JP3740271B2 - Text line entry area reading method and mail address recognition method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention extracts a character line when recognizing a destination in order to classify a mail item or a delivery product, selects a portion where the address is written from the extracted character line, performs character line recognition, The present invention relates to a method for recognizing a mail or the like as a result of address recognition.
[0002]
[Prior art]
In a mail automatic reading device, there is Japanese Patent Application Laid-Open No. 5-151389 as a method for detecting an address description area on a mail, and this is a technique for detecting a postal code position and detecting the address description area from this position. is there. As a method of cutting out a character line from an address description area, there is Japanese Patent Application Publication No. Sho 60-41396, which detects the height of a block pattern of a written character or the like and detects a pattern of the same height as a line. It is. Japanese Patent Application Publication No. Sho 63-18785 discloses a method for discriminating vertical and horizontal writing of cut out character lines, which calculates the horizontal and vertical peripheral lengths of character groups in the address area, and compares these peripheral lengths. Thus, the line direction of the character group is determined. In the address line character recognition method, various means for reading a place name such as a prefecture name, a city group name, and a town name of the address have been devised. For example, in Japanese Patent Application Publication No. 2-64882, the address from the head of the address to the town name is devised. The address line is recognized by a method of performing different character extraction processing between the character string and the character string at the chome address portion.
[0003]
[Problems to be solved by the invention]
In a mailing address, the character size, character spacing, and description means (printing, brush, ballpoint pen, etc.) described in advance are unknown. Therefore, the characters described as the address are described in a wide variety of forms, and it is impossible to accurately extract all the address lines described in various characters with one known setting in advance. . In order to set parameters when extracting an address line, it is not easy to determine the character form described as the address line from the image data.
[0004]
In addition, the format of the address of mails, etc. includes the vertical writing with the long side in the vertical direction and the address in the vertical direction, the horizontal writing with the long side in the horizontal direction and the horizontal writing in the horizontal direction, and the long side There are mainly many vertical and horizontal writing formats in which the address is written in the vertical direction. Among these types of formats, when an image is input with the long side of the mail piece placed vertically, in the direction of the address line, the vertically placed vertical book and the horizontally placed horizontal book are determined to be the same longitudinal direction. This is because the direction of each character is unknown. In many cases, the address of a vertically placed vertical book is written in large characters, and the address of a horizontal and horizontal book is often written in small characters. It is difficult to perform an address line extraction process for these two formats with the same parameter setting and to obtain a correct extraction result for both.
[0005]
[Means for Solving the Problems]
The present invention is an effective technique for enabling accurate address line extraction and improving address recognition accuracy even if the size, character spacing, and description method of the letters described as mail address are unknown. .
[0006]
FIG. 1 explains the means for solving the problem to be solved by the present invention. Figures 101 and 105 show the images of the mailing surface, which is the digitized long side of the mail. The format of the 101 address is vertical and the format of 105 is horizontal. As a result of performing character line extraction with a parameter setting that extracts a character line with a large character for the address surface with 102 as 101, the correct character line can be extracted because 101 is described with a character with a large address . If address area extraction is performed using the character line extraction result of 102, a partial image such as 103 is obtained. If address recognition is performed using this, a correct address recognition result is obtained. However, if character line extraction is performed with the parameter setting for extracting a character line of a small character like 104, nothing is extracted because there is no character line composed of small characters.
[0007]
Further, 106 is a result of extracting character lines with the parameter setting for large characters on the address surface of 105, and character lines described in small characters are extracted as one line by combining a plurality of lines. If the address area extraction is performed based on the character line extraction result, a partial image 107 is obtained. When the character direction is determined for this partial image and the image is rotated, 108 partial images are obtained. However, a correct address recognition result cannot be obtained using the 108 partial images. 109 is the result of character line extraction for 105 address planes with small character parameter settings. Since the address line was written in small characters, the character line was correctly extracted. If address area extraction is performed based on the 109 character line extraction results, 112 address areas are generated. The character direction is determined for this address area candidate, and the result of rotation is a partial image 113. When address recognition processing is performed on the 113 partial images, a correct address recognition result is obtained.
[0008]
As described above, when character line extraction is performed using a plurality of setting values for the address surface, there is a high possibility that a correct result is included in the plurality of character line extraction results. Therefore, all character line extraction results are stored, and a plurality of address area description area candidates are generated from the character line extraction results. A recognition process is performed on the plurality of description area candidates, and a mail address recognition result is determined based on the recognition result.
[0009]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings.
[0010]
FIG. 2 shows a hardware configuration for realizing the present invention. An object of the embodiment of the present invention is automatic sorting by mail address recognition. Therefore, as shown by 201, there is a scanner for digitizing the mail address as an input image and connected to the bus. As shown by 205, there is a display system that displays an image of the address plane being processed, an address recognition result, a processing mode, and an error code. There is an operation system at 206, where an operator operates the machine. A mail address dictionary is installed in 202, which stores standard patterns of characters for address recognition, resident address data of delivery areas, and hierarchical structure information of prefectures and municipalities. A memory 203 is provided, and a CPU 204 is provided. The address recognition program is operated using the memory and the CPU. The address recognition program operating using this memory and CPU is input to the sorting unit 207 indicating the address recognition result, and guides the mail to the bin corresponding to the recognition result. Further, as indicated by 208, a network connection unit is provided, which is connected to a network line, so that it is possible to exchange an address dictionary, an address recognition program, image data of the address surface, etc. from a remote location.
[0011]
FIG. 3 shows the processing configuration of the address recognition program. As shown in 301, the address line generation module first processes the address line generation module in 302 using binary image data scanned with the long side vertical and the upper left as the stamp attachment position as shown in 301. Generate region candidates.
[0012]
This address line generation module includes a character line extraction unit indicated by 303, an address area candidate generation part indicated by 304, and an address area candidate selection part indicated by 305. In the character line extraction unit, a combination for each connected component in the vicinity of the input image on the address plane is verified, the distance between the connected components, the increase in area when integrated, and the linearity between the connected components are verified, and the integration process To extract a character line. Thereafter, the address area candidate generation unit verifies the size of the character line and the increase in area when the character line is combined, and generates an address area candidate when combining a plurality of character line extraction results. Next, the address area candidate selection unit verifies the positional information and size of the generated address area candidates, the character line size in the address area candidates, and variations in the character lines, and assigns a score of the address area to each address area candidate. Sort by score. The information is input to the address line recognition module 306 in the rearranged order, and address line recognition processing is performed. For this address line recognition processing, for example, processing known in Japanese Patent Application Publication No. 2-6882 is used.
[0013]
Next, the address recognition result determination unit indicated by 307 evaluates the address area candidate to which the address recognition result is added. The address recognition result determination unit repeats the address line recognition process until an acceptance result is obtained with reference to the recognition results added to the plurality of input address area candidates. When the acceptance recognition result is obtained, the address line recognition process is stopped, and the recognition result that is the acceptance is output as the recognition result of the mail address.
[0014]
FIG. 4 illustrates a detailed configuration in the address line generation module shown in FIG. The 402 character line extraction unit has upper case character line extraction with 403 parameters set for upper case extraction, and upper case character line test set for 404 upper case extraction, and 405 has parameters for lower case extraction. There is a set of character line extraction for lowercase letters, and 406 has a character line test for lowercase letters set for lowercase letters. In the character line extraction unit 402, uppercase character line extraction processing is first performed, and uppercase character line verification processing is performed on the result. Only the character line extraction results that have passed this verification process are registered in the table as uppercase character line extraction results. Next, lowercase character line extraction processing is performed in the same manner, and lowercase character line verification processing is performed on the result. The character line extraction result that has passed this verification process is registered in the table as a lowercase character line extraction result.
[0015]
These upper-case character line extraction processing and lower-case character line extraction processing differ only in the parameter settings within the character line extraction processing. This parameter setting is a setting of a distance between connected components that are referred to when the connected components are integrated, and a threshold value for an increase in area when the connected components are integrated. Also, the character setting for uppercase characters and the character row verification for lowercase characters differ only in the parameter settings inside the character line verification process. This parameter setting refers to the size of the rectangle, the connected components in the rectangle, and the average size of the connected components in the rectangle that are referenced when performing the verification process on the character line rectangle candidates obtained in the character line extraction process. , Setting of a threshold of variation of connected components in a rectangle.
[0016]
Next, in 407, address line area candidate generation is performed. Here, by referring to the character line extraction result registered in the table by the character line extraction unit, the address area candidate that is appropriate as the area in which the address line is described on the basis of the positional relationship and alignment of the character lines Is created and registered in the table.
[0017]
Next, address area candidate selection is performed at 408. Here, the table is referred to, and the address area candidate is addressed using the generated address area candidate size, position, and information on the text line rectangle contained therein. Area-likeness scoring is performed, sorting is performed in the order of the scores, and addressable area candidates to be input to the addressed line recognition unit are arranged and output in the order that they are likely to be addressed.
[0018]
FIG. 5 shows a processing flow of character line extraction shown in FIG. In this flow, by changing the parameter setting, it is possible to change the character size to be subjected to the character line extraction process or change the line direction. In this character line extraction, connected components are integrated in two stages, and a character line is extracted. A rectangle generated by the first stage integration is called a primary merge rectangle, and a rectangle generated by the second stage integration is called a row candidate rectangle.
[0019]
First, in the integration of the first stage, the connected component that becomes the core is selected at 501 and set to the upper end, lower end, right end, and left end coordinates of the primary merge rectangle. Next, at 502, the coordinates of the primary merge rectangle are compared with other connected components, and the nearest connected component is selected. The distance between the connected component selected in 503 and the primary merge rectangle is determined. If the distance is less than a certain value, the increase in area is determined in 504 when the connected component and the primary merge rectangle are merged. In this case, the connected component is registered in the primary merge rectangle in 505, and the coordinates of the primary merge rectangle are also updated. Next, in 506, it is determined whether or not the primary merge rectangle has been updated. If the primary merge rectangle has been updated, it is compared again with other connected components and the process is repeated. If the primary merge rectangle has not been updated, there is no connectable component that can be integrated, and the processing related to the primary merge rectangle ends. Next, in 507, the existence of connected components that are not registered in all the primary merge rectangles is checked. If there is a connected component, the connected component to be used as a nucleus is set as an unregistered connected component in 508, and the neighboring connected components are set. And repeat the process. Here, the threshold value of the determination processing of 503 and 504 is a parameter setting value. By changing these parameters, the vertical and horizontal extraction directions can be switched, and the character size and character spacing in the character line to be extracted can be changed.
[0020]
Next, a processing flow for generating a row candidate rectangle based on the primary merge rectangle generated by the above processing will be described. In 509, the primary merge rectangle to be the core is set to the coordinates of the top, bottom, right, and left edges of the line candidate rectangle, and in 510, it is checked whether it is aligned with other primary merge rectangles vertically or horizontally and aligned. In step 511, the distance between the line candidate rectangle and the primary merge rectangle is checked. If these two checks are satisfied, the row candidate rectangle is updated at 512. In step 513, the update of the line candidate rectangle is checked. If the rectangle is updated, another primary merge rectangle is compared, and the process is repeated in the same manner. If it has not been updated, there is no primary merge rectangle to be integrated with this line candidate rectangle, so this line candidate rectangle is registered at 514. Next, in 515, the primary merge rectangle that is not reflected in the row candidates is checked, and if there is any that is not reflected, it is set as the primary merge rectangle, and the integration process is repeated in the same manner. Here, the threshold values of the judgment portions 510 and 511 are parameter settings. By changing these parameters, the vertical and horizontal extraction directions can be switched, and the character size and character spacing in the character line to be extracted can be changed.
[0021]
FIG. 6 shows a processing flow of a character line test for testing whether a line candidate rectangle after character line extraction is appropriate as a character line. In each row candidate rectangle, the width is checked at 601 and the height is checked at 602. If the size is appropriate, a connected component included in the row candidate rectangle is searched at 603. In the connected components searched here, the number of connected components is checked at 604, the average size of the connected components is checked at 605, and the presence of an extremely large connected component is checked at 606. The line candidate rectangles that have passed these checks are registered as character lines in the table at 607. In step 608, it is checked whether the line candidate rectangle is the last one, and the same check is repeated until the last line candidate rectangle. In this character line verification process, the threshold value in the determination process at 601, 602, 604, 605, and 606 is set as a parameter. By changing these parameter settings, it is possible to change the size of the character line to be verified, the character size in the character line, and the variation in character size.
[0022]
FIG. 7 shows a processing flow of address line area candidate generation. First, in 701, a character line as a core is determined and set as a candidate address area rectangle. In 702, the address area candidate rectangle and the adjacent character line are searched, and in the adjacent character line, the text line is checked for a different size in 703, and the address area candidate when adding the character line in 704 is checked. The area increase is checked, and if both checks are valid, a character line is added in 705 to update the address line area rectangle. In 706, it is checked whether or not the address area rectangle is updated. If there is an update, a nearby character line is searched and the same processing is repeated. If no update has occurred, register the address area candidate in the table in 709, check if there is a character line not registered in the address area candidate in 707, and if it exists, set the character line to be the core in 708 Correct and repeat the process.
[0023]
FIG. 8 shows a processing flow of address area candidate selection. Referring to the table, check the size of each addressable area candidate with 801, add a score with 802 if appropriate, and check with 803 whether the address area location is included in the sender's description. If it is, the score is deducted at 804. Next, search the address line in the address area at 805, check the number of address lines at 806, add a score at 807 if appropriate, add a score at 809 if check the arrangement of the address lines at 808 810, check if the address line size is complete, and if it is complete, score 811. In 812, it is checked whether the address area candidate is the last, and if it is not the last, the same scoring is repeated, and scoring is performed for all the address area assistants. When scoring is completed for all addressable area candidates, the addressable area candidates are sorted in the order of points in 813 and input to the address line recognition module. In the address line recognition module, recognition processing is performed in the sorted order, and when an appropriate recognition result is obtained, the address recognition result of the mail piece being processed is determined.
[0024]
【The invention's effect】
As described above, according to the present invention, the accuracy of extracting mail address lines is improved, and as a result, the mail address recognition accuracy is improved. This invention is effective as a process for address forms of various mail items.
[Brief description of the drawings]
FIG. 1 is a diagram of a solution to an address line extraction problem.
FIG. 2 is a diagram of a hardware configuration in an embodiment of the present invention.
FIG. 3 is a diagram showing a module configuration for address recognition processing;
FIG. 4 is a diagram showing a module configuration in address line extraction processing;
FIG. 5 is a flowchart of address line extraction processing;
FIG. 6 is a flowchart of address line verification processing;
FIG. 7 is a flowchart of address area candidate generation processing;
FIG. 8 is a flowchart of address area candidate selection processing;
[Explanation of symbols]
301: Address image 302 ... Address line generation module 303 ... Character line extraction unit 304 ... Address area candidate generation unit 305 ... Address area candidate selection unit 306 ... Address line recognition module 307 ..Address recognition result determination unit 308 ... Address recognition result

Claims (5)

The image information on the character description surface is converted into an electrical signal and input, and the input image is repeatedly extracted multiple times with multiple parameter settings for at least one of the character size and the character spacing in the character line. A character line description area reading method comprising: extracting a plurality of character lines and extracting a character line description area by combining the plurality of character lines extracted with different parameter settings . The image information of the address description surface such as postal items is converted into an electric signal and input, and the input image is set multiple times by setting a plurality of parameters for at least one of the character size and the character interval in the character line. Repeat extraction to extract a plurality of character lines, a combination of a plurality of character lines extracted the with different parameter settings to extract the address described region, the character identification character line of said extracted address described region A method for recognizing mail addresses, etc. During the address area extraction candidate region address lines from the image data of one address plane is described to output a plurality extracted, the position of the region with respect to a plurality of area candidates issued extract, size Using the whole or a part of the size of the character line included in the area or the relative position information of the character line, the validity as the area where the address is described is evaluated, and the evaluation Area candidates are arranged in order of validity evaluation values, character lines in the area candidates of the arranged address lines are character-identified, and a mail address recognition result from areas candidates in which a plurality of address lines are described 3. A method for recognizing a mail or the like according to claim 2, wherein the mail address is determined.   A first step of scoring the likelihood of the address line description area based on the position information on the address surface of the area candidate in which the address line is described in the validity evaluation, and the area where the address line is described The second step of counting the number of character lines in the candidate and scoring the likelihood of the address line description area based on the number, and the coordinates of the character line rectangles in the candidate area of the address line A third step of scoring the likelihood of the address line description area according to the value deviation, and the address line description according to the size difference between the character line rectangles in the candidate area where the address line is described 4. A postal item according to claim 3, wherein the area candidates are rearranged in the order of scores, and the candidates for the areas where the addresses scored in the first to fourth steps are listed. Address recognition method.   In the extraction of the above character line, the parameter setting for the character line described in large characters with respect to the extraction result obtained by setting the character line extraction parameter setting as the extraction target for the character line extraction parameter setting. The character line extraction result that passed the character line verification process is stored in the memory, and the parameter setting of the character line extraction is set as the extraction target character line described in small characters Perform the character line verification process that is set for the character line described in small characters for the obtained extraction result, and store the character line extraction result that has passed the character line verification process in the memory, 3. A method for recognizing a mail address or the like according to claim 2, wherein a candidate for an area in which the address line is described is generated with reference to a character line extraction result stored in the memory.

Images