventional method requires 19 (category) ×
Whereas 5 (patterns) = 95 matchings with the standard pattern are required, with this method, about 10 matchings are sufficient, and a significant reduction in processing amount is possible.
In addition, assuming that the place names in the whole country are read,
If the average value of 000 character types and character string length is 10 characters, 1
The conventional method requires about 40,000 matches to read one character string, whereas the present method requires several hundred matches.

Also, when the above-described method is applied to the multiple hypothesis testing method, the processing amount of character identification can be similarly reduced. In addition, since the character category is limited, when a pattern is cut out at an incorrect position, the probability that the pattern can be rejected by character identification is high, and subsequent pattern verification can be interrupted, further reducing the processing amount be able to.

This method is also effective in improving the reading accuracy. Parzen Density Estimation method and k-NN method that match multiple standard patterns in each character category,
It is known that it is effective for improving the accuracy of character identification. However, an increase in the number of standard patterns leads to an increase in the amount of processing, and the conventional method cannot sufficiently increase the number of standard patterns. However, in this method, since the character categories to be identified are greatly reduced, the number of standard patterns can be increased, and reading accuracy can be improved.

[0019]

FIG. 4 shows an example of the embodiment of the present invention.
This will be described with reference to the data flow diagram of FIG. In this specification, the data flow diagram is
Follows Saxon notation (J. Martin, "Software Structuring Technique", Modern Science, ISBN4-7649-0124-2 C3050P
5562E).

The present embodiment is an address recognition process in which an image of a character line on which an address is described is input and an address character string as a result of character string recognition is output. First, the character extraction 401 extracts character pattern candidates from a character line image. Character identification 40
2 is to perform character identification of each character pattern candidate,
Pass the result to 04. The search processing 404 dynamically generates a search tree for searching for an address tree stored in the address dictionary 408 according to the character identification result. After the generation of the search tree is completed, the character string candidate selection 405 selects and outputs a path from the starting node of the search tree to the leaf, that is, a recognized character string. The standard pattern selection 403 selects a standard pattern of a character category required for expanding the next node of the search tree from the standard pattern dictionary 406, and outputs it to the character identification processing 402.

FIG. 5 schematically shows an example of a process of generating a search tree generated by the present invention. Here, an example is given of a process for recognizing which of the addresses shown in FIG. 2 is the pattern sequence shown in FIG. 506 is a starting node of the search tree. From the starting node 506, a node 501, that is, a node corresponding to A in FIG. To develop the next node from node 501, change the first character pattern to “medium” “place”
Matches with the standard pattern of the two-character type. If the character is similar to any of the characters, a node corresponding to the similar character is newly generated. If it is determined that the character is similar to "medium" as in this example, a node 502 corresponding to the node B in the address tree in FIG. 3 is generated (in FIG. Branches that were not selected as a result are also indicated by broken lines). Next, in order to generate a node next to the node 502, the second pattern is a two-letter category of "living" and "mountain",
That is, matching is performed with a standard pattern of a character category that can be used as a node next to B in FIG. If the input character pattern is sufficiently similar to both characters, nodes corresponding to both characters are generated. Hereinafter, the expansion of the nodes is similarly repeated. If it is rejected like 503,
That is, if the input character pattern is not similar to any character at the next node, the search for that branch is stopped. When the node corresponding to the leaf of the address tree such as the node 506 is reached, the search for the branch is stopped. When all the branches for which the search should be continued are gone, all the search processes are completed.

After the search is completed, a character string on the path from the starting node to the leaf in the generated search tree is set as a character string recognition result. If a plurality of leaves are reached, a candidate is selected through more detailed judgment. For example, there is a method of selecting a candidate, such as selecting a candidate having a maximum average similarity of the character identification results of the patterns on the route.

In the above-described embodiment, a series of processing has been described using an address consisting of a city name and a town name.
The invention of the present application can be applied to a series of processing of addresses including notation of street addresses, names of apartments and buildings, etc., following city names and town names. This is because the street, street name, apartment name, and name are also finite character strings, and can be represented by the tree described above.

The present invention can be applied to the case where multiple hypothesis testing is used for character segmentation. In the multiple hypothesis test method, as shown in FIG. 6, a hypothesis on how to extract characters is expressed in a network. Hereinafter, such a network is referred to as a cut-out hypothesis network. On the extraction hypothesis network, the extracted pattern is represented by an edge, and the boundary of the pattern is represented by a node. The hypothesis on how to extract characters in a character line is expressed by a path from the leftmost starting node 601 to the rightmost end node 602 of the extracted hypothesis network.

