JP3313272B2 - Address reading method and identification function weight vector generation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an address reading / separating apparatus which reads an address written on a document such as a mail and sorts the document.

[0002]

2. Description of the Related Art The following three functions are required to read an address from a character sequence (address character line) in which the address is described.

(1) Cut out a character pattern. (Character cutout) (2) The character type (character code) of each character pattern is identified. (Character identification) (3) The character identification result is interpreted as an address character string.
(Character string collation) Of these, character lines written by hand have large variations in character pitch, and thus character extraction in (1) is particularly difficult.

Conventionally, it has been known that a multiple hypothesis testing method is effective in accurately extracting a character from a handwritten character line. This is a method in which a plurality of hypotheses are set for character extraction, character pattern candidates are extracted, and then a correct hypothesis is determined by character identification or character string collation. For example, Fujisawa et al.'S method of testing a hypothesis based on the results of character identification (1558, IEICE General Conference 1558)
“Character cutout method that automatically separates touched handwritten characters”). Also, as a method for determining the success or failure of cutting out a character pattern candidate based on the outline of a character, the method of Ishidera et al. (1995 IEICE General Conference D-576).
"Character segmentation method for reading handwritten addresses"). Also, according to the character identification result and the character string collation result,
Murase et al.'S method for testing hypotheses (IEICE (D) vol. J69-D, No. 9 "Character extraction and recognition from handwritten character strings using linguistic information") and Oi's method (Technical Report PRU92-40 "Chome / Block Recognition Method in Address Reading").

[0005] The multiple hypothesis test method is a method in which candidates for character patterns are extracted and tested using the results of character identification and character string collation. For example, for the address character string shown in FIG. 19B, character pattern candidates as shown in FIG. 20A are generated. In FIG. 20A, the relationship between character pattern candidates is represented by a graph in which a boundary between patterns is a node and the pattern is an arc. In this manner, the relationship between the character pattern candidates represented by a graph is referred to as a cut-out graph. By expressing the relationship between the character pattern candidates in this way, the extraction of the character pattern can be replaced with the problem of finding the optimal path from the start point to the end point of the extraction graph.

[0006]

As described above, in the prior art, for example, the addresses shown in FIGS. 19A and 19B both include the pattern shown in FIG. 19C. Whether the character (c) should be cut into one character (“plate”) or two characters “water counter” cannot be determined by the conventional method using the character identification result. This is because, in either of the extraction methods, a high similarity is obtained as a result of character identification of the extracted pattern. Also, if both "Takamachi, Nakazumi City" and "Mizutani, Nakazumi City" can be used as the address, which of the above two cutting methods is correct is determined by the conventional method using character string matching. Can not. If the positional relationship between the size of the character pattern and the preceding and succeeding patterns is used, characters can be correctly cut out even in such a case. For example, (a) is written with a large space between characters, whereas the gap g of the pattern in (c) is small. Therefore, it can be determined that (c) is one character. On the other hand, (b) is written with a small character interval,
It is appropriate to interpret the gap g of the pattern in (c) as the space between characters. Therefore, in (b), it can be determined that there are two characters.

On the other hand, if an inappropriate arc is deleted from the character identification result in FIG. 20A, a cut-out graph as shown in FIG. 20B is obtained. However, in the portion of FIG. 19C, a plurality of extraction methods remain as candidates, and there is a possibility that character extraction is erroneous in the conventional method. However, in the conventional method, it is difficult to determine how to cut out characters based on such a relationship with the preceding and following patterns.
For example, the method of Fujisawa et al. And the method of Ishidera et al. Determine the validity of each pattern, and cannot use the relationship with the preceding and succeeding patterns. Also, Oi's method,
The method of Murase et al. Uses the relationship between the front and back of each pattern in character string collation, but cannot use information on relative feature amounts such as the spacing between the preceding and following characters.

The problem to be solved by the present invention is related to the multiple hypothesis test method, in which a character pattern is extracted from an address character line that cannot be correctly detected as a candidate for extraction by only character identification results and character string collation. The character is accurately cut out using the relative feature amount of the pattern.

[0009]

In order to solve the above-mentioned problems, the present invention introduces an evaluation value (rough penalty) based on the size of a character pattern and the positional relationship with preceding and succeeding patterns in a multiple hypothesis testing method. I do. Further, the present invention is characterized in that information of size and positional relationship can be easily handled by providing a penalty calculating means corresponding to various cutoff error hypotheses. Here, the principle of the present method will be described with reference to an example of a pattern (hereinafter, referred to as “pattern 4-6”) corresponding to an arc stretched between nodes 4 and 6 in FIG. First, in the vicinity of the pattern 4-6, the hypothesis 202
It is assumed that it should have been cut out like a rectangle shown by a broken line, and the validity of this hypothesis is evaluated based on the size of the character pattern and the positional relationship with the preceding and following patterns. Similarly, the pattern 4-6 is evaluated based on seven hypotheses in the cutout dictionary 201. In this case, the hypothesis 203 is determined to be valid for the following reason.

(1) The interval between the pattern 4-5 and the interval between the patterns 5-6 is sufficiently larger than the interval before and after (the interval between the pattern 6-7 and the pattern 3-4 with respect to the pattern 4-6).

(2) Pattern 4-5 and Pattern 5-6
Are all wide enough.

As described above, when any one of the hypotheses 201 is determined to be valid, the pattern (in this example, pattern 4-6) is cut out and deleted from the graph.
Such a determination is made by a linear discriminant function obtained by learning samples collected in advance. As a result, arcs can be reduced from the cut-out graph, and a character pattern can be cut out with high precision from a character line that was difficult to handle in the conventional method. In addition, by systematizing the cutout error in this way, it becomes easy to learn the parameters necessary for the test of the cutout candidate.

The method of realizing the present invention is described above. For this purpose, according to the first aspect of the present invention, in an address reading device to which the present invention is applied, a character in which an address is described from an image on an image is described. Perform character line extraction processing to extract lines,
If the character pattern cannot be uniquely determined, pattern extraction processing for extracting a plurality of character pattern candidates is performed, and character identification processing for character identifying the extracted pattern is performed. Subsequently, for each of the various types of conceivable character segmentation errors, the validity of the hypothesis that each read pattern is a pattern segmented in the error manner is determined for each pattern and a plurality of before and after the hypothesis. An address dictionary matching process that calculates an approximate penalty based on the relative feature amount with the pattern, then narrows down the character pattern according to the character identification result and the approximate penalty, and matches it with a dictionary of address character strings stored in advance. The address is read.

In the second aspect, in order to calculate each of the outline penalties, an identification function for identifying a correctly extracted pattern candidate and an erroneously extracted pattern candidate is used.

According to a third aspect of the present invention, in the discrimination function, at least an interval between each pattern and a pattern before and after each pattern is processed as one of the feature amounts.

According to a fourth aspect of the present invention, in the discrimination function, at least a gap in a pattern before and after each pattern is set as one of the feature quantities.

In the fifth aspect, in order to create the identification function, first, an image of a character line is used as input data, and a plurality of character pattern candidates are cut out when a method of cutting out a character pattern is not uniquely determined. Whether or not these character pattern candidates are correctly cut out, and if they are wrong, by classifying what kind of error, the cutout error of each character pattern candidate is determined,
According to this method, the pattern candidates are classified and stored, and the classification function is learned and determined using the classified and stored pattern candidates.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The configuration of an embodiment of the present invention will be described with reference to a data flow diagram of FIG. As used herein, data flow diagrams follow the Gaine-Saxon notation (J. Martin "Software Structuring Techniques", Modern Science, ISBN 4-7649-0124-2.
C3050 P5562E). This example is an address recognition process in which an entire image obtained by photoelectrically converting a surface on which an address is described is input and an address character string of an address read result is output. First, a character line extraction process 301 extracts an address character line from the entire image. Next, the character pattern candidate extraction processing 302
Character pattern candidates are extracted from the address character line, and a cut-out graph is generated. Outline penalty calculation processing 303
Calculates the approximate penalty (p) of each pattern candidate. The character identification processing 304 identifies each character pattern candidate, and outputs codes of a plurality of candidate characters and the degree of similarity between the candidate characters. The pattern certainty calculation process 305 calculates the certainty (pattern certainty) of each pattern candidate based on the similarity of the character identification result and the approximate penalty. The address dictionary matching process 306 selects and discriminates pattern candidates based on the pattern certainty factor, and matches candidate characters of the character identification result with the address character string dictionary.

FIG. 2 shows the procedure of the processing of this embodiment. First,
In step 401, a character line is extracted from the entire image. Next, in step 402, character pattern candidates are extracted from the character line, and a cut-out graph is generated. Next, for all character pattern candidates in the cut-out graph, steps 405 to 407 are performed in the control loop 403.
repeat. In the control loop 403, first, in step 405, the approximate penalty of each character pattern candidate is calculated. Next, in step 406, each character pattern candidate is identified. Next, in step 407, the certainty factor of the pattern corresponding to each arc on the cut-out graph is calculated. After the control loop 403 ends, in step 404, the character identification result is compared with the address dictionary.

Next, the input and output of the character line extraction processing 301 will be described with reference to FIG. Character line extraction processing 301
Is a process for extracting an address character line from the postal matter image 501 in the entire image. The address character line is a rectangular area 502 that includes from the name of the prefecture and city of the address to the street and address.
That is. When the address extends over two or more lines, rectangular areas 502 corresponding to the number of lines are output. As a method for extracting address character lines in this way, Toshiba Review 1993 Vol. 48 N
o. Use the method described in Chapter 3 Section 3 of "Image Processing Technology for Postal Machines".

Next, the principle of the character pattern candidate extraction processing 302 will be described with reference to FIG. First, character boundary candidates are extracted from the address character line. The vertical bars to which numbers (boundary numbers) from 0 to 9 are assigned in the drawing are candidates for the boundary. A candidate for the boundary is a gap between integrated rectangles as described in IEICE Technical Report IE88-138, "Character extraction method in a printed document containing a character string with an indefinite pitch". Here, a value obtained by subtracting the x coordinate of the right end of the pattern on the left side of the pattern from the x coordinate of the left end of the pattern on the right side of the boundary is the boundary gap, and the right end of the pattern on the left side of the boundary from the x coordinate of the left end of the right pattern of the boundary. Is defined as boundary coordinates.
For example, the boundary coordinate corresponding to the boundary number 4 is the x coordinate of 604, and the boundary gap is the width represented by 605. Next, a combination of boundaries is checked so that the difference between the boundary coordinates does not exceed the character size estimated from the height of the character line, and a pattern sandwiched between the boundaries is registered as a character pattern candidate.
In the example of FIG. 4, the difference 601 between the boundary coordinates does not exceed the estimated character size, and the difference 603 between the boundary coordinates does.
Therefore, the pattern 0-1 and the pattern 0-2 are registered, but the pattern 0-3 is not registered.

FIG. 5 shows a character pattern candidate extraction process 302.
Shows the format of a pattern table that represents the arcs of the cut-out graph generated by. Each record of the pattern table corresponds to one pattern candidate. Each record of the pattern table has an area 701 for storing the outline of the pattern, and the outline of the pattern is described by a chain code or the like. In addition, there is an area 703 for storing the number 702 of the left boundary of the pattern and the number of the right boundary, and the number of the boundary is 0 in the left end of the character line, as shown in FIG. Decide to increase in increments of one. Further, there is an area 704 for storing candidate characters as a result of character identification. In the present embodiment, up to three candidate characters are stored in descending order of similarity. There is an area 705 for storing the similarity of the candidate characters stored in the area 704. The candidate characters and the similarity are left-justified. If the number of candidate characters is 3 or less, a null code is set in the blank of the candidate character area. Zeros are filled in the margins of the areas of similarity. Note that there is an area 706 for storing the pattern certainty factor.

FIG. 6 shows a character pattern candidate extraction process 302.
Shows the format of the boundary table that represents the nodes of the cut-out graph generated by. Each record in this table represents one boundary. Each record has a boundary number 801 and boundary coordinates 80
2. This is an area for storing the boundary gap 803.

The character identification processing 304 is realized by using a known method as described in, for example, Mori “Pattern Recognition” (ISBN4-88552-075-4C3055, IEICE) pp. 32-109. Categories for character identification include kanji, hiragana, katakana, Arabic numerals and symbols.
It shall be used for the place name or the street / block. As the output of character identification, a plurality of candidate characters and the similarity of the input pattern to the standard pattern of each candidate character are obtained.

Next, the configuration of the outline penalty calculation processing 303 will be described with reference to the data flow diagram of FIG. A character pattern candidate which is an input of this process is represented by one record in the pattern table and a boundary table. Here, as processing for evaluating the assumption that each pattern candidate is erroneously cut out, each of the pattern error candidates is evaluated by a cut-out error hypothesis evaluation processing 901, 902,
903 and so on. In the present embodiment, seven types of extraction error hypothesis evaluation processes corresponding to the types of extraction errors E1 to E7 shown in FIG. 8 are used. The output pi of the extraction error hypothesis evaluation process indicates that the larger the output pi, the more likely the extraction error hypothesis is. The outputs pi of these cut-out error hypothesis evaluations are added in the process 904 and output as the approximate penalty p.

FIG. 8 is a diagram for explaining the types of hypotheses of cutout errors. In the figure, a thick black line represents a pattern candidate of interest, a dashed rectangle represents a circumscribed rectangle of a correct character pattern, and a white line represents a part of a pattern around the pattern candidate of interest. For example, E1 represents a hypothesis that the pattern candidate of interest is actually a part of the left side of a character cut out by mistake. E7 indicates a hypothesis that the pattern candidate of interest is actually one pattern in which two characters are erroneously set.

FIG. 9 is a data flow diagram showing the configuration of the cut-out error hypothesis evaluation process. A pattern candidate which is an input of this processing is represented by one record in the pattern table and a boundary table. Here, the hypothesis evaluation processing 1101 is the hypothesis evaluation processing 901 and 9 shown in FIG.
02, 903, etc., corresponding to any of the cut-out error hypothesis evaluation processes. The feature amount extraction processing 1102 extracts feature amounts such as the size of the character pattern and the positional relationship with the preceding and following patterns from the input character pattern candidate. Features are
Let it be an n-dimensional vector represented by the following equation.

F = (f1, f2,..., Fn) Next, in processing 1103, pi is obtained from the feature value F. pi is a value of a linear discriminant function for identifying a correctly extracted candidate pattern and an erroneous candidate pattern such as Ei, and can be defined by the following expression.

Pi = F · Vi + ci (F · V
i: inner product of Vi and F) Here, Vi: weight vector of the linear discriminant function ci: constants Vi and ci are learned in advance by a method described later and stored in the parameter dictionary 1104. As another example different from this example, the parameter dictionary may be switched according to the candidate character of the character identification result.

FIG. 10 illustrates the feature values used in this example. In the figure, a pattern 1201 represented by a thick black line indicates a character pattern candidate of interest, and patterns 1202 and 1203 represented by white indicate peripheral patterns. The dashed rectangle indicates a rectangle circumscribing each pattern.
In this example, the number of dimensions of the feature amount is n = 6, and each dimension fj is defined as follows.

F1: height of the pattern of interest f2: width of the pattern of interest f3: interval between the pattern of interest and the pattern on the left f4: interval between the pattern of interest and the pattern on the right f5: Maximum gap in the pattern of interest f6: Number of connected components in the pattern of interest In this example, the same feature is used in all the cut-out error hypothesis evaluation processes, but different features are used in each process. May be used. Further, each feature amount may be normalized by the overall feature amount of the character line such as the character line height h.

FIG. 11 shows the principle of the method for evaluating a cut-out error hypothesis. Coordinate axes 1301 and 1302 virtually show a feature space that is actually n-dimensional. Pattern group 13
03 indicates the distribution of the feature amount F of the correctly extracted character pattern candidate, and the pattern group 1304 indicates the distribution of the feature amount F of the character pattern candidate for which the extraction error hypothesis Ei is true. Wi in the figure is a pattern group 1303
And 1304, which are identification function weight vectors for identifying the groups 1303 and 1304, which are orthogonal to Wi. Here, assuming that a is the Euclidean distance from the origin to the hyperplane B, the hyperplane B is expressed as a set of F satisfying the following equation.

(Wi · F) = a · | Wi | (Wi · F: inner product of Wi and F, | Wi |: norm of Wi) Further, the value d of the linear discriminant function that identifies the groups 1303 and 1304 is If d> 0, F is identified as group 1304, and if d <0, F is identified as group 1303.

D = (Wi · F) −a · | Wi | Wi and a · | Wi | are Toriwaki “recognition engineering” (ISBN4
-339-01059-6 C3355 P2781E, Corona) pp. 113
-It can be stopped by a method like 119. However, it is not appropriate to use the value of d as pi as it is because the distribution of each Ei in the feature space is different. Therefore, the value of the linear discriminant function normalized as in the following equation is used as pi.

Pi = d / (s · | Wi |) = (Wi · F) / (s · | Wi |) −a / s (s: variance of d with respect to the set of 1303 and 1304) Vi stored in the dictionary 1104,
ci is determined as follows.

Vi = Wi / (s · | Wi |) ci = a / s Next, the outline of the pattern certainty factor calculation processing 305 will be described. The pattern certainty factor is a measure indicating how probable each arc on the cut-out graph, that is, each pattern candidate, is obtained by the following equation.

(Pattern certainty factor) = {c1. (First-order candidate character similarity) -c2.p} In the equation, p is a rough penalty. Further, c1 and c2 are constants, and are appropriately adjusted for each system.

Next, the configuration of the address dictionary matching process 306 will be described.
This will be described with reference to the data flow diagram of FIG. The pattern candidate, the pattern certainty degree candidate character, and the similarity, which are the inputs of this processing, are passed by the pattern table and the boundary table. First, pattern candidates whose pattern certainty factor is equal to or less than a certain value are selected in a pattern candidate selection process 1401. In the example of FIG. 20A, patterns 0-2, 0
-3, etc., because the similarity of the character identification result is small, the pattern certainty is also reduced and deleted, as shown in FIG. In addition, since the pattern 4-6 has a large outline penalty, the pattern certainty becomes small and is deleted. Next, in a dictionary matching process 1403, candidate characters of a character identification result corresponding to each pattern candidate are compared with an address character string stored in an address character string dictionary 1404 in advance.
Outputs the candidate (candidate character string) of the address character string that has been successfully verified. The candidate address character string sort 1405 sorts the candidate character strings in descending order of the degree of verification certainty and outputs the result as an address read result. The collation certainty is the degree of matching when the candidate character of the character identification result is collated with the candidate character string, and the larger the value of the collation certainty, the more likely the candidate character string is a probable candidate.

FIG. 13 shows an outline of the dictionary collation processing 1403. In this processing, an address character string received by the automaton generated based on the character identification result is selected from the address character string dictionary 1404. The method of obtaining the address character string accepted by the automaton conforms to the method of Marukawa et al. (Information Processing Society of Japan, Vol. 35, No. 6, "Error Correction Algorithm for Handwritten Kanji Address Recognition"). Here, the center frame in FIG. 13 schematically shows the automaton 1501 generated from the candidate characters of the character identification result after the complementary pattern selection. The boundary between the patterns is a state, and the candidate characters of the character identification result are transitions. The number of each state is the same as the number of the section of the cut-out graph. The automaton is realized by a table having the same format as the pattern table. The bold line in the automaton 1501 indicates what transition has been made by the automaton 1501 through the character string 1503 (“Mizutani-cho 12, Nakazumi City”) in the address character string dictionary 1404. As described above, when a character string in the address character string dictionary 1404 is received, the character string is output as one of the candidate character strings. The verification certainty mc is a sum of certainty tc (transition certainty) corresponding to each transition at the time of verification.

Mc = Σtc The transition confidence is obtained by the following equation.

Tc = {c1.sm-c2.p} .jm where sm: similarity of candidate characters corresponding to each transition jm: difference between state numbers before and after the transition The constants c1 and c2 in the formula are pattern certainty. Use the same one used to determine the degree. In the example of FIG. 13, the character string “11, Sumimachi, Nakazumi-shi” is also accepted, and the address recognition result 15
02 is output. However, in the case of the latter, since the candidate character is received using a candidate character having a similarity smaller than the former, the verification certainty is reduced.

FIG. 14 shows a hardware configuration in this embodiment. The example of the embodiment of the present invention described above is applied to a recognition device 1604 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. The image is input by the scanner 1601. Here, a delay line 1602 is provided on the mail conveyance path to secure the address reading time. The mail is sorted by the sorter 1603 based on the address reading result. The scanner 1601 and the recognition device 1604 are connected by an input / output cable 1612.
The sorter 1603 and the recognition device 1604 are connected by an input / output cable 1613. A recognition device 1604 includes a bus 1611 for connecting each part inside the recognition device, and an input / output interface 16 for controlling communication with the scanner 1601.
05, an arithmetic processing unit 1606 for controlling the overall recognition device and address recognition processing, an input / output interface 1607 for communicating with the sorter 1603, a keyboard 1608 for performing operations such as activation, and an execution status are displayed as necessary. CRT 1609, table required for address recognition,
Memory 1610 for storing programs, dictionaries, etc.
Is provided.

FIG. 15 is an example of a display screen of a sample collection tool for collecting a sample for learning the parameter dictionary 1104 used for the extraction error hypothesis evaluation 1101 in FIG. In FIG. 15, reference numeral 1701 denotes a CRT for displaying a screen, and 1702, a window for displaying a character line image. The entire character line image is displayed in the window, and the pattern currently focused on is displayed in a different color (portion drawn by a thick black line in the figure). The operator observes the image in the window 1702, determines whether the image is a correctly cut out pattern, is incorrect, or determines which error in FIG. 8 is wrong.
Click with 4. The buttons are arranged on the panel 1703. Upon receiving the button click event, the sample collection tool stores the feature amount of the pattern of interest in a file corresponding to the type of error, and displays a new pattern in the window 1702.

FIG. 16 shows the parameter dictionary 1 in FIG.
4 is a data flow diagram of a system for learning 104. Character line image D collected in advance
Using the B (database) 1801, the sample collection tool 1802 extracts the correct pattern DB (database) 1803 and the various pattern databases E1 to E in the hypothesis of the cut error in FIG.
7 corresponding to the extraction error pattern DB (1804,
1805). The learning tool 1806 receives as input the pattern DB 1803 of the cut-out correct answer and the database 1804 of the cut-out error pattern corresponding to E1, and obtains V1 and c1 by the method described with reference to FIG.
Output to the parameter dictionary 1808. Hereinafter, E2 to E
Pattern database (1)
805 etc.) in the same manner.
, Etc.) to obtain Vi and ci, and obtain the parameter dictionary 180
8 is output.

FIG. 17 shows the format of a table for storing parameter dictionaries. Each record in the table pdic [i]
, The parameters Vi, ci corresponding to each Ei are stored. For example, the head 1903 of the table corresponds to pdic [1] and stores V1 and c1. The i-th record 1904 from the top corresponds to pdic [i] and stores Vi and ci. In each record, ci is stored in area 1901, and Vi is stored in area 1902.

FIG. 18 shows a processing procedure of the rough penalty calculation processing. First, in step 2001, a variable p is initialized to 0. Next, in a control loop 2002, steps 2003 and 2004 are repeated while incrementing the variable i. Step 2003 is a step of activating the segmentation error hypothesis evaluation. Step 2004 is a step of adding pi obtained as a result of the segmentation error hypothesis evaluation to p. Step 2008 outputs the variable p as the approximate penalty. Steps 2005 to 2006
Shows a processing procedure of a subroutine for performing cutout error hypothesis evaluation. Step 2005 is a step of substituting the value of ci (pdic [i] .c) read from the parameter dictionary into the variable pi. Step 2006 is a control loop for calculating the inner product of F obtained by the feature extraction and Vi read from the parameter dictionary, and while increasing the variable j by the number of dimensions of the feature amount, the value of each dimension of Vi (pd
ic [i] .v [j]) and the value of each dimension of F (f [j]) is added to pi.

[0047]

According to the method of the present invention described above, the size of the character pattern and the positional relationship between the preceding and succeeding patterns can be used even from an address character line in which it is not possible to correctly test a candidate for extraction by only character identification results and character string collation. Then, characters can be accurately cut out.

[Brief description of the drawings]

FIG. 1 is a configuration diagram illustrating an example of an embodiment of the present invention.

FIG. 2 is a diagram showing a processing procedure in an example of an embodiment of the present invention.

FIG. 3 is a character arrangement diagram showing a relationship between a whole mail image and character lines.

FIG. 4 is a diagram showing a relationship between a pattern and a boundary.

FIG. 5 is a table showing a format of a pattern table for storing arcs of a cut-out graph;

FIG. 6 is a table showing a format of a node table in which nodes of a cut-out graph are stored.

FIG. 7 shows data indicating a procedure of a rough penalty calculation process.
Flow diagram.

FIG. 8 is a data flow diagram showing the types of cutoff error hypotheses.

FIG. 9 is a diagram illustrating a procedure of a cutout error hypothesis evaluation process.

FIG. 10 is a diagram showing feature amounts used in a cutout error hypothesis evaluation process.

FIG. 11 is a diagram showing the principle of a cutout error hypothesis evaluation process.

FIG. 12 is a data flow diagram illustrating a procedure of an address dictionary matching process.

FIG. 13 is a view showing the principle of dictionary matching processing.

FIG. 14 is a hardware configuration diagram showing an example of an embodiment of the present invention.

FIG. 15 is a view showing an example of a display screen of a sample collection tool.

FIG. 16 is a system configuration diagram for learning a parameter dictionary.

FIG. 17 is a table showing a storage format of a parameter dictionary.

FIG. 18 is a flowchart showing the procedure of a rough penalty calculation process.

FIG. 19 is a diagram showing an example of an address character string to be read.

FIG. 20 is a diagram showing an example of a cutout graph and a cutout error hypothesis.

[Explanation of symbols]

301: character line extraction processing, 302: character pattern candidate extraction processing, 303: outline penalty calculation processing, 30
4 ... Character identification processing, 305 ... Calculation of pattern certainty factor,
306: Address dictionary collation processing, 501: Mail image 502: Address / address writing area 601: Boundary coordinate difference 602: Boundary coordinate difference 603: Boundary coordinate difference 604: Boundary coordinate 605 ... Boundary gap 701 ... Pattern contour storage area 702 ... Pattern left boundary number field 703 ... Pattern right boundary number field 704 ... Candidate character storage area 705 ... Candidate character similarity storage area 706 ... Pattern certainty storage area 801 Boundary number field 802 Boundary coordinate field 803 Boundary gap field 1101 Extraction error hypothesis evaluation processing 1201 Attention character pattern candidate 1202 Peripheral pattern 1203 Peripheral pattern 1301 Coordinate axis 1302 Coordinate axes 1303... Correctly cut out character pattern candidate group 1304. A character pattern candidate group for which the error hypothesis is true 1404... An address character string dictionary 1502... An address recognition result 1604... A recognition device 1611... A bus 1612... An I / O cable 1613. Screen 1702: Character image display window 1703: Panel

Continuing from the front page (72) Inventor Hiromichi Fujisawa 1-280 Higashi-Koikekubo, Kokubunji-shi, Tokyo Inside the Hitachi, Ltd. Central Research Laboratory (72) Inventor Hisao Ogata 1-280 Higashi-Koikekubo, Kokubunji-shi, Tokyo Hitachi, Ltd. Central Inside the research institute (72) Yoshihiro Shima 1-280 Higashi Koikebo, Kokubunji-shi, Tokyo Central Research Laboratory, Hitachi, Ltd. (72) Inventor Masato Teramoto 1 Ikegami, Haruoka-cho, Owariasahi-shi, Aichi Pref. Hitachi, Ltd. Office Systems Division (56) reference Patent flat 4-299485 (JP, a) JP Akira 61-175878 (JP, a) JP flat 7-6203 (JP, a) (58 ) investigated the field (Int.Cl. ⁷ G06K 9/62 G06K 9/34

Claims (4)

(57) [Claims] An image signal described on a document is converted into an electric signal.
Image input means for converting and inputting
Means for reading the address, Possible character segmentation
Record the weight vector of the discriminant function for each type of error.
Recording meansCharacter that extracts a character line on which an address is described from an image in an address reading device having
If the line extraction process and how to extract the character pattern cannot be uniquely determined
Pattern extraction to extract multiple character pattern candidates
Processing, character identification processing for character identifying the extracted pattern,Regarding the cut out pattern, the respective patterns and
And relative features of multiple patterns before and after
Any of the above types of error using the function weight vector
Is calculated, and from the identification function
The process of calculating the approximate penalty indicating the validity of the extraction method
And  Character pattern according to character recognition result and outline penalty
To narrow down the address candidates, and
And sequentially execute address dictionary matching processing for matching
Characteristic address reading method. 2. The address reading method according to claim 1, wherein at least an interval between patterns before and after each pattern is one of the feature amounts. 3. The address reading method according to claim 1, wherein at least a gap in each pattern is one of the feature quantities. 4. A method for creating a discriminant function weight vector for each outline penalty calculation process according to claim 1, wherein an image of a character line is input and a plurality of characters are extracted when a method of extracting a character pattern cannot be uniquely determined. Pattern cutout processing for cutting out pattern candidates, and whether character pattern candidates are correctly cut out,
Error How to cut out that if you are wrong stores pre-Me
A cut-out error type discriminating process for classifying what kind of error among the types, a pattern storing means for classifying and storing each character pattern candidate according to a result of the cut-out error discriminating process, characterized by sequentially executing the process of learning the identification function weight vectors for each said pattern using a character pattern candidates are identified function
Number weight vector generation method.