JP2001314820A - Device for detecting address region - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve such a problem that a regions showing the structural contents of the same place of residence as an address like a place of residence of a sender is erroneously detected as an address region. SOLUTION: The device for detecting an address region comprises a means 21 for producing candidates for an address region, a means 22 for calculating the characteristic content of the candidate for an address region, a means 31 for storing an address probability of the characteristic content and the candidate for an address region to the characteristic content and a sending probability which is the probability of a region having the structural contents of a place of residence other than an address in such a way that these probabilities are made to correspond to each other, a means 23 for converting the characteristic content of each of the candidates for an address region to the address probability of each of the candidates for an address region and the sending probability, referring to the storage means 31, a means 24 for deciding the sequential order of the candidates for an address region based on the address probability and the sending probability obtained and a means 25 for selecting a specified number of the candidates for an address region in the order starting with the highest address probability.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a destination area detecting device for detecting a destination area on which a destination of a postal matter is described.

[0002]

2. Description of the Related Art Conventionally, such an address area detecting apparatus divides a postal image into several partial areas of a postal image including an area which is an address in the postal image (hereinafter referred to as an addressing area). The address area and the partial area of the postal image are called address area candidates). In this case, in addition to the address area such as the sender's address, an area having the same address structure as the address area (hereinafter, referred to as a “sender area”) must be confused and the address area must be output so as not to be detected. Must.

Therefore, conventionally, only a measure of the likelihood of a destination area is provided for each destination area candidate, and a given number of destination areas are output from a large scale. For example, 1996 International Conference on Pattern Recognition, Volume 3 (Proceedings of 13th International Conf.
erence on Pattern Recognition, Volume III Track
C), Adrian P. Whichello and Hong Yan's paper “Loca
ting Address Blocks and Postcodes in Maii-Piece Im
ages "(p. 716 to p. 720), as described on p. 719, it is more likely that an address area candidate having a smaller distance from a point below the center of the postal image is more likely to be an address area. There is known a method of selecting a destination area candidate in ascending order from a lower center point of a postal image using the mail address.

[0004]

In the prior art, whether or not a mailing area is an addressing area is determined only by a measure of the likelihood of an addressing area. No countermeasures have been taken to distinguish it from a partial image of a postal image that has the same structure as a mailing address. Therefore, the sender's address may be mistaken for the address, and the accuracy of detecting the address area is not always high.

SUMMARY OF THE INVENTION In view of the above-mentioned conventional problems, the present invention can reduce the risk of erroneously detecting an area having the same address structure as an address, such as a sender address, and improve the accuracy of address area detection. It is an object of the present invention to provide a destination address detecting device.

[0006]

According to the present invention, in order to achieve the above object, there is provided a means for generating a plurality of address area candidates including an address area which is an area of address information based on a partial image of a postal image. Means for extracting, for each generated address area candidate, a feature quantity representing an index when evaluating whether the address area is the address area; and address probability indicating the probability that the feature quantity and the address area candidate are the address area. Means for comparing and storing the sending probability indicating the probability of an area having an address structure other than the address of the address area candidate and the pattern probability indicating the probability of an area having no address structure other than the address in the address, Means for converting the feature amount of each address area candidate extracted by the extracting means into an address probability, a sending probability, and a pattern probability by referring to the storage means, and using the address probability and the sending probability among the converted probabilities. hand Means for calculating a score for each addressing area candidate and determining an order based on the score; and means for selecting a predetermined number of addressing area candidates from the addressing area candidates based on the determined order. It is characterized by having.

According to the present invention, in addition to the probability that an address area candidate generated based on a partial image of a postal image is an address, a sending probability having a structure of an address other than the address such as the address of the sender is newly added. Take into account. Then, the order of the destination area candidates is determined using the two destination probabilities and the sending probability, and a predetermined number of destination area candidates are selected in descending order of the probability of the destination based on the obtained order. As a result, erroneous detection of an area having an address structure other than the address is reduced, and detection accuracy of the address area is increased.

[0008]

Next, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a first embodiment of a destination area detecting device according to the present invention. In FIG. 1, 1 is an input device such as a scanner, 2
Is a data processing device that operates under program control, 3 is a storage device that stores information, and 4 is an output device. The input device 1 is a device for inputting a postal image to be processed. The data processing device 2 includes a destination area candidate generation unit 21, a feature amount extraction unit 22, a destination / submission probability calculation unit 23, an order determination unit 24, and an output candidate determination unit 25. The storage device 3 includes an address / submission distribution dictionary 31. The output device 4 outputs a predetermined number of destination area candidates from the destination area candidates in descending order of the probability of being the destination.

Next, the internal configuration of the data processing device 2 will be described. First, terms used in the present specification are defined. First, an area is defined as a partial image of a postal image input by the input device 1. Further, the address area is defined as an area indicating the address part in the postal image input by the input device 1. An area that represents an address instead of an address, such as the sender's address, is called a sender area. An area that is neither an address nor an area representing an address other than the address is called a pattern area. The pattern area is an area that is neither an address area nor a sending area.

The address area candidate generating means 21 generates a plurality of areas including the address area from the postal image input from the input device 1. Hereinafter, each of these areas is referred to as a destination area candidate. For example, taking a postal image as shown in FIG. 2 as an example,
As shown in FIG. 3, the destination area candidates G1, G2, G3, G4 and G5 are generated by the destination area candidate generating means 21. FIG. 4 shows only the locations of the five destination area candidates generated by the destination area candidate generating means 21. Of the plurality of destination area candidates generated by the destination area candidate generation means 21, only one destination area candidate indicating a destination area is provided. Each address area candidate other than the address area is a sender area or a pattern area.

The feature extraction means 22 extracts, from each address area candidate obtained by the address area candidate generation means 21, a feature required to calculate the address / submission probability based on the address / address distribution dictionary 31. It is a means to do. As an example of the feature amount, as shown in FIG. 5, when the origin 0 and the x-axis and the y-axis are set in the postal image, the x-coordinate value and the y-coordinate value of the absolute position of the center point of the destination area candidate are set. Here, the coordinate value is represented by the number of pixels, but this is not essential. The address area candidates G1, G2, G3, G4 and G in FIG.
FIG. 6 shows the feature values in the case of No. 5. For example, the feature quantity of the destination area candidate G1 is x = 600 and y = 80, and the feature quantity of the destination area candidate G2 is x = 480 and y = 80. This feature amount is used to calculate an address / sending probability based on the address / sending distribution dictionary 31.

Here, definitions of the address probability, the sending probability and the pattern probability will be described. First, a probability that a destination area candidate can be regarded as a destination area based on a feature amount extracted from the destination area candidate for one destination area candidate is defined as a destination probability of the destination area candidate. Further, the probability that the destination area candidate can be regarded as the sending area based on the feature amount extracted from the destination area candidate is defined as the sending probability of the destination area candidate. Furthermore,
The probability that the destination area candidate can be regarded as a pattern area based on the feature amount extracted from the destination area candidate is defined as the pattern probability of the destination area candidate. The pattern probability of each destination area candidate is obtained by subtracting the destination probability and the sending probability of the destination area candidate from 1. Address area candidates G1, G2, G3, G in the case of FIG.
The addressing probability, sending probability, and pattern probability of 4 and G5 are, for example, the probabilities as shown in FIG. In the example of FIG. 7, for example,
The destination probability of the destination area candidate G2 is 0.10, the sending probability is 0.20, and the pattern probability is 0.70.

The address / sending probability calculating means 23 uses the address / sending distribution dictionary 31 based on the feature quantity of the address area candidates obtained by the feature quantity extracting means 22 to calculate the address probability of each address area candidate, Calculate the sending probability and the pattern probability. This operation is performed for all the address area candidates. The order determining means 24 is means for arranging the address area candidates in an order suitable for the address area based on the addressing probabilities and the sending probabilities of the respective address area candidates obtained by the address / sending probability calculating means 23. The order for all the address area candidates is obtained by the following processing. First, a score of a destination area candidate is calculated from the destination probability and the sending probability for a certain destination area candidate. When the score of the destination area candidate is represented by m, the way of assigning the score m of the destination area candidate is obtained from the destination probability p and the sending probability q of the destination area candidate by calculation according to the following equation (1).

M = f (p) −g (q / (p + q)) (1) where
The functions f and g are represented by f (u) = u,
(U) = u. Equation (1) is good because q / (p +
This is because q) indicates a possibility that the destination area candidate is not a destination area when it is assumed that the destination area candidate is the destination area or the sending area. This score calculation is executed for all the address area candidates, and the order of the greatest score of all the address area candidates is determined based on the obtained scores. This is the order of all the address area candidates obtained in the order determination processing. In the calculation of the score m, in equation (1), for the argument u, f
(U) = u, g (u) = u, but f (u), g
(U) may be anything as long as it is a monotonically increasing function of the argument u.

The following equation (2), equation (3), equation (4), equation (5), and equation (6) may be used instead of equation (1) for calculating the score.

M = f (p) / g (q / (p + q)) (2)

M = f (p) -g (q) (3)

M = f (p) / g (q) (4)

M = f (p) + g (p / (p + q)) (5)

M = f (p) × g (p / (p + q)) (6) Equation (2) is better because q / (p + q) as in equation (1).
This is because the possibility that the destination area candidate is not the destination area when the destination area candidate is assumed to be the destination area or the sending area is indicated. Equations (3) and (4) are good because f (p) represents the likelihood of a destination area as a destination area candidate, and g (g)
This is because (q) indicates the likelihood of the destination area of the destination area candidate. Also, the reason why the equations (5) and (6) are good is that 1−q / (p + q) = p / (p + q).
This is the same reason as in Expressions (1) and (2).

The order determining means 24 determines the order of the destination area candidates in addition to the method of calculating scores from the addressing probabilities and the sending probabilities of the respective address area candidates and arranging all the address area candidates in the order of larger points. You may do it. For example, it is assumed that there are n destination area candidates. When a certain order is given to n address area candidates, the i-th (i = 1, i = 1,
.., N), assuming that the addressing probability for the addressing area candidate is pi, the sending probability is qi, and the pattern probability is ri, the probability P that the addressing area comes earlier than the sending area in the above order.
Is

[0022]

(Equation 1) Can be calculated by The order of all the n addressing area candidates is obtained, and the order of each of the addressing area candidates is calculated by Expression (7).
, And calculates the order that maximizes the probability P.

The output candidate deciding means 25 comprises the order deciding means 2
Based on the order for the destination area candidates obtained in step 4, the destination area candidates are selected by the number of candidates given from the beginning of the order,
Output those address area candidates.

Next, the storage device 3 will be described. In the storage device 3, an address / sender distribution dictionary 31 is provided. As shown in FIG. 8, the address / submission distribution dictionary 31 includes a conversion table of values of feature ranges, destination probabilities, destination probabilities, and pattern probabilities of the destination area candidates having the feature amounts belonging to the ranges. However, FIG. 8 shows a part of the conversion table. Here, in FIG.
"0" means that the value is 100 or more and less than 110.
When the feature value is converted into the probability value using the conversion table of FIG. 8, for example, the x coordinate value of the center position of the destination area candidate is 1
Assuming that the value is 10 or more and less than 120 and the y coordinate value is 50 or more and less than 55, the addressing probability, the sending probability and the pattern probability of the addressing area candidate are 0.70, 0.
20, 0.10.

Next, a specific operation of the present embodiment will be described. First, a postal matter to be read as an address is read by the input device 1 including a scanner or a camera, and image information on the surface on which the address of the postal matter is written is read. For example, it is assumed that the postal image of FIG. The information of the postal image read by the input device 1 is input to the destination area candidate generating means 21, and the destination area candidate generating means 21 outputs a plurality of address area candidates including the address area based on the partial image of the postal image as described above. Is generated.

In the destination area candidate generating means 21, for example,
Pattern Recognition, 1992, Vol. 25, No. 12 (Pattern Recognition, Vol. 25, No. 12, 19)
92) Anil K. Jain and Sushil K. Bhattacharjee's paper "ADDRESS BLOCK LOCATION ON ENVELOPES USING GA
BOR FILTERS "(p. 1449 to p. 1277)
p. A plurality of address area candidates including the address area are created using the method described in FIG. 1470, and as shown in FIG.
5, five destination area candidates of G2, G3, G4 and G5 are generated. FIG. 4 shows only the locations of the five address area candidates obtained by the address area candidate generating means 21, but G1 is a stamp area, a pattern area, G2 is a postmark area, a pattern area, and G3 is a picture area. Therefore, the pattern area, G4 is the sender's address area because it is the sender's address area, and G5 is the address area.
However, whether each address area candidate is an address area, a sending area, or a pattern area is determined by the address area candidate generating means 21.
Is not known at the time immediately after generating the destination area candidate.

The destination area candidates G 1, G 2, G 3, G 4 and G 5 generated by the destination area candidate generation means 21 are supplied to the feature quantity extraction means 22. FIG. 10 shows a feature amount extracting unit 22.
3 is a flowchart showing the flow of the processing of FIG. First, it is determined whether or not the processing has been executed for all the address area candidates (step A1).
One is selected (step A2). You can choose any order. Next, a feature amount is calculated for the selected destination area candidate (step A3). The feature amount is the coordinate value x, y of the absolute position of the center point of the destination area candidate when the origin 0 and the x-axis and the y-axis are defined in the postal image as described in FIG. The feature amount extraction means 22 repeats the processing of steps A1 to A3, and calculates feature amounts for all the address area candidates.
Address area candidates G1, G2, G3, G4, G in FIG.
The coordinate values x and y of the absolute position of the center point of No. 5 are as shown in FIG. In FIG. 6, for example, the feature amount x of the destination area candidate G2
Is 480 and the feature amount y is 80. In this embodiment, the unit of the coordinate value is the number of pixels. However, this is not an essential problem.

It should be noted that various feature values other than the absolute position of the address region candidate may be used as the feature value of the destination region candidate. For example, there are a relative position to another destination area candidate, a position obtained by normalizing the size of the postal image, and a width and height of a circumscribed rectangle of the destination area candidate. Further, a character line is extracted from the address area candidate by, for example, a method described in Japanese Patent Application No. 7-105575, and the coordinate value of the extracted character line, for example, the x-coordinate of the left end of the character line is determined. Row arrangement information such as whether or not they match in all rows may be used. Further, for example, in 1992, processing and recognition of a line image, co-authored by Takeshi Yasui and Tomoharu Nagao, p. 14 to p. 17 to obtain a feature amount such as an average value or a mode value of the number of white pixel / black pixel inversions in the horizontal direction and the number of continuous black pixels in the horizontal direction with respect to the row image. A value obtained by averaging feature amounts for all character lines may be used.

As shown in FIG. 8, the address / submission distribution dictionary 31 is a conversion table of the range of the feature amount of the destination area candidate and the address probability, the transmission probability, and the pattern probability corresponding to the range.
FIG. 8 shows a conversion table between a combination of the feature quantities x and y itself and a probability value. However, a new feature quantity is calculated by performing a calculation based on a plurality of feature quantities including feature quantities other than x and y. May be calculated, and a conversion table between the calculated value and the probability value may be created.

In the address / sending probability calculating means 23, the address
The likelihood that each address area candidate having the feature amount calculated by the feature amount extraction unit 22 with reference to the sender distribution dictionary 31 is an address area, a sender area, or a pattern area, that is, each address area candidate. A process for calculating the address probability, the sending probability, and the pattern probability of the area candidate is performed. For example, if it is found from the destination area candidate that the position of the center point of the destination area candidate is (x = 115, y = 53) by the feature amount extracting means 22, FIG.
X is the range 110 of the feature amount x
１２０120, and y belongs to the range of 50 to 55 of the feature quantity y. Therefore, when the addressing probability of this addressing area candidate is 0.70, the sending probability is 0.20, and the pattern probability is 0.10, You can ask. The address / sending probability calculating means 23 can obtain the address probability p, the sending probability q, and the pattern probability r of each address area candidate from the feature amounts in FIG. 6 as shown in FIG.

Next, the order determining means 24 performs a process of determining the order of all the address area candidates based on the addressing probabilities and the sending probabilities of the respective address area candidates obtained by the address / sending probability calculating means 23. FIG. 11 is a flowchart showing the flow of the processing of the order determining means 24. First, it is determined whether or not the processing has been executed for all the address area candidates (step B1), and if there is an unprocessed candidate, one is selected (step B2). You can choose any order. Next, for the selected destination area candidate, the score of the destination area candidate is calculated by p and q in accordance with Equation (1), where p is the destination probability of the destination area candidate and q is the sending probability (step B).
3). However, the functions f and g are f
(U) = u and the function g (u) = u.

Thereafter, the processing of steps B1 to B3 is repeated to calculate the score of each address area candidate. When the scores of all the address area candidates are calculated, the address area candidates are sorted in descending order of the score (step B4). This is the order of all the address area candidates obtained in the order determination processing.
In step B3, for example, in the case of FIG. 6, when the score is calculated based on equation (1), the result shown in FIG. 9 is obtained, and the order of G5, G4, G3, G1, G2 is obtained. Can be Note that f (u) = u and g (u) = u
However, f (u) and g (u) may be anything as long as they are monotonically increasing functions of the argument u. As described above, instead of the equation (1), the equations (2) to (6) are used for the score calculation.
May be used.

