JP5003051B2 - Automatic mail sorting machine and automatic mail sorting method - Google Patents

Description

  The present invention collects images of postal items to be classified, reads addresses from the collected postal images, and automatically classifies postal items based on classification specifying information derived from the read addresses In particular, the present invention relates to an automatic mail sorting machine and a mail automatic sorting method for reading addresses in parallel from one mail image using a plurality of address information reading units having different recognition algorithms.

  Read the address (postal code, prefecture, city, street name, street address, company name, destination name, etc.) from the postal items to be classified, and based on the category identification information (for example, category code) derived from the read address An automatic mail sorting machine for automatically sorting mail items is known. This type of postal mail sorting machine usually collects images of postal items to be sorted by the postal sorting machine body, and sends the collected postal images to an address information reading unit (OCR: Optical Character Reader). The address is being read.

  There is also an automatic mail sorting machine that sends an image of a mail piece to an operator input unit when the address information reading unit fails to read the address and requests the operator to input a correct value. Such a complementary input function by an operator is generally called a video coding disc. When an operator visually recognizes and inputs a name (postal code or address string) included in a postal image displayed on a display, The category specifying information is derived based on the input information.

  In recent years, various recognition algorithms have been developed in the technical field of character recognition, and the recognition rate has been improved. However, characters written as mail addresses include not only handwritten characters and printed characters, but also various types of fonts, so it is difficult to handle with one recognition algorithm, and the recognition performance is limited. there were.

Therefore, it has been proposed to read addresses in parallel from a single mail image using a plurality of address information reading units having different recognition algorithms. For example, the character recognition device described in Patent Document 1 collates a character pattern obtained from a postal matter image with a reference pattern for reference in a composite similarity extraction process by a first recognition unit, and recognizes a recognition result for a read character. As the first character candidate is obtained, the shape of the character pattern obtained by reading and the shape data for reference are collated by the contour feature matching process by the second recognition unit, and the second character candidate is obtained as a recognition result for the read character. The character candidate selection index is read by applying these recognition results to the matrix table, and one of the two recognition results is selected according to the read character candidate selection index.
JP 7-271899 A

However, the character recognition device described in Patent Document 1 has several problems.
The first problem is that even if there is a defect or bias in the recognition result, it cannot be easily improved. The reason is that the discrimination table (matrix table) for selecting the recognition result is fixed data, and one of the two recognition results is uniquely selected. Such inconveniences and biases in the recognition result can be improved by updating the discrimination table. However, the character recognition device described in Patent Document 1 does not consider updating the discrimination table by the user.

  The second problem is that the capacity of the discrimination table is large and it takes time to refer to the discrimination table. The reason is that the discrimination table of the character recognition device described in Patent Document 1 is a matrix table that defines selection indexes for each character candidate. Specifically, confidence values of similarity that are recognition elements of the first character candidate in the first recognition unit are arranged in the X direction for each character candidate, and become recognition elements of the second character candidate in the second recognition unit. The shape features are arranged in the Y direction for each character candidate, and the character candidate selection index based on the contrast of both arrangements is registered in the table column in advance, and its capacity becomes enormous.

  The present invention has been considered in view of the above circumstances, and uses a plurality of address information reading units with different recognition algorithms to read addresses in parallel from a single mail image, while providing operator input information. The feature quantity of each address information reading unit is stored in the database, and the discrimination table is updated based on the stored information in the database, thereby improving the recognition result defects and bias voluntarily and increasing the classification completion rate of postal items. An object is to provide an automatic mail sorting machine and an automatic mail sorting method that can be improved.

  In order to achieve the above object, the automatic mail sorting machine of the present invention collects images of mail pieces to be sorted, reads addresses from the collected mail images, and based on the classification specifying information derived from the read addresses, A mail sorting machine that automatically sorts mail items, using a plurality of address information reading units with different recognition algorithms, a parallel reading unit that reads addresses in parallel from one mail image, and a discrimination table While referring to the recognition result integration unit that derives the category identification information that seems to be correct by comparing the recognition results and feature quantities from each address information reading unit, and when it fails to derive the category identification information that seems to be correct The correct address information reading unit or the operator input unit that requests the operator to input correct answer category specifying information, and the input information by the operator and each address information reading unit A database for storing a symptom amount, based on the stored information in the database, it is constituted; and a learning unit for updating the determination table.

  In this way, in the automatic mail sorting machine that reads addresses from a single mail image in parallel using a plurality of address information reading units with different recognition algorithms, it is possible to improve the mail mail sorting completion rate. The reason is that the input information by the operator and the feature quantity from each address information reading unit are accumulated in the database, and the discrimination table is updated based on the accumulated information in the database, so that the defect and the bias of the recognition result can be independently improved. Because.

Further, the automatic mail sorting machine according to the present invention accumulates the correct address information reading unit information that is input information by the operator, the address area detection information that is the feature amount from each address information reading unit, and the likelihood in the database. The learning unit learns the prototype of each class using the correct address information reading unit as the feature space class, address area detection information and likelihood as the feature space feature vector, based on the accumulated data in the database, The prototype can be copied to the determination table.
In this way, not only the capacity of the discrimination table can be reduced, but also the reference time for the discrimination table can be shortened. The reason is that the data written in the discrimination table is not matrix data in which a selection index for each character candidate is defined, but is prototype data of feature quantities having each address information reading unit as a feature space class.

In the automatic mail sorting machine according to the present invention, the recognition result integration unit obtains the Euclidean distance from the prototype determined by the discrimination table using the feature quantity from each address information reading unit as an input vector, and the Euclidean distance is the smallest. It can be set as the structure which makes the recognition result of an address information reading part the correct answer.
In this way, it is possible to capture the characteristics of each address information reading unit (recognition algorithm) and select the recognition result of the address information reading unit most suitable for the recognized character. Further, it is possible to easily cope with an increase or decrease in the number of address information reading units.

Further, the automatic mail sorting machine of the present invention compares the Euclidean distance of each address information reading unit with the rejection threshold when the recognition result integration unit compares all of the address information when all the Euclidean distances are larger than the rejection threshold. It can be set as the structure which makes the recognition result of a reading part incorrect.
In this way, it is possible to avoid erroneous categorization information derivation and request the operator to input the correct address information reading unit or correct value.

Further, in the automatic mail sorting machine of the present invention, the recognition result integration unit compares the recognition result from each address information reading unit, and when all the recognition results are the same, the recognition result of all the address information reading units Can be set as a correct answer.
In this way, it is possible to speed up the process of deriving the category specifying information by omitting the reference table reference process and the Euclidean distance calculation process.

Further, the mail automatic classification method of the present invention collects images of mail pieces to be classified, reads addresses from the collected mail images, and automatically sorts mail items based on classification specifying information derived from the read addresses. Automatic postal sorting method, which uses a plurality of address information reading units with different recognition algorithms, reads addresses in parallel from one postal image, and reads each address information while referring to the discrimination table Compare the recognition results and feature quantities from the part to derive the category identification information that seems to be correct, and if the derivation of category identification information that seems to be correct fails, the correct addressee information reading unit or the correct category identification information Requests input from the operator, accumulates information input by the operator and feature amounts from each address information reading unit in the database, and based on the stored information in the database There a way to update the determination table.
In this way, when reading addresses from a single mail image in parallel using a plurality of address information reading units with different recognition algorithms, the operator's input information and feature values of each address information reading unit are stored in the database. In addition, by updating the determination table based on the stored information in the database, it is possible to voluntarily improve defects and bias in recognition results, and improve the postal matter classification completion rate.

The automatic mail classification method of the present invention accumulates in the database the correct address information reading unit information that is input information by the operator, the address area detection information that is the feature amount from each address information reading unit, and the likelihood. Based on the data stored in the database, the correct address information reading unit is used as a feature space class, address area detection information and likelihood as feature space feature vectors, and prototypes of each class are learned. Can be copied.
This makes it possible not only to reduce the capacity of the discrimination table, but also to shorten the reference time of the discrimination table, compared to the case where matrix data that defines selection indexes for each character candidate is used.

The automatic mail classification method of the present invention obtains the Euclidean distance from the prototype determined by the discrimination table using the feature quantity from each address information reading unit as an input vector, and the recognition result of the address information reading unit having the smallest Euclidean distance. Can be made correct.
In this way, it is possible to capture the characteristics of each address information reading unit (recognition algorithm) and select the recognition result of the address information reading unit most suitable for the recognized character. Further, it is possible to easily cope with an increase or decrease in the number of address information reading units.

The automatic mail classification method of the present invention compares the Euclidean distance of each address information reading unit with the reject threshold, and if all the Euclidean distances are larger than the reject threshold, the recognition results of all the address information reading units are rejected. The answer can be correct.
In this way, it is possible to avoid erroneous categorization information derivation and request the operator to input the correct address information reading unit or correct value.

Further, the automatic mail sorting method of the present invention compares the recognition results from the respective address information reading units, and when all the recognition results are the same, the recognition results of all the address information reading units can be made correct. .
In this way, it is possible to speed up the process of deriving the category specifying information by omitting the reference table reference process and the Euclidean distance calculation process.

  As described above, according to the present invention, a plurality of address information reading units having different recognition algorithms are used to read addresses in parallel from a single mail image, while the operator input information and each address information By storing the feature values of the reading unit in the database and updating the discrimination table based on the stored information in the database, it is possible to voluntarily improve defects and bias in recognition results and improve the completion rate of mail classification it can.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, in the drawings, the address information reading unit is appropriately expressed as OCR and the database as DB.

[Automatic mail sorting device]
FIG. 1 is a block diagram showing a configuration of an automatic mail sorting machine according to an embodiment of the present invention.
The automatic mail sorting machine according to the embodiment of the present invention shown in this figure collects images of mail pieces to be sorted, reads addresses from the collected mail images, and creates classification specifying information derived from the read addresses. An automatic mail sorting machine that automatically sorts mail pieces based on a mail sorting machine body unit 1, a parallel reading unit 2, a recognition result integration unit 3, a recognition result output unit 4, and an operator input unit. 5, a database 6, and a learning unit 7.

  The postal sorting machine body 1 collects images of postal items to be sorted using a scanner or the like, sends the collected postal images to the parallel reading unit 2, and returns the classification specifying information returned from the recognition result output unit 4. Based on this, the mail is sorted. The category specifying information is not particularly limited as long as the category of the mail can be specified. In the present embodiment, a predetermined classification code is used.

The parallel reading unit 2 reads addresses in parallel from one mail image using a plurality of address information reading units 2a and 2b having different recognition algorithms. When the reading process of one postal matter image is completed, the recognition result, the feature amount, and the processed image are output from the parallel reading unit 2. In this embodiment, a classification code is output as a recognition result, and ABF area information (Address Block Finding information: address area detection information) and a likelihood value are output as feature amounts.
In this embodiment, the address is read by the two address information reading units 2a and 2b, but the address may be read by three or more address information reading units.

The recognition result integration unit 3 compares the recognition results and the feature amounts from the respective address information reading units 2a and 2b while referring to the determination table, and derives a classification code that seems to be correct.
The recognition result output unit 4 returns the classification code to the postal sorting machine main unit 1 when the recognition result integration unit 3 succeeds in deriving the classification code that seems to be correct, and the recognition result integration unit 3 seems to be correct. When the derivation of the classification code has failed (hereinafter, appropriately rejected), the recognition result, the feature amount, and the processed image of each address information reading unit 2a, 2b are sent to the operator input unit 5.

  The operator input unit 5 displays the recognition results and processed images of each address information reading unit 2a, 2b, and inputs the correct address information reading unit (correct OCR number) or correct value (postal code or address character string) to the operator. Request. Here, when the correct address information reading unit is input, the classification code that is the recognition result of the address information reading unit is sent to the postal sorting machine main unit 1, and when the correct value is input, it is input to the classification code. And is sent to the mail sorting machine main unit 1.

  The database 6 stores information input by the operator and feature amounts from the respective address information reading units 2a and 2b. For example, correct OCR numbers are accumulated as input information by the operator, and ABF area information and likelihood are accumulated as feature amounts from the respective address information reading units 2a and 2b.

  The learning unit 7 learns based on the stored information in the database 6 and updates the discrimination table. For example, based on the data stored in the database 6, the correct OCR number is used as the feature space class, the ADF area information and the likelihood is used as the feature space feature vector, and the prototype of each class is learned, and this prototype is copied to the determination table. To do.

When the automatic mail sorting machine is configured in this way, the use of a plurality of address information reading units 2a and 2b having different recognition algorithms, the independence of the recognition results is not found when reading addresses from a single mail image in parallel. Can be improved, and the mail completion rate can be improved.
In addition, since the data written in the discrimination table is not matrix data in which a selection index for each character candidate is defined, but is prototype data of feature quantities having each address information reading unit 2a, 2b as a class of feature space. In addition to reducing the capacity of the table, it is possible to shorten the reference table reference time.

  Then, the recognition result integration unit 3 uses the feature amounts from the respective address information reading units 2a and 2b as input vectors to determine the Euclidean distance from the prototype determined by the discrimination table, and the address information reading unit 2a having the smallest Euclidean distance. The recognition result of 2b can be a correct answer. Thereby, the feature of each address information reading part 2a, 2b can be caught, and the recognition result of address information reading part 2a, 2b most suitable for a recognition character can be selected.

  Further, the recognition result integration unit 3 compares the Euclidean distances of the respective address information reading units 2a and 2b with the reject threshold value, and when all the Euclidean distances are larger than the reject threshold value, all of the address information reading units 2a and 2b are compared. The recognition result can be an incorrect answer. In this way, it is possible to avoid derivation of an erroneous classification code, and to request the correct address information reading unit or the input of the correct value in the operator input unit 5.

  Further, the recognition result integration unit 3 compares the recognition results from the respective address information reading units 2a and 2b. If all the recognition results are the same, the recognition results of all the address information reading units 2a and 2b are regarded as correct answers. can do. Thereby, the reference table reference processing and the Euclidean distance calculation processing can be omitted, and the classification code derivation processing can be speeded up.

  Next, an automatic mail sorting method according to an embodiment of the present invention will be described with reference to FIG.

[Automatic mail sorting method]
FIG. 2 is a flowchart showing the sorting process procedure of the automatic mail sorting machine according to the embodiment of the present invention.
As shown in this figure, when the mail sorting process starts, the mail sorting machine main unit 1 collects images of the supplied mail items with a scanner or the like, and sends the images to the parallel reading unit 2 (S11).
The parallel reading unit 2 reads the address in the mail image sent from the postal sorting machine main unit 1 in parallel by the plurality of address information reading units 2a and 2b, and derives the classification code (S12).

The recognition result integration unit 3 receives the recognition results, feature amounts and processed images of the address information reading units 2a and 2b from the parallel reading unit 2, and recognizes the address information reading units 2a and 2b while referring to the determination table. The result and the feature quantity are compared to derive a category code that seems to be correct (S13).
The recognition result output unit 4 determines whether or not the recognition result integration unit 3 has succeeded in deriving the classification code (S14). If the determination result is YES, the classification result is sent to the postal sorting machine main unit 1. (S15) When NO, the recognition result, the feature amount, and the processed image are sent to the operator input unit 5.

The operator input unit 5 displays the recognition results and processed images of each address information reading unit 2a, 2b, and inputs the correct address information reading unit (correct OCR number) or correct value (postal code or address character string) to the operator. A request is made (S16).
When the input by the operator is completed (S17), the classification code is sent to the postal sorting machine main unit 1 (S18), and the OCR number input by the operator to the database 6 and the ABF which is the feature amount of each address information reading unit 2a, 2b. Area information and likelihood are accumulated (S19).
When 10,000 data items are accumulated in the database 6 (S20), the above-described prototype is calculated by applying these data to the learning function (S21), and the discrimination table is updated by copying this prototype (S21). S22).

Next, specific examples of the present invention will be described with reference to FIGS.
FIG. 3 is an explanatory diagram illustrating an example of a screen displayed by the operator input unit of the automatic mail sorting machine according to the first embodiment of the present invention, and FIG. 4 is a database of the automatic mail sorting machine according to the first embodiment of the present invention. FIG. 5 is an explanatory diagram showing an example of stored data, FIG. 5 is an explanatory diagram showing an example of a prototype set in the discrimination table of the automatic mail sorting machine according to the first embodiment of the present invention, and FIG. 6 is an embodiment of the present invention. It is explanatory drawing which shows the example of the characteristic information sent from each address information reading part of the mail automatic sorting machine which concerns on Example 1. FIG.

  In the postal sorting machine main unit 1, 256-gradation grayscale images of the supplied mail are collected by a scanner or the like, and the images are sent to the parallel reading unit 2. In this embodiment, a case where two address information reading units 2a and 2b are connected to the parallel reading unit 2 will be described.

The parallel reading unit 2 reads the addresses in the mail image sent from the postal sorting machine body unit 1 using the address information reading units 2a and 2b connected in parallel. When reading the address, the grayscale image of 256 gradations is binarized using an appropriate threshold value, and recognition processing is performed on the binarized image. Each address hierarchy (postal code, prefecture, branch name, chome, street address, company name, destination name, etc.) is subject to address reading.To derive the classification code as the final result, the classification code must be unique. Therefore, it is necessary to have read information of each layer that can be determined by the following.
Each address information reading unit 2a, 2b of the parallel reading unit 2 outputs the classification code, ABF coordinate information (x coordinate, y coordinate), likelihood and processed image to the recognition result integrating unit 3.

  The recognition result integration unit 3 determines which division area the ABF coordinate information applies to when the postal image is divided into nine pieces of information inputted from the respective address information reading units 2a and 2b, and uses the ABF coordinate information. It is converted into 9-divided screen information (ABF area information). Next, the recognition results (classification codes) of the respective address information reading sections 2a and 2b inputted from the parallel reading section 2 are compared. If all the classification codes are the same, the classification code is recognized as the recognition result output section 4. Output to. Further, when the classification codes recognized by the respective address information reading units 2a and 2b are different, the rejection determination is made, and the ABF area information, the likelihood, and the processed image are sent to the recognition result output unit 4.

  The correct classification code input to the recognition result output unit 4 is sent to the postal sorting machine main unit 1, and based on the classification code, the supplied mail is classified into a designated classification box, and the classification is completed. In the case of rejection, the ABF area information, likelihood, and processed image of each address information reading unit 2a, 2b are sent to the operator input unit 5.

In the operator input unit 5, the operator visually checks the mail image. The operator input unit 5 includes a display for displaying a mail image, a mouse for inputting information, and a keyboard. FIG. 3 shows an image display example in the operator input unit 5. The left side of the display screen shown in FIG. 3 is an area 5a for displaying a binarized image processed by one address information reading unit 2a, and the right side is an area 5b for displaying a binarized image processed by the other address information reading unit 2a. It is. In the upper row, there are a column 5c for the operator to input a correct OCR number (1 or 2 in this embodiment) and a column 5d for inputting a correct value (classification code, postal code, address, etc.). Further, when the address information reading units 2a and 2b can read several address hierarchies, the regions 5e and 5f below the regions 5a and 5b displaying the processed images of the address information reading units 2a and 2b are displayed. The result of reading the category code is displayed.
Then, the operator types the correct OCR number or correct value. The keyed correct answer classification code is sent to the postal sorting machine main unit 1, and the postal matter keyed with the correct answer value is classified based on the keyed correct answer value to complete the classification.

The correct OCR number input by the operator input unit 5 is sent to the database 6 and stored together with the ABF area information and likelihood of each address information reading unit 2a, 2b as shown in FIG.
The data stored in the database 6 is applied to the learning function of the learning unit 7 when 10,000 data are stored.

  In this embodiment, as shown in FIG. 4, since the two address information reading units 2a and 2b have ABF area information and likelihood as feature quantities, respectively, the address information selected at that time using a four-dimensional feature vector. Features of the reading units 2a and 2b can be expressed. In the feature space spanned by this feature vector, there are two classes corresponding to each address information reading unit 2a, 2b, and a prototype is set as a representative pattern of the class. The prototype can be easily obtained by using, for example, a widely known k-means method. Also, each feature quantity is standardized so that the pattern distribution aspect in the feature space does not change on the feature quantity scale. The prototype obtained is shown in FIG. If each class cannot be linearly separated by one prototype, the number of prototypes is increased and each class is separated.

The prototype obtained by learning is reflected in the determination table. The discrimination table obtained by learning is updated by copying the prototype to the existing discrimination table of the recognition result integration unit 3.
After the update, referring to the updated discrimination table, the output values of the address information reading units 2a and 2b are compared to derive the classification code. The discrimination table is referenced by standardizing the four-dimensional feature amounts from the respective address information reading units 2a and 2b and using the standardized values as input vectors and the prototypes of the address information reading units 2a and 2b set in the discrimination table. This can be done by obtaining the Euclidean distance. For example, in the case of the input vector as shown in FIG. 6, the Euclidean distance from the prototype of the address information reading unit 2a is 2.24, the Euclidean distance from the prototype of the address information reading unit 2b is 2.48, and the address information reading unit Since the Euclidean distance 2a has the smallest value, the output recognition result is that of the address information reading unit 2a. Also, a rejection threshold for the Euclidean distance is determined, and if all the Euclidean distances are larger than the value, the rejection is made. Moreover, it is preferable to reject also when the Euclidean distance of each address information reading part 2a, 2b is equal.

  Further, if the desired recognition accuracy cannot be obtained even if the updated discrimination table is used, the ABF area information and the likelihood of each address information reading unit 2a, 2b are further accumulated to perform learning again, and the discrimination table Update.

  The present invention collects images of postal items to be classified, reads addresses from the collected postal images, and automatically classifies postal items based on classification specifying information derived from the read addresses Applicable to machine and postal mail sorting method. In particular, the present invention is useful in an automatic postal sorting machine and an automatic postal sorting method that read addresses in parallel from one postal matter image using a plurality of address information reading units having different recognition algorithms.

It is a block diagram which shows the structure of the mail automatic sorting machine which concerns on embodiment of this invention. It is a flowchart which shows the division | segmentation processing procedure of the mail automatic sorting machine which concerns on embodiment of this invention. It is explanatory drawing which shows the example of the screen which the operator input part of the mail automatic sorting machine which concerns on Example 1 of this invention displays. It is explanatory drawing which shows the example of the data accumulate | stored in the database of the mail automatic sorting machine which concerns on Example 1 of this invention. It is explanatory drawing which shows the example of the prototype set to the discrimination | determination table of the mail automatic sorting machine which concerns on Example 1 of this invention. It is explanatory drawing which shows the example of the characteristic information sent from each address information reading part of the mail automatic sorting machine which concerns on Example 1 of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Postal sorting machine main-body part 2 Parallel reading part 2a Address information reading part 2b Address information reading part 3 Recognition result integration part 4 Recognition result output part 5 Operator input part 6 Database 7 Learning part

Claims (8)

An automatic mail sorting machine that collects images of postal items to be classified, reads addresses from the collected mail images, and automatically classifies mails based on the classification specific information derived from the read addresses ,
A parallel reading unit that reads addresses in parallel from one mail image using a plurality of address information reading units with different recognition algorithms;
While referring to the discrimination table, the recognition result integration unit for deriving the category identification information that seems to be correct by comparing the recognition result and the feature amount from each address information reading unit,
An operator input unit for requesting the operator to input a correct answer address information reading unit or a correct value when derivation of the category specific information that seems to be correct has failed;
A database for storing information input by the operator and feature amounts from each address information reading unit;
A learning unit for updating the discrimination table based on the accumulated information in the database;
Equipped with a,
The database stores correct address information reading unit information that is input information by an operator, address area detection information and likelihood that is a feature amount from each address information reading unit,
The learning unit learns a prototype of each class using the correct address information reading unit as a feature space class, address area detection information and likelihood as a feature space feature vector based on the accumulated data in the database. automatic mail sorting machine, wherein the copy it to the determination table. The recognition result integration unit obtains the Euclidean distance from the prototype determined by the discrimination table using the feature quantity from each address information reading unit as an input vector, and the recognition result of the address information reading unit having the smallest Euclidean distance is set as a correct answer. The automatic mail sorting machine according to claim 1 . The recognition result integration unit compares the Euclidean distance of each address information reading unit with a reject threshold, and if all the Euclidean distances are larger than the reject threshold, the recognition result of all the address information reading units is incorrect. Item 2. Automatic mail sorting machine according to item 2 . The recognition result integration unit, compares the recognition result from the address information reading unit, when all the recognition results of the same, more of claims 1 to 3 to correct the recognition results of all of the address information reading unit The automatic mail sorting machine described in Crab. An automatic postal classification method that collects images of postal items to be classified, reads addresses from the collected postal images, and automatically classifies postal items based on classification specific information derived from the read addresses ,
Using multiple address information reading units with different recognition algorithms, the address is read in parallel from one postal image,
While referring to the discrimination table, compare the recognition results and feature quantities from each address information reading unit to derive the category specific information that seems to be correct,
If derivation of the category specific information that seems to be correct fails, request the operator to input the correct addressee information reading part or correct value,
Accumulate information input by the operator and feature values from each address information reading unit in the database,
Based on the accumulated information in the database, update the discrimination table ,
In the database, correct address information reading part information that is input information by the operator, address area detection information and likelihood that is a feature amount from each address information reading part,
Based on the data stored in the database, the correct addressee information reading unit is used as the feature space class, address area detection information and likelihood as the feature space feature vector, and the prototype of each class is learned, and this prototype is copied to the discrimination table. An automatic mail sorting method characterized by: 6. The postal automatic system according to claim 5, wherein a feature value from each address information reading unit is used as an input vector to obtain a Euclidean distance from a prototype determined by a discrimination table, and a recognition result of the address information reading unit having the smallest Euclidean distance is taken as a correct answer. Classification method. The postal automatic classification according to claim 6, wherein the Euclidean distance of each address information reading unit is compared with a reject threshold, and if all the Euclidean distances are larger than the reject threshold, the recognition results of all the address information reading units are incorrect. Method. Comparing recognition results from the address information reading unit, when all the recognition results of the same, automatic mail classification according to any one of claims 5-7 to correct the recognition results of all of the address information reading unit Method.

Images