JP2002312719A - Character segmentation device, character segmentation method and program for character segmentation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character segmentation device and a character segmentation method, by which characters are segmented correctly without the prerequisite of row extraction even in the case that a plurality of rows are inputted simultaneously. SOLUTION: The character segmentation device is provided with a character string image storage part 11 for storing character string images, a connection component extraction part 12 for binarizing the character string images and extracting connection components, a minimum unit preparation part 13 for judging whether or not they are in contact for all the connection components, turning the connection component itself to a minimum unit at the time of non- contact and cutting the connection component and newly preparing the minimum unit at the time of contact, a two-dimensional connection relation preparation part 14 for obtaining the connection relation of the minimum units with each other, a character segmentation candidate output part 15 for preparing a combination pattern by a plurality of the minimum units in the connection relation with each other and outputting the pattern with the high possibility of being a correct character as a character segmentation candidate, and a character segmentation establishment part 16 for performing a recognition processing, a language knowledge processing and layout analysis on the basis of the character segmentation candidate and outputting a character segmentation result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus and a method for extracting a character from an optically scanned character string image, and more particularly to a character extraction apparatus and a character for extracting a character without assuming extraction of a character line. The present invention relates to a clipping method and a character clipping program.

[0002]

2. Description of the Related Art When a character string written on a mail or a form is read by an optical character reading device (OCR) and character recognition is performed, preprocessing is performed prior to recognition processing. Lines are extracted line by line, and then several character cutout candidates that are candidates for cutting out one character are obtained. Then, a recognition process is performed on the character cutout candidate, and a linguistic knowledge process is performed to obtain a character cutout result and a character string reading result.

[0003]

Conventionally, in generating such a character segmentation candidate, it has been necessary to extract lines one by one from a character string pattern input first.
In other words, in the conventional technology, it is assumed that lines are already extracted line by line, so if multiple lines are input at the same time, if it fails to cut out one line at a time, it is difficult to generate character extraction candidates correctly. Met.

[0004] For example, the technique for creating a character segmentation candidate disclosed in Japanese Patent Application No. 7-8793 is based on the premise that characters are correctly input line by line.
Further, in the technologies disclosed in Japanese Patent Application Nos. 8-52844, 11-30686 and 11-57792, the process of extracting character lines (character strings) is an initial stage. It is given in. With these techniques, it is difficult to correctly generate a character cutout candidate if it fails to cut out one line at a time from the input pattern using any of the methods.

On the other hand, Japanese Patent Application No. 7-236232 discloses a technique for accurately extracting characters one by one even if a plurality of character strings meander. A process for extracting a character string is performed in the processing at the stage, and if extraction of the character string fails, it is also difficult to correctly generate a character extraction candidate.

For example, when a character string as shown in FIG. 2A is input, it is not easy to correctly determine whether the character string is composed of one line or two lines. ), The first line is `` Tsukitsukicho '' and the second line is `` Na Hita 3 '', and character extraction candidates that can be correctly recognized as `` Ariake Machi 3 '' It is impossible to generate and similarly difficult to read.

On the other hand, Japanese Patent Application No. 7-203148 discloses a method capable of correctly reading a character string without assuming a correct line-by-line input. Sometimes, when a correct candidate is not generated, it is impossible to read a character string correctly. Also, according to this method, creation of character candidates is
Since all combinations of connected components of black pixels are considered, explosion of combinations may occur and processing may be broken, and it is difficult to cut out characters correctly when a touched character pattern is included There was a problem of becoming.

Therefore, from the viewpoint of improving the reading accuracy of a character string, even when a plurality of lines are input at the same time, without assuming that line extraction is performed, the combination does not explode and the contacted character pattern is correctly detected. It can be said that a device that can be cut and can perform correct character segmentation is required.

The present invention has been made in view of the above problems, and a main object of the present invention is that even when a plurality of lines are inputted at the same time, the combination explodes without assuming that line extraction is performed. It is another object of the present invention to provide a character cutout device, a character cutout method, and a character cutout program that can cut a contacted character pattern correctly and perform correct character cutout.