FIG. 7 shows a process of generating a search tree when the character extraction process 401 of FIG. 4 generates an extraction hypothesis network as shown in FIG. In the multiple hypothesis testing method, a pattern following a certain pattern is not uniquely determined in advance. Therefore, in addition to the search process shown in FIG. 5, when developing the next node from each node, a process of extracting a pattern to be next character-identified and selecting it from the hypothesis network is required. Reference numerals 701 to 705 in FIG. 7 schematically show the selection process. In the search tree as shown in FIG. 7, each node corresponds to the node in the address tree in FIG. 3, as in FIG. For example, 7
10 and 711 correspond to A, and 712 and 713 correspond to B. However, the departure node 714 does not have a correspondence to the address tree as in the example of FIG.

FIG. 8 shows the format of a search tree table for storing search tree information. In the search tree table, each record is 3
It is represented by a table with two data items. Each record corresponds to a branch of the search tree. FIG. 8 shows a search tree as shown in FIG. 7, in which record 801 corresponds to branch 714, record 802 corresponds to branch 706, record 803 corresponds to branch 707, record 804 corresponds to branch 708, and record 805 corresponds to branch 709. First data item adrNode (next contact)
Stores a pointer to a node of the address tree corresponding to the node on the terminal side of the branch. For example, the branch of the search tree corresponding to 802 is 706, and child nodes 712 and 71 of 706
Since the node of the address tree corresponding to No. 3 is B, a
A pointer to B is stored in drNode. The second data item ptnCrd (pattern certainty factor) stores a value of the likelihood of child node expansion. As the value of ptnCrd, for example, a probability density or the like obtained in the process of character identification processing described later is used. The third data item strCrd (character string certainty factor) stores a value indicating the likelihood of a series of expansion from the root of the search tree to the child node. strCrd
As the value of (character string certainty factor), for example, an average value of ptnCrd of a branch on a path from the root to the child node or the like is used. The fourth data item ptn (input pattern) stores a pointer to a pattern used when expanding child nodes. If the parent node is the starting node, a null pointer (a value that cannot be used as a pointer) is stored, and the fact that there is no higher node is stored. The fifth data item prv (previous branch) stores a pointer to the parent node of the node. In the case of the starting node, a null pointer is stored in prv. This search tree table can be said to be a table in which the connection relation of the characters in the address is stored in a tree structure. In this example,
According to the processing procedure described later, a branch in which the next node cannot be expanded like 715 is not stored in the table.

FIG. 9 shows a data format of an address tree table for storing an address tree. In the address tree table, each record is 3
It is represented by a table with two data items. Each record corresponds to a branch of the address tree. Further, a unique integer value (id) is assigned to each node. First data item pr
tNode (parent contact) stores the id of the node on the upper side of the branch. In the second data item cldNode (child contact), the id of the lower node of the branch is stored. In the figure, the ids of the nodes are indicated by symbols "A" and "B" to correspond to FIG. 3, but an integer value unique to each node is actually stored. In the third data item code (character code), a character code corresponding to the branch is stored.

FIG. 10 shows the processing procedure of the present invention. The first step is a process of generating a cut-out hypothesis network from a character line image, that is, a character cut-out process 1001. For the character extraction process, for example, a known method such as that described in Murase et al., “Character Extraction and Recognition from Handwritten Character Strings Using Linguistic Information” (IEICE (D) vol. J69-D, No. 9) Used. The second step is a character identification / search processing 100
2. As will be described later, the character identification and the search process operate in cooperation with each other, and write data in the search tree table. The third step is a character string candidate selection process 10 for finding the most likely character string candidate from the search tree.
03. In the character string candidate selection process, adrNode
Finds a record corresponding to the leaf of the address tree from the search tree table, and selects a record having the largest strCrd value among these records. Next, by sequentially following the pointer prv from this record, a path from the starting node to the leaf is obtained, and a character string on the path is set as a character string candidate as a recognition result.

FIG. 11 shows the processing procedure of the character identification / search processing. First, in step 1101, the starting node is registered in the first record. Next, in step 1102, a pointer pc to the branch data of interest and a new branch data write destination pw are set in the second record of the search tree table. Next, in a loop 1103, each of pattern candidates at the left end of the character line (there can be a plurality of patterns as indicated by 603 and 604 in FIG. 6) is classified into the character category (address tree table) corresponding to the branch starting from the starting node of the address tree. Character identification processing is performed on the data item “code”). If the character identification is accepted, the branch of the search tree is registered at the address indicated by the pointer pw. Next, in a loop 1104, expansion of child nodes of the branches registered in the search tree table is repeated, and if the character identification result is accept, a new branch is added to the search tree table.
Thus, when there are no more branches to be searched, that is, when the values of the pointers pc and pw become equal, the loop is ended. Next, in step 1105, the most probable character string in the search tree is obtained and output as a recognition result.

In the loop 1104, a processing loop 1106 for expanding the next node from the unexpanded node, a step 1107 for sorting the expanded nodes by the character string certainty factor, and expanding more child nodes than a predetermined constant srcWd. In this case, the step of selecting only the upper srcWd nodes is executed. According to this step, even if many character identification results are accepted, an increase in the time required for processing can be suppressed by reducing the number of nodes.

FIG. 12 shows a hardware configuration in this embodiment. The present invention described above is applied to, for example, a recognition device 1205 for reading an address of a mail address reading / sorting machine as shown in FIG. The thick line in the figure indicates the flow of mail. A scanner 1201 controls image input. The scanner 1201 is connected to a preprocessing device 1204 that controls an image of an address character line. The preprocessing device 1204 connects to the bus 1210 of the recognition device via the input / output interface 1207. 1202
Is a delay line (delay transport path) provided on the mail transport path to secure the address reading time. 120
Reference numeral 3 denotes a sorter for sorting mail items based on the address reading result. The sorter 1203 is an input / output interface 120
8 to the bus 1210 of the recognition device. 120
Reference numeral 6 denotes an external storage device for storing a program required for the present invention, a standard pattern for character identification, an address tree, and the like. A hard disk or the like is used as the external storage device. An external storage device is connected to a bus 1210 of the recognition device via an input / output interface 1209. A recognition device 1205 has a bus 121 for connecting each part inside the recognition device.
0, an arithmetic processing unit 1211 for controlling the entire recognition device and for address recognition processing, and input / output interfaces 1207, 1208, and 1 for inputting and outputting to and from external devices.
209, a memory 1212 for storing programs, tables, dictionaries, etc. necessary for address recognition.

When the device is started, first, the external storage device 120
From 6, the program, standard pattern dictionary, address tree is memory 1
212. After loading is completed, the pre-processing device 12
When the character line image is supplied from 04 to the recognition device 1205, the character line image is copied to the memory 1212. Further, the arithmetic processing unit 1212 stores the cut-out hypothesis network and the search tree in the memory 1 according to the method described in the present invention.
The address is recognized while being generated on the 212, and the recognition result is transmitted to the sorter 1203. Sorter 1203 sorts mail items according to the recognition result.

The processing procedure described with reference to FIGS. 10 and 11 is based on the premise that the input character line is correctly cut out. For this reason, it cannot respond to the example as shown in FIG. 13A shows an example in which one character at the beginning of the address is completely missing, FIG. 13B shows an example in which noise is mixed at the beginning of the line, and FIG. 13C shows one example of the character at the beginning of the line. This is an example in which a part is missing. In the following, a description will be given of a method for expanding the above-described character identification / search processing method and coping with the lack of a line head and the inclusion of noise.

FIG. 14 shows a processing procedure corresponding to the lack of a head of a line. Step 1401 is a character cutout process similar to the flow described in FIG. Step 1402 is:
This is a reading process assuming that a character line is correctly cut out as described with reference to FIG. If an acceptable recognition result is obtained in step 1402,
As in the flow described with reference to FIG.
Execute 05. If no candidate to be accepted is obtained, step 1403 is executed.

Step 1403 is a process corresponding to a case where some characters at the beginning of a line are completely missing. in this case,
It is necessary to assume that the pattern corresponding to the root of the address tree is missing from the character line, and it is necessary to start character identification and search after checking which node of the address tree corresponds to the beginning of the line. is there. First, a pattern that can be the beginning of a line is obtained from the cut-out hypothesis network. Here, the “pattern that can be the beginning of a line” is a pattern that is in contact with the departure node of the network. For example, FIG.
For the input image as in (C), a character segmentation hypothesis network as shown in FIG. 15 is generated, the starting node is 1501, and the "patterns that can be the beginning of a line" are patterns 1503 and 1504. Next, all possible patterns are identified by all possible character categories in the address. If there is a character category to be accepted, a node corresponding to the category is obtained from the address tree. Then, a partial tree of the address tree having the node as a root is regarded as an address tree, and character identification and search are performed in the same processing procedure as described with reference to FIG. If step 1403
If a recognition result that is acceptable in is obtained, character string candidate processing 1405 is executed as in the flow described with reference to FIG. If no candidate to be accepted is obtained, step 1404 is executed.

Step 1404 is a process corresponding to a case where noise is mixed at the head of the line as shown in FIG. 13B and a case where a part of the character at the head of the line is missing as shown in FIG. 13C. In this case, the result of character identification of the pattern of the first character is rejected, and it is expected that no candidate to be accepted will be given in step 1403. For this reason, it is necessary to find the search start node based on the pattern serving as the second character. First, a pattern that can be the second character of the line, that is, a pattern that can be reached via one edge from the starting node of the cut-out hypothesis network (in the example of FIG. 15, patterns 1505, 1506, and 1507) is
Identify by all possible character categories in the address. If there is a character category to be accepted, a node corresponding to the category is obtained from the address tree. Then, a partial tree of the address tree having the node as a root is regarded as an address tree,
Character identification and search are performed in the same processing procedure as described with reference to FIG.

FIG. 16 shows an example of the principle of the character identification processing used in the present invention. FIG. 16 schematically shows a vector space of a feature used for character identification. Features include, for example, forest “pattern recognition” (ISBN4-88552-075-4 C3055,
A known device as described in IEICE (pp. 32-109) can be used. A plurality of character patterns are collected in advance for each character category, and these feature amounts are stored in the standard pattern dictionary 406 as standard patterns. In the character identification process 402, if it is determined whether or not the input pattern is similar to a certain character, Parzen D
ensity Estimation ("Statistical Pattern Recogniti
on, "ISBN 0-12-269851-7), it is determined whether or not the conditional probability density distribution of the character pattern in the character category is equal to or greater than a certain value. It is checked how many standard patterns of the character category exist within a given radius r.

[0038]

According to the present invention, the character category of the standard pattern to be compared with each pattern can be narrowed down using the knowledge of the character string of the address, and the processing amount at the time of address recognition can be reduced. Also,
When multiple hypothesis testing is used for character extraction, the number of patterns to be subjected to character identification can be reduced.

[Brief description of the drawings]

FIG. 1 shows an example of an address character string to be read.

FIG. 2 shows an address character string to be read.

FIG. 3 shows an example of a tree structure expression of an address character string.

FIG. 4 shows a data flow of one embodiment of the present invention.

FIG. 5 shows an example of a search tree generation process generated in the present invention.

FIG. 6 shows an example of a cut-out hypothesis network.

FIG. 7 shows an example of a search tree generation process generated in the present invention when the multiple hypothesis test method is used.

FIG. 8 shows a format of a search tree table.

FIG. 9 shows a format of an address tree table.

FIG. 10 shows a processing flow of an embodiment of the present invention.

FIG. 11 shows a processing flow of character identification / search.

FIG. 12 shows a configuration of the present invention.

FIG. 13 shows an example of an input image that is difficult to recognize.

FIG. 14 shows a processing flow corresponding to a missing line at the beginning of a line and noise contamination.

FIG. 15 shows an example of a cut-out hypothesis network when a head of a line is missing.

FIG. 16 shows an example of the principle of character identification.

[Explanation of symbols]

401 ... character extraction, 402 ... character identification processing, 40
3 ... standard pattern selection processing, 404 ... search processing, 4
05 ... character string candidate selection processing, 407 ... search tree, 40
8 ... Address tree.

 ──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Hiroshi Sakami 1-280 Higashi Koikekubo, Kokubunji-shi, Tokyo Inside the Central Research Laboratory, Hitachi, Ltd. Central Research Laboratory

Claims (2)

[Claims] An address reading method for converting image information written on a document into an electric signal, inputting the converted signal, and reading an address described in the image, wherein a character having an address described on the image is provided. Character line extraction processing to extract lines, pattern extraction processing to extract character patterns from character lines, and reference to an address tree that stores the concatenation of characters in an address in a tree structure A character category limiting process for limiting a character category of a character to be identified in the pattern, a character identifying process for identifying a character of the extracted pattern, and an address for searching a tree structure stored in the address tree storage unit according to the character identification result. An address reading method, comprising a tree search process. 2. The method according to claim 1, wherein when the pattern extraction processing cannot uniquely determine the character extraction method, a plurality of extraction method hypotheses are output, and the address tree search processing controls the search in accordance with the plurality of character extraction hypotheses. 2. The address reading method according to claim 1, wherein: