JPH05307639A - Device for detecting address area of postal matter - Google Patents

Abstract

PURPOSE:To make it possible to detect an address area on a postal matter including a stamp, a postmark, a zip code, an advertizing sentence, sender's address and name, etc., in addition to an address without requiring algorithm for extracting a feature and to improve the reading performance of address reading. CONSTITUTION:The original image of a postal matter P on which an address is written is photoelectrically converrted by a photoelectric conversion part 4a, an address area on the postal matter P out of its original image is outputted by an address area detecting part 42 by means of a neural network, a teacher signal pattern to be a required result for the address area of the matter P is instructed by a personal computer, and the branch weight of the neural network is changed based upon the instructed teacher signal pattern and the address area outputted by the neural network in the detecting part 42.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an address area detecting device for mail, which detects an address area of mail.

[0002]

2. Description of the Related Art In recent years, in recognition devices for automatically recognizing characters, recognition techniques have been developed up to the area of handwritten Chinese characters.

In the field of processing mail such as an optical reading and sorting device using such a recognition device, the postal code on the mail is read, and depending on the read result, the mail is divided into a plurality of delivery areas. There is a postal code automatic reading / sorting machine for sorting, or a delivery sorting machine for further reading the address (address) on the mail, and sorting the mail for each of a plurality of delivery areas according to the read result.

However, in the optical character reader for reading the address of the mail used in the delivery sorting machine as described above, the detection of the address area is very important. Besides, stamps, postmarks, postal codes, advertisements, sender's address, name, etc. are mixed, and there is no fixed law in the place where the address is written, the format, etc., so the address area should be detected correctly. Was extremely difficult.

However, whether or not the address area is correctly detected has a great influence on the subsequent processing of the delivery sorting machine, and is one of the most important techniques. Various methods have been attempted as a method for detecting the address area, and one using a feature extraction algorithm has been proposed as one of them, but a parameter that completely indicates the characteristics of the address area has not been found. Is the current situation. Therefore, conventionally, there is a problem that the address cannot be read because the address area is not correctly detected. Therefore, there are drawbacks such that a feature extraction algorithm is required and the addressing area detection performance of the mail is poor.

[0006]

Conventionally, as described above, there are drawbacks in that a feature extraction algorithm is required and the performance of detecting the address area of a mail is poor.

Therefore, according to the present invention, by letting the neural network detect the address area of the mail for which it is difficult to construct the feature extraction algorithm, the feature extraction algorithm is not required, and the mail can be improved in performance easily. It is an object of the present invention to provide an address area detection device for an object.

[0008]

SUMMARY OF THE INVENTION The address area detecting apparatus for mail of the present invention is a supply means for supplying an original image of a mail or an compressed image of a mail with an address, and a mail supplied by this supply means. Output means for outputting the address area of the mail to the original image of the object or its compressed image using a neural network, instructing means for instructing a desired result teacher signal pattern for the address area of the mail, and this instructing means It is composed of changing means for changing the branch weight of the neural network in the output means according to the teacher signal pattern and the address area output by the output means.

The address area detecting apparatus for mail of the present invention is
A supply means for supplying the original image of the mail with the address described or its compressed image, and the address area of the mail for the original image of the mail or the compressed image supplied by this supplying means is output using a neural network. And output means for changing the branch weights of the neural network in the output means so that the output means outputs a desired result.

[0010]

According to the present invention, in the above-mentioned structure, the original image of the mail with the address written or its compressed image is supplied by the supply means, and the original image of the mail supplied or its compressed image is supplied. The address area of the mail is output by the output means using a neural network, the teacher signal pattern as a desired result for the address area of the mail is instructed, and the instructed teacher signal pattern is output by the output means. The branch weight of the neural network in the output means is changed according to the address area.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 2 shows an automatic mail sorting machine as an address reading sorting machine according to the present invention, which has the following configuration.

That is, the postal items P, such as postcards and sealed letters, are set in the supply unit 1 in a lump and in a standing position. The postal matter P of the supply unit 1 is sequentially taken out and conveyed one by one from the frontmost postal matter P by the take-out unit 2 and given to the postal matter P by the reading unit 4 arranged in the middle of the take-in conveyance path 3. Address information (address information) is read. The reading unit 4 includes a photoelectric conversion unit 4a for photoelectrically converting the address information on the postal matter P, and a determination unit 4b for determining the address information by recognizing the character pattern according to the output of the photoelectric conversion unit 4a. It is well known. Next, the postal matter P that has passed through the reading unit 4 is conveyed to the sorting unit 11 via the conveying path 3.

The sorting section 11 is provided with a plurality of sorting conveyance paths 5 vertically arranged by a gate mechanism section (not shown).
a to 5g are selectively distributed based on the result of reading the address information.

A plurality of pockets (collection boxes) 6 are provided below the sectioned conveying paths 5a to 5f along the lower surface thereof.
Are arranged in a plurality of stages (A stage to G stage), and the mail P sent to the sorting conveyance paths 5a to 5g is stored in a predetermined pocket 6 based on the reading result of the reading unit 4. It is designed to be collected separately. In addition, the transport paths 3, 5a,
A transport detector group (not shown) that detects the transport of the postal matter P on the transport path is provided above.

As shown in FIG. 3, the address information A and the like are written on the surface of the mail P. Address information A above
Are each composed of a plurality of lines of character information. In such a postal matter P, “prefecture name”, “ward, city name”, “town name”, which should be read as the address information A,
The address such as "address" is described. An operator panel 7 as an operation panel is provided above the supply unit 1.
Is provided.

FIG. 4 shows an electric circuit. That is, the control unit 31 uses the division designation table stored in the memory 32 to read the division designation data corresponding to the address information supplied from the discrimination unit 4b, and the read division designation data (pocket 6, ...). The mail P is transported to the pockets 6, ... For example, by outputting a sorting gate control signal in response to a transportation detection signal from a transportation detector (not shown) in the middle of the transportation path 3, the driver 33 can be used to control the sorting gates 9. There is. In the memory 32, the addresses of the pockets 6, ... (Division designation data) corresponding to each address information are stored as a division designation table. Further, the control unit 31 controls the entire transport system by driving the transport mechanism unit 36 such as the transport path 3 with the driver 35.

An external device such as a personal computer 71 is connected to the control unit 31, and the neural network in the address area detection unit 42, which will be described later, is learned by using the keyboard, mouse, and display unit of the personal computer 71. The teacher signal pattern of is input. Next, the photoelectric conversion unit 4a and the determination unit 4b of the reading unit 4 will be described with reference to FIG.

That is, the photoelectric conversion unit 4a is for obtaining a pattern signal (read signal) by optically scanning the postal matter P and photoelectrically converting it, and for example, a light source for irradiating the postal matter P with light. , And a self-scanning CCD image sensor that receives the reflected light and converts it into an electric signal.

The output of the photoelectric conversion section 4a is supplied to the discrimination section 4b. The discrimination unit 4b includes a binarization circuit 41, an address area detection unit 42, a character recognition unit 43, a town name / block recognition unit 44, an address dictionary 45, and an address recognition unit 46.

The binarization circuit 41 includes the photoelectric conversion unit 4
The read signal from a is binarized. The binarized signal for the entire postal matter P from the binarization circuit 41 is each pixel value (1 or 0) for the original image, as shown in FIG. Are indicated by diagonal lines. The output of the binarization circuit 41 is the address area detection unit 4
2 and the selection circuit 51, which will be described later, in the character recognition unit 43.

The address area detecting section 42 is arranged to detect the mail P
The neural network is used to detect the address area in which the address information is described from among the information described in (1), and the data indicating this address area is output.

As shown in FIGS. 6 and 7, the address area detecting unit 42 is composed of what is called a multilayer perceptron, and is composed of a hierarchical network composed of an input layer 61, an intermediate layer 62 and an output layer 63. Has been done.

As a result, as shown in FIG. 7, each pixel value of the original image is input to each unit of the input layer 61, and the output layer 6
Each unit value of 3 becomes each pixel value of the output image. The output takes a binary value of "0" and "1", and "1" (hatched portion) indicates the address area.

However, the intermediate layer 62 may have any number of layers,
Or it may be omitted. Each layer is composed of an appropriate number of units, there is no coupling between units of the same layer, and the coupling between layers is the coupling in the positional direction of the input layer 61, the intermediate layer 62, and the output layer 63.

Each unit has a sum of input signals to the unit as an input and has a differentiable and non-linear input / output characteristic such as a sigmoid function. That is, it has a threshold characteristic that the input signal is output if it is above a certain level, and is not output if it is below that level.

The coupling element between each unit generally has a variable value called a branch weight, and the product of the transmission signal and the branch weight is transmitted, such as a variable resistor in the electric element. It is a thing.

The branch weight initially has a random value, but by connecting an external device (not shown) such as a personal computer, each branch weight is changed from the teacher signal pattern input from it, The desired result is achieved. This is called neural network learning, and is a method typified by backpropagation. The determination result of the address area detection unit 42, for example, data indicating the determined address area is output to the selection circuit 51 in the character recognition unit 43.

The character recognition section 43 supplies from the selection circuit 51 a selection circuit 51 which outputs the binarized signal from the binarization circuit 41 corresponding to the address area supplied from the address area detection section 42. A character detection / cutout circuit 52 for detecting and cutting out the characters of the address information corresponding to the addressed area (address area) one by one, that is, for detecting and cutting out the character information corresponding to the address (address). That is, the character information that is supplied from 52, that is, the detected and cut out character information is normalized,
Normalization circuit 53 for sampling, and normalization circuit 5
The character information supplied from No. 3 is constituted by a recognition circuit 54 for recognizing the character by, for example, a method of matching the character in the dictionary 55 with a reference pattern.

The output of the character recognition unit 43, that is, the recognition circuit 54 is supplied to the town name / block recognition unit 44. Above town name
The block recognition unit 44 recognizes the town name / block by the address read from the address dictionary 45 for the recognition character supplied from the character recognition unit 43.

The output from the town name / block recognition unit 44 is supplied to the address recognition unit 48. The address recognition unit 48 recognizes the address by the town name / block supplied from the town name / block recognition unit 44, and outputs the classification designation data for the address, that is, the data indicating the pockets 6, ... Is.

Learning of the neural network in the address area detecting unit 42 will be described. For example, now
An external device such as a personal computer 71 is connected to the control unit 31. Then, the personal computer 71
Set the learning mode with the keyboard and instruct to start learning. Further, the postal items P for learning are set in the supply unit 1.

In such a state, the control unit 31 drives and controls the transport mechanism unit 36 so that the mail items P accumulated in the supply unit 1 are taken out one by one by the take-out unit 2.
It is transported on the transport path 3. During the transportation, the information on the postal matter P is read by the photoelectric conversion unit 4a in the reading unit 4, and this read signal is supplied to the binarization circuit 41. Then, the binarization circuit 41 binarizes the read signal from the photoelectric conversion unit 4a and supplies it to the personal computer 71 and the address area detection unit 42.

The personal computer 71 displays an image based on the supplied binary signal on the display unit. A teacher signal pattern as an address area for the binarized image displayed on the display unit is designated by using a mouse or a keyboard. The teacher signal pattern (address area which is the correct answer) indicated by the shaded area (corresponding to “1”) as shown in FIG. 8 from the personal computer 71 is supplied to the address area detecting section 42.

As a result, in the neural network of the address area detection unit 42, each pixel value (1 or 0) as a binarized signal from the binarization circuit 41 and each as a teacher signal pattern from the personal computer 71. By being supplied with the teacher signal corresponding to the pixel value,
When the network outputs the same output as the correct answer, the coupling does not change, but when the network outputs an output different from the correct answer, the branch weight is modified so that the output becomes the correct answer.

For example, when the correct answer is "1" but "0" is output, the coupling from the unit that has been sending a positive signal to the corresponding unit is increased and a negative signal is being sent. Reduce coupling from the unit. If the correct answer is "0" but "1" is output, the correction is also reversed. This correction is applied from the output layer 63 to the input layer 61.

As a result, the neural network of the address area detecting section 42 receives the correct answer from the personal computer 71 for each pixel value (1 or 0) as the binarized signal from the binarizing circuit 41. Learning is performed using the region as a teacher signal pattern.

Further, the learning of the neural network of the address area detecting unit 42 is performed in the same manner as described above by the binarized signal for the mail P and the teacher signal pattern serving as the correct address area supplied from the personal computer 71. Be seen. The above operation is repeated for several pieces of mail P, ..., A neural network of the address area detecting unit 42 suitable for address area detection is constructed.

By performing the address area detection process using the neural network constructed in this way,
The address area detection unit 42 detects the area accurately. The reading operation of the address information on the postal matter P by the reading unit 4 in such a state will be described.

That is, the photoelectric conversion unit 4a causes the postal matter P
A pattern signal obtained by optically scanning the upper side and performing photoelectric conversion is output to the binarization circuit 41. The binarization circuit 41 outputs the read signal from the photoelectric conversion unit 4a to the destination area detection unit 42 and the selection circuit 51 as a binarized signal.

The address area detection unit 42 detects the address area by processing the binarized signal for the entire mail P supplied from the binarization circuit 41 using a neural network, and detects the detected address area. The data indicating the area is output to the selection circuit 51.

The selection circuit 51 includes the address area detection section 42.
The binarization circuit 41 corresponding to the address area supplied from
The binarized signal from is output to the character detection / cutout circuit 52.
The character detection / cutout circuit 52 detects and cuts out the characters of the address information corresponding to the address area (address area) supplied from the selection circuit 51 one by one, that is, detects the character information corresponding to the address (address) and normalizes it. Output to the circuit 53.

The normalization circuit 53 is a character detection / cutout circuit 52.
The character information that is supplied from, ie, detected and cut out, is normalized, sampled, and output to the recognition circuit 54. The recognition circuit 54 converts the character information supplied from the normalization circuit 53 into
The character is recognized by a matching method with the reference pattern for the character in the dictionary 55, and the recognition result is output to the town name / block recognition unit 44.

The town name / block recognition unit 44 recognizes the town name / block by the address read from the address dictionary 45 with respect to the recognition character supplied from the character recognition unit 43, and the recognition result to the address recognition unit 48. Output. The address recognition unit 48 recognizes the address by the town name / block supplied from the town name / block recognition unit 44, and outputs classification designation data for the address to the control unit 31.

Next, the operation in such a configuration will be described. First, the operator collects the postal matter P in the supply unit 1 and gives an instruction to start from the operator panel 7. Then, the control unit 31 drives and controls the transport mechanism unit 36, so that the postal matter P accumulated in the supply unit 1 is taken out one by one by the take-out unit 2 and is transported through the transport path 3. During this conveyance, the address information on the postal matter P is read by the reading unit 4. Then, the control unit 31 reads out the classification designation data corresponding to the address information supplied from the reading unit 4, and controls the conveyance of the mail P to the pockets 6, ... Corresponding to the classification designation data.

Further, as the teacher signal pattern, as shown in FIG. 8, all the address areas are represented by "1" signals, but not limited to this, as shown in FIG.
It may be a teacher signal pattern in which the circumscribing frame of the addressed area is used as the teacher signal. In this case, the desired output pattern is a circumscribed rectangle, unlike the entire address area, and the boundaries of the address area are expressed in a more explicit manner.

Further, as shown in FIG. 10, a teacher signal pattern for displaying the upper left point and the lower right point of a rectangle circumscribing the addressed area may be used. The output of the neural network may not show the desired rectangle of the address area in some cases, but this makes it possible to uniquely determine the circumscribed frame.

As described above, the original image of the postal matter P in which the address is described is photoelectrically converted by the photoelectric conversion unit 4a, and the addressing area of the postal matter P with respect to the original image of the postal matter P is processed by using a neural network. The address area detection unit 42 outputs the
The personal computer instructs the teacher signal pattern as a desired result for the address area of the mail P, and the neural network is output by the instructed teacher signal pattern and the address area output by the neural network in the address area detector 42. The branch weight of the network is changed.

As a result, in addition to the address, stamps, postmarks,
It is possible to detect the address area, which was difficult to detect by the conventional method, on the mail P including the postal code, advertisement, sender's address, name, etc. without the feature extraction algorithm. It is possible to improve the reading performance.

[0049]

As described in detail above, according to the present invention,
By letting the neural network detect the address area of the mail, which is difficult to construct the feature extraction algorithm,
It is possible to provide a mailing address area detection device that does not require a feature extraction algorithm and is simple and can improve performance.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a reading unit according to an embodiment of the present invention.

FIG. 2 is a sectional view showing a configuration of an address reading / sorting machine according to an embodiment of the present invention.

FIG. 3 is a diagram showing a state of description of address information and the like of the mail of FIG.

FIG. 4 is a block diagram showing a main part of an electric circuit in the address reading / sorting device of FIG.

5 is a diagram showing each pixel value for the original image from the binarization circuit in FIG.

6 is a diagram showing a schematic configuration of a neural network used in the address area detection unit of FIG.

7 is a diagram showing a schematic configuration of a neural network used in the address area detection unit of FIG.

FIG. 8 is a diagram for explaining a teacher signal pattern used in a neural network.

FIG. 9 is a diagram for explaining an example of a teacher signal used in a neural network.

FIG. 10 is a diagram for explaining an example of a teacher signal used in a neural network.

[Explanation of symbols]

P ... Mail, A ... Address information, 4a ... Photoelectric converter, 4b ...
Discrimination unit, 31 ... Control unit, 41 ... Binary circuit, 42 ... Address area detection unit, 61 ... Input layer, 62 ... Intermediate layer, 63 ... Output layer, 71 ... Personal computer.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Internal reference number FI technical display location G06K 9/66 9289-5L

Claims (2)

[Claims] 1. A supply means for supplying an original image of a postal matter having an address or a compressed image thereof, and a postal address area for the original image of the postal matter or its compressed image supplied by the supplying means. Output means for outputting using a neural network; instruction means for instructing a teacher signal pattern as a desired result for the address area of the mail; teacher signal pattern of the instruction means and address area output by the output means And a changing unit for changing the branch weight of the neural network in the output unit according to the above. 2. A supply means for supplying an original image of a postal matter having an address or a compressed image thereof, and an address area of a postal matter for the original image of the postal matter or the compressed image supplied by the supplying means. A postal matter characterized by comprising: output means for outputting using a neural network; and changing means for changing a branch weight of the neural network in the output means so that the output means outputs a desired result. Address area detection device.