[0010]

In order to achieve the above object, a character cutout device according to the present invention is a character cutout device for generating a character cutout candidate from a character string image. An image storage unit, a connected component extraction unit that binarizes the character string image to extract connected components, and determines whether or not all connected components are in contact with each other, and determines whether the connected components are not in contact with each other. A minimum unit creation unit that creates the minimum unit by cutting the connected component when it is determined that the connected component is in contact with the connected component itself as a minimum unit when A two-dimensional connection relationship creating unit that determines a two-dimensional connection relationship between the minimum units arranged in a state, and creates a combination pattern by combining a plurality of the minimum units in a connection relationship with each other, A character segmentation candidate output unit for outputting a likely combination pattern is Shii characters as character segmentation candidates, the one having at least.

Further, according to the character extracting apparatus of the present invention, a character extracting apparatus for generating a character extracting candidate from a character string image is provided. A connected component extraction unit that extracts a connected component by converting the value into a value and determines whether or not all connected components are in contact with each other. When it is determined that the connected components are not in contact with each other, the connected component itself is a minimum unit. And a minimum unit creating unit that creates a new minimum unit by cutting the connected component when it is determined that the connected component is in contact,
A two-dimensional connection relationship creating unit that determines a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space, and a combination pattern created by combining a plurality of the minimum units having a connection relationship with each other. A character cutout candidate output unit that outputs a combination pattern likely to be a correct character as a character cutout candidate, and performs a recognition process, a linguistic knowledge process, and a layout analysis based on the character cutout candidate, and outputs a character cutout result And a character cutout confirming unit.

In the present invention, the minimum unit creation unit refers to the average value of the width and height of all connected components extracted from the character string image and the width and height of the connected component of interest, It can be configured to be statistically determined whether or not the connected components that are in contact, further, with reference to the image pattern representing the connected component shape of interest, whether the connected components that are in contact It is also possible to adopt a configuration that is determined statistically.

In the present invention, in the two-dimensional connection relation creation unit, when obtaining the adjacent relation of the minimum unit extracted from the character string image, the minimum unit sharing a Voronoi side by Voronoi division is adjacent to the minimum unit. When it is determined that there is a relationship, and when the adjacent minimum units share two Voronoi sides, it is preferable that it is determined that the minimum units are not adjacent.

Also, in the present invention, the two-dimensional connection relationship creating unit may include a width, height, and center coordinate value of each of the plurality of minimum units determined to be adjacent to each other, and the plurality of minimum units. By referring to the width, height and center coordinate values of the circumscribed rectangle created by combining and the average value of the width and height of all the minimum units extracted from the character string image, statistically evaluated The connection relationship may be determined, and further, the connection relationship may be determined with reference to the image patterns representing the shapes of the plurality of minimum units.

Further, in the present invention, the character segmentation candidate output unit traces the connection relationship obtained by the two-dimensional connection relationship creation unit when creating the combination pattern, thereby combining a plurality of minimum units. Preferably, a pattern is created.

Also, in the present invention, the character segmentation candidate output section may include a width of the combination pattern,
With reference to the height and the center coordinate value, and the average value of the width and height of all the minimum units extracted from the character string image, or
Width, height and center coordinate value of the combination pattern,
From the width, height and center coordinate values of the minimum unit adjacent to the combination pattern, the circumscribed rectangle width, height and center coordinate value created by combining the combination pattern and the minimum unit, and from the character string image With reference to the average value of the width and height of all the extracted minimum units, statistical evaluation is performed for the same number of times as the number of the minimum units adjacent to the combination pattern, and the combination pattern is used as a character cutout candidate. Whether or not to output can be determined, further, with reference to an image pattern representing the shape of the combination pattern and an image pattern representing the shape of the minimum unit adjacent to the combination pattern,
A configuration may be adopted in which it is determined whether to output the combination pattern as a character segmentation candidate.

A character extracting method according to the present invention is a character extracting method for generating a character extracting candidate from a character string image, wherein a character string image storing step of inputting and storing a character string image, and binarizing the character string image. A connected component extraction step of extracting connected components, and determining whether or not all connected components are in contact, and when it is determined that the connected components are not in contact, the connected component itself is used as a minimum unit, A minimum unit creation step of cutting the connected component to create a new minimum unit when it is determined that the connected component is in contact with the connected component, and a two-dimensional configuration of the minimum units arranged in a two-dimensional space. Two-dimensional connection relationship creating step to determine a typical connection relationship, to create a combination pattern by combining a plurality of the minimum units in a connection relationship with each other,
A character extraction candidate output step of outputting a combination pattern having a high possibility of being a correct character as a character extraction candidate.

The character extracting method according to the present invention is a character extracting method for generating a character extracting candidate from a character string image, wherein a character string image storing step of inputting and storing a character string image is performed. A connected component extracting step of extracting connected components by converting the value to a value, and determining whether or not all the connected components are in contact. When it is determined that the connected components are not in contact, the connected component itself is the minimum unit. And a minimum unit creation step of cutting the connected component to create a new minimum unit when it is determined that the connected components are in contact with each other, and the minimum unit between the minimum units arranged in a two-dimensional space. A two-dimensional connection relationship creation step for obtaining a two-dimensional connection relationship, and combining the plurality of minimum units having a connection relationship with each other to create a combination pattern, A character cutout candidate output step of outputting a combination pattern having a high possibility as a character cutout candidate, and a character cutout that performs recognition processing, linguistic knowledge processing, and layout analysis based on the character cutout candidate and outputs a character cutout result And at least steps.

A program according to the present invention includes a character string image storing process for inputting and storing a character string image, a connected component extracting process for binarizing the character string image and extracting a connected component, It is determined whether or not the connected component is in contact, and when it is determined that the connected component is not in contact, the connected component itself is used as a minimum unit, and when it is determined that the connected component is in contact, the connected component is determined. A minimum unit creation process for creating a new minimum unit by cutting off, a two-dimensional connection relationship creation process for finding a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space, and connecting with each other A character cutout candidate output process that creates a combination pattern by combining a plurality of the minimum units having a relationship and outputs a combination pattern that is likely to be a correct character as a character cutout candidate If is intended to execute on a computer.

Further, the program of the present invention includes a character string image storing process for inputting and storing a character string image, a connected component extracting process for binarizing the character string image and extracting a connected component, and It is determined whether or not the connected component is in contact, and when it is determined that the connected component is not in contact, the connected component itself is used as the minimum unit, and when it is determined that the connected component is in contact, the A minimum unit creation process of cutting a connected component to create a new minimum unit, and a two-dimensional connection relationship creation process of finding a two-dimensional connection relationship between the minimum units in a state arranged in a two-dimensional space, A character cutout candidate that creates a combination pattern by combining a plurality of the minimum units connected to each other and outputs a combination pattern that is likely to be a correct character as a character cutout candidate And power processing, based on the character segmentation candidate performs the recognition processing and language knowledge processing and layout analysis, in which to execute the character cut-out determination processing of outputting the character extraction result to the computer.

[0021]

Next, an embodiment 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 character segmenting device according to an embodiment of the present invention. The configuration of the embodiment shown in FIG.
A character string image storage unit 11 that inputs and stores a character string image obtained by a character string image reading device 18 that reads a character string image using optical means, and binarizes the character string image to extract connected components. The connected component extraction unit 12 determines whether or not all connected components are in contact. When it is determined that the connected components are not in contact, the connected component itself is used as the minimum unit, and the connected components are contacted. The minimum unit creation unit 13 that cuts the connected component to create a new minimum unit when it is determined that the minimum unit is present, and obtains a two-dimensional connection relationship between the minimum units arranged in the two-dimensional space. A combination pattern is created by combining the two-dimensional connection relationship creation unit 14 with a plurality of minimum units having a connection relationship with each other, and a combination pattern that is likely to be a correct character is a character cutout candidate. A character cutout candidate output unit 15 to and outputting, based on the character segmentation candidate performs recognition processing and language knowledge processing and layout analysis, and character segmentation determination unit 16 for outputting a character segmentation results, but the computer 1
7, a display means 19 for displaying the result is connected to the character cutout determination section 16, and the computer 17 is provided with the above means 11-16.
Is stored as a program.

Next, the operation of this embodiment will be described. Basically, the present invention converts the input character string pattern into
This is a method in which a character segmentation candidate is determined by first dividing the pattern into fine partial patterns and then integrating the partial patterns again, and in that respect, there is no great difference from a conventionally known method. However, a significant difference from the conventionally known method is that the present invention does not assume that the input has already been correctly cut out line by line.
Given such a character segmentation candidate, if a method of reading a character string using simultaneously the character recognition cost, linguistic cost, and layout cost described in Japanese Patent Application No. 7-203148, a plurality of lines are input simultaneously. Even if it is done, without assuming that line extraction is performed, the character pattern that touched can be cut correctly without explosion of the combination,
Correct character segmentation can be performed.

The character string image storage section 11 inputs and stores, for example, a character string image scanned optically. This can be easily realized using a normal scanner and memory. The connected component extraction unit 12 inputs a character string image, binarizes the image, and extracts a connected component for character extraction.

There are several methods for realizing this processing. For example, binarization can be easily realized by using binarization based on Otsu's discriminant analysis. For the extraction of connected components, for example, extraction of connected components can be easily realized by normal labeling processing and contour tracking processing.In addition, contour tracking is performed after extracting connected components by normal labeling processing. Thus, the chain code of the outline of the connected component can be simultaneously extracted. Further, the connected component itself can be stored as an image pattern, and a circumscribed rectangle of the connected component can be stored. To explain with reference to the drawing, when a binarized pattern as shown in FIG. 2A is input, eight connected components (CC1 to CC
Up to 8) are extracted. In the example of FIG. 3, the circumscribed rectangle of each connected component is also displayed for easy viewing, and the order of the connected components is determined in descending order of the center coordinates of the circumscribed rectangle.

A pattern as shown in FIG. 2A is input, and the data of the connected component obtained by the connected component extraction unit 12 includes an ID which is a serial number of the connected component and a circumscribed rectangle as shown in FIG. The width and height (W _i , H _i ) and the coordinates (X _i , Y
_i ), the connected component image itself x _i , the chain code tracking start point coordinates (C _xi , C _yi ), the chain code length CL _i, and the chain code C _i are output as connected component information. Here, the subscript i means the i-th connected component, and corresponds to the ID number. Further, in FIG. 4, for simplicity, the image pattern x _i is a multi-dimensional vector information, the description of the contents of the chain code C _i is omitted. The data length of the image pattern can be known from the width and height of the circumscribed rectangle.

Now, the connected component obtained by the connected component extraction unit 12 may be extracted as one connected component due to contact between a plurality of characters as in the example of CC4 in FIG. In other words, when the character of the input character string image is touched by the connected component extraction unit 12, the touched part cannot be separated. Therefore, the connected component touched by the minimum unit creation unit 13 is separated. The processing of the minimum unit creation unit 13 is roughly divided into two processings. The first is the detection of the connected component that is making contact, and the second is the disconnection of the contact point.

In order to separate the connected connected components,
First detect which connected components are in contact.
I have to. There are several ways to detect this.
However, for example, out of all the connected components obtained from the input image,
Consider the tangent rectangle and calculate the average of the width and height of all circumscribed rectangles as (W
_Cmean, H_Cmean).
The width and height of the component (W_i, H_i) To detect contact
It can be carried out. Where the subscript i is the ith connected component
Means that The connected components of interest are in contact
Whether the contact detection feature
(W_Cmean, H_Cmean, W _i, H_iConsider the four-dimensional vector
It is possible to use this method to perform statistical learning.
You. Or (W_i/ W_Cmean, H_i/ H_Cmean) Two-dimensional vector
Can also be used.

For example, a character string image for learning is prepared, and (W _Cmean , H _Cmean ) and (W _i , H _i ) of all connected components are obtained in advance, and the teacher of the connected components that are in contact with each other is obtained. Data
T _i = 1, can be learned every teacher data of connected components otherwise the T _i = 0. For example, when performing regression analysis as statistical learning, it can be considered as a problem of estimating the teacher data Ti using (W _Cmean , H _Cmean , W _i , H _i ) as an explanatory variable. Similarly, it can be considered as a problem of estimating the teacher data Ti using (W _i / W _{C mean} , H _i / H _Cmean ) as an explanatory variable. In other words, a contact detection feature may be given to the explanatory variable, and teacher data corresponding to the feature may be estimated.

When using a multilayer perceptron type neural network learning by back propagation,