FIG. 12 is a flowchart showing another order determining method of the order determining means 24. First, all possible orders of a plurality of destination area candidates are created (step C).
1). Next, it is determined whether or not the processing has been completed for all the obtained orders (step C2). If there is any unprocessed one, one unprocessed one is selected from all the orders (step C2). Step C3). You can choose any order. Next, with respect to the order of the selected destination area candidates, the probability P that the destination area for the order comes earlier than the source area is calculated based on the equation (7) (step C4). In this way, the processing of steps C2 to C4 is repeated, and the processing of obtaining the probabilities P for all the orders is performed. In addition, a process is performed to determine the order of the destination area candidates that maximizes the probability P from all the orders of the destination area candidates (step C).
5). This is the order for all the address area candidates obtained in the order determination processing.

The output candidate determining means 25 outputs the destination area candidates from the head of the order based on the order of the destination area candidates obtained by the order determining means 24. If the number of destination area candidates to be output is predetermined to be, for example, two, G5,
G4 is output. Also, if the number of destination area candidates to be output is 1
If so, G5 is output.

Next, a method for creating the address / submission distribution dictionary 31 will be described. The address / submission distribution dictionary is a list in which the address ranges, the address probabilities, and the pattern probabilities of the address region candidates having the characteristic amounts within the range of the characteristic amounts are arranged for various characteristic amount ranges. Now, the range of the feature amount is represented by A, the destination probability of the destination area candidate having the feature amount within the range is p (A), the sending probability is q (A), and the pattern probability is r.
(A). Also, functions hp (A), hq (A), and hr (A) for each range A are prepared, and are respectively referred to as a frequency of a destination area, a frequency of a sending area, and a frequency of a pattern area.

FIG. 13 is a flowchart showing a method of creating the address / submission distribution dictionary 31. First, m mail images are prepared in advance as learning data. Further, for the feature quantity, that is, for the absolute position x, y of the center point of the destination area candidate, x
The minimum value xs, maximum value xl, and minimum value ys, maximum value yl of the possible values of y are appropriately determined in advance. For example, xs is 0, x is the maximum value of the width of the postal image of m pieces of learning data (xl is the direction of the x-axis in FIG. 5), ys is 0, and m postal images of the learning data are yl as yl. (The vertical is the direction of the y-axis in FIG. 5). As the width and the height, pixel values are used in the same manner as the characteristic values x and y. And an integer k
Xs + (ix (xl-xs)) / k to xs +
((I + 1) × (xl−xs)) / k and ys +
(J × (yl−ys)) / k〜ys + ((j + 1) ×
(Yl-ys)) / k (i = 0, 1, 2,..., K-1,
Consider all combinations of j = 0, 1, 2,..., k−1). In the present embodiment, k is 1
A total of 100 ranges is used as 0. In the present embodiment, ranges are set at equal intervals for each feature amount, but this is not an essential problem.

As a specific method, first, m postal images are input (step D1 in FIG. 13). Next, 0 is substituted for the address frequency hp (A), the sending frequency hq (A), and the pattern frequency hr (A) for the range A of all the feature amounts, and the frequency is initialized (step D2). Further, it is determined whether or not the processing of step D5 has been executed for all m postal images (step D3).
One unprocessed postal image is selected from the individual postal images (step D4). The order does not matter. Further, for the selected postal image, the postal image is input and the destination area candidate generating means 21 generates n destination area candidates (step D).
5). Next, the feature amount extracting means 2 for the address area candidate
1, the feature quantity, that is, the absolute position x, y of the center point of the destination area candidate is calculated (step D6).

Here, for a certain destination area candidate, it is assumed that the feature amount of the destination area belongs to the feature amount range A and whether the destination area candidate is the destination area, the sending area, or the pattern area. Is manually determined, and if the destination area candidate is the destination area, the value of hp (A) is increased by 1. If the destination area candidate is the sending area, the value of hq (A) is increased by 1, and the address area is increased. If the candidate is a pattern area, hr
A process of increasing the value of (A) by 1 is performed (step D7). In step D7, the frequency is updated by performing the above-described processing on the destination area candidates of the postal image. Hereinafter, the processing of steps D3 to D7 is repeated, and the same processing is performed for all postal images. When the processing is completed for all postal images, the addressing probability p (A), the sending probability q (A), and the pattern probability r for the range of the feature amount are calculated using Expressions (8), (9), and (10). (A) is calculated (step D8).

P (A) = hp (A) / (hp (A) + hq (A) + hr (A)) (8)

Q (A) = hq (A) / (hp (A) + hq (A) + hr (A)) (9)

R (A) = hr (A) / (hp (A) + hq (A) + hr (A)) (10) In this way, the feature amount and each probability are obtained, and the range A of all probabilities is obtained. And p (A), q (A), r (A) are stored in the address / submission distribution dictionary 31 in association with each other.

FIG. 14 is a block diagram showing the configuration of the second embodiment of the present invention. In the present embodiment, an address probability conversion dictionary 32 is provided in the storage device 3 of the first embodiment,
An address probability conversion means 26 is provided in the data processing device 2. The address probability converting means 26 calculates the address probability, the sending probability, and the pattern probability obtained by the address / sending probability calculating means 23,
This is a means for correcting to another addressing probability, sending probability and pattern probability based on the addressing probability, sending probability and pattern probability. The address probability conversion dictionary 32 is a dictionary for storing information necessary for converting the old address probability and the like into the new address probability and the like by the address probability conversion means 26.

The addressing probability and the sending probability of each addressing area candidate in the first embodiment are calculated independently for each area, and it is assumed that there is only one addressing area among the addressing area candidates. Not reflected. In the present embodiment, this point is improved, and the detection accuracy of the destination area is further improved. First, the necessary terms are defined below. Based on the feature value obtained by the feature value extracting means 22 for a certain destination area candidate, the probability that the destination area candidate is the destination area among all the destination area candidates is determined by the probability of the target destination area candidate. Define again as address probability. Further, under the condition that a destination area candidate other than the destination area candidate is the destination area, the probability that the destination area candidate is the sending area is defined again as the sending probability. Further, the probability that the destination area candidate is a pattern area is defined as a pattern probability under the condition that the address area other than the destination area candidate is a destination area.

The probabilities that the destination area candidates are the sending area and the pattern area are defined as (1-p ') where p' is the destination probability defined here, q 'is the sending probability, and r' is the pattern probability.
q 'and (1-p') r '. Hereinafter, in order to avoid confusion, the probability obtained by the address / submission probability calculation means 23 is referred to as an old probability, and the probability obtained by the address probability conversion means 26 is referred to as a new probability. For example, when the address area candidates of G1, G2, G3, G4, and G5 as shown in FIG. 4 are obtained, the new address probability p ', the new delivery probability q', and the new pattern probability r 'are shown in FIG. It becomes street. In FIG.
Among the address area candidates G1, G2, G3, G4, G5, for example, the new address probability of G1 is 0.04, and the new address probability of G2 is 0.08.

The address probability conversion means 26 calculates the old address probability p, the old address probability q, and the old address probability p obtained by the address / delivery probability calculation means 23.
The old pattern probability r is converted into a new address probability p ′, a new submission probability q ′, and a new pattern probability r ′ based on the destination probability conversion dictionary 32. The address probability conversion means 26 calculates a new address probability p 'from the old address probability p obtained by the address / delivery probability calculation means 23 using, for example, the following equation (11).

P '= p / S (11) where S is the total sum of the old address probabilities for all the address area candidates in all the address area candidates.

Alternatively, for example, a conversion table of the range of the old address probability p of the address area candidate, the range of the maximum value pmax of the old address probability of the address area candidates other than the address area candidate, and the new address probability p 'of the address area candidate is stored in the conversion table. It is prepared in the address probability conversion dictionary 32. The address probability conversion dictionary 32 at this time is, for example, as shown in FIG.
It can be expressed as follows. However, since the data amount is large, only a part is shown (the part indicated by... Is omitted).
Now, the range of the probability is represented as A, and p and p in the range are
The new address probability of the address area candidate having max is represented by p '(A), the new sending probability is represented by q' (A), and the new pattern probability is represented by r '(A). Also, the function hp for each range A
(A), hq (A), and hr (A) are prepared, and are respectively referred to as a frequency of a destination area, a frequency of a sending area, and a frequency of a pattern area.

FIG. 17 is a flowchart showing a method of creating the address probability conversion dictionary 32. First, m mail images are prepared in advance as learning data. Also, the probabilities p and pma
For x, the minimum value as, the maximum value al, of the possible values of p,
The minimum value bs and the maximum value bl of the possible values of pmax are appropriately determined in advance. For example, 0 for as, 1 for al,
0 is set as bs and 1 is set as bl. Then, for a certain integer k, the feature amount p is as + (i × (a
l−as)) / k〜as + ((i + 1) × (al−as)
)) / K and the feature value pmax is bs + (j ×
(Bl-bs)) / k-bs + ((j + 1) .times. (Bl-
bs)) / k (i = 0, 1, 2,..., k-1, j = 0,
Consider all combinations of 1, 2, ..., k-1)
Consider two ranges. In the present embodiment, k is set to 20 and a total range of 400 is used. In the present embodiment, ranges are set at equal intervals for each feature amount, but this is not an essential problem.

As a specific method, first, m postal images are input (step E1 in FIG. 17). Next, 0 is substituted into the address frequency hp (A), the sending frequency hq (A), and the pattern frequency hr (A) for the range A of all the feature amounts, and the frequency is initialized (step E2). Further, it is determined whether or not the processing has been performed on all of the m postal images (step E3), and if there is an unprocessed one, one unprocessed one is selected from the m postal images. (Step E4). The order does not matter. Next, n destination area candidates are generated for the selected postal image by the destination area candidate generating means 21 (step E5). Further, the feature quantity extracting means 21 calculates the feature quantity of each of the destination area candidates, that is, the absolute positions x and y of the center points of the destination area candidates for the n destination area candidates (step E6). Further, the old address probabilities, old sending probabilities, and old pattern probabilities are calculated by the address / probability calculating means 23 using the feature amounts obtained in step D5 for the n address area candidates (step E7).

Here, for a certain destination area candidate, it is assumed that the old destination probability of the destination area and the maximum value of the old destination probability of the destination area belong to the probability range A. It is manually determined whether the destination area is a destination area or a pattern area. If the destination area candidate is a destination area, the value of hp (A) is increased by one. If the destination area candidate is a destination area, hq (A) is used. Is increased by one, and if it is a pattern area, a process of increasing the value of hr (A) by 1 is performed (step E8). In step E8, the frequency is updated by performing the above-described processing on the n destination area candidates of the postal image. Hereinafter, the processing of steps E3 to E8 is repeated, and the same processing is performed for all postal images.

When the processing is completed for all postal images, the new address probability p 'for the range of the feature amount is calculated using the following equations (12), (13), and (14).
(A), new sending probability q '(A), new pattern probability r' (A)
Is calculated (step E9).

P ′ (A) = hp (A) / (hp (A) + hq (A) + hr (A)) (12)

Q ′ (A) = hq (A) / (hp (A) + hq (A) + hr (A)) (13)

R ′ (A) = hr (A) / (hp (A) + hq (A) + hr (A)) (14) In this way, a new probability is obtained, and all the probability ranges A and p '(A), q' (A), and r '(A) are stored in the address probability conversion dictionary 32 in association with each other. FIG. 16 is a table for converting the address probabilities p 'based on the values of the old address probabilities p of the two address area candidates. However, the number of the address area candidates based on the table may of course be larger. Alternatively, not only the old addressing probability but also the old sending probability and the old pattern probability may be used.

Next, for each destination area candidate, the new submission probability q 'and the new pattern probability r' for the destination area candidate are expressed by the following equation (15), where the old transmission probability is q and the old pattern probability is r. Calculate using 16).

Q '= q / (q + r) (15)

R '= r / (q + r) (16) As an example, when the old address probabilities p, old address probabilities q, and old pattern probabilities r are obtained as shown in FIG. ,
Expressions (11) and (15) in the address probability conversion means 26
Then, the new address probability p ', the new sending probability q', and the new pattern probability r 'are obtained as shown in FIG. Thereafter, the new address probability, the new sending probability, and the new pattern probability are regarded as the address probability, the sending probability, and the pattern probability, respectively, and the same processing as in the first embodiment is performed. For example,
In the case of the example of FIG. 15, when the score is calculated using the mathematical expression (1), a result as shown in FIG. 18 is obtained, and the destination area candidate G
A score of 0.07 is the highest score. Where the function f,
g is f (u) = u and g (u) = u for the argument u.

[0058]

As described above, according to the present invention, when a destination area is detected from a destination area candidate of a postal matter, the destination area having a structure of an address which is not an address such as an address of a sender is detected. Is newly taken into account and the output of the sender area is made difficult by subtracting it from the address probability, which is an address, so that an area that is not an address but has an address structure is erroneously detected as an address area. The risk can be reduced, and the detection accuracy of the address area can be increased.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a first exemplary embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of a postal image.

FIG. 3 is a diagram when a plurality of address area candidates are generated from the postal image of FIG. 2;

FIG. 4 is a diagram showing the address area candidates in FIG. 3 excluding an image.

FIG. 5 is a diagram for explaining a feature amount of a feature amount extraction unit.

FIG. 6 is a diagram illustrating an example of a feature amount for each destination area candidate by a feature amount extracting unit.

FIG. 7 is a diagram illustrating an example of each probability for each destination area candidate by a destination / submission probability calculation unit.

FIG. 8 is a diagram showing a specific example of a conversion table of an address / submission distribution dictionary.

FIG. 9 is a diagram illustrating an example of a score result obtained by an order determining unit.

FIG. 10 is a flowchart illustrating a flow of a process of a feature amount extracting unit.

FIG. 11 is a flowchart showing a flow of processing of an order determining unit.

FIG. 12 is a flowchart showing another processing flow of the order determining means.

FIG. 13 is a flowchart illustrating a method of creating an address / submission distribution dictionary.

FIG. 14 is a block diagram illustrating a configuration of a second exemplary embodiment of the present invention.

FIG. 15 is a diagram showing an example of a new probability for each destination area candidate by the destination probability conversion means.

FIG. 16 is a diagram showing an example of an address probability conversion dictionary.

FIG. 17 is a foot chart showing a method of creating an address probability conversion dictionary.

FIG. 18 is a diagram illustrating an example of a score result for each destination area candidate by an order determining unit.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input device 2 Data processing device 3 Storage device 4 Output device 21 Address region candidate generation means 22 Feature amount extraction means 23 Address / submission probability calculation means 24 Order determination means 25 Output candidate determination means 26 Address probability conversion means 31 Address / submission distribution Dictionary 32 Address probability conversion dictionary

Claims (7)

[Claims] 1. A means for generating a plurality of address area candidates including an address area which is an area of address information based on a partial image of a postal image, and the address area is an address area for each generated address area candidate. Means for extracting a feature quantity representing an index when evaluating whether or not the feature quantity and the address probability indicating the probability of being the address area of the address area candidate and the probability of an area having an address structure other than the address of the address area candidate. Means for comparing and storing the sending probability indicating the probability of the area having no address structure other than the address in the destination and the address, and storing the feature amount of each address area candidate extracted by the extracting means. Means for referring to the storage means to convert to the addressing probability, the sending probability and the pattern probability; and calculating the score for each addressing area candidate using the addressing probability and the sending probability among the converted probabilities. Based on Address area detecting device characterized by comprising a means for selecting a predetermined number of address area candidate from among the address area candidates based on have means for determining the order, was determined order. 2. The method according to claim 1, wherein the order determining means sets a destination probability for each destination area candidate to p and a sending probability to each destination area candidate.
q, two kinds of increasing functions f (x) with x given in advance as an argument
And g (x), the score for the address area candidate
2. The address area detection device according to claim 1, wherein the address area is calculated as f (p) -g (q), and the order of the address area candidates is determined in descending order of the score. 3. The order deciding means, for each destination area candidate, p, a destination probability for the destination area candidate, and
q, two kinds of increasing functions f (x) with x given in advance as an argument
And g (x), the score for the address area candidate
2. The destination area detection device according to claim 1, wherein the destination area detection apparatus calculates as f (p) -g (q / (p + q)) and determines the order of the destination area candidates in descending order of the score. 4. The method according to claim 1, wherein the order determining means sets, for each destination area candidate, a destination probability to the destination area candidate and a sending probability to the destination area candidate.
q, two kinds of increasing functions f (x) with x given in advance as an argument
And g (x), the score for the address area candidate
2. The destination area detecting device according to claim 1, wherein the destination area candidate is calculated as f (p) / g (q), and the order of the destination area candidates is determined in descending order of the score. 5. The order deciding means, for each destination area candidate, a destination probability for the destination area candidate, and a sending probability for the destination area candidate.
q, two kinds of increasing functions f (x) with x given in advance as an argument
And g (x), the score for the address area candidate
2. The destination area detecting device according to claim 1, wherein the destination area is calculated as f (p) / g (q / (p + q)), and the order of the destination area candidates is determined in descending order of the score. 6. The order deciding means, for each destination area candidate, a destination probability for the destination area candidate, and a sending probability for the destination area candidate.
q, when a constant given in advance is c, the score for the destination area candidate is calculated as p-cq, and the order of the destination area candidates is determined in descending order of the score.
2. The address area detection device according to 1. 7. The order deciding means, for each destination area candidate, a destination probability for the destination area candidate, and a sending probability for the destination area candidate.
q, when a constant given in advance is c, the score for the destination area candidate is calculated as p-cq / (p + q), and the order of the destination area candidates is determined in descending order of the score. 2. The address area detection device according to claim 1, wherein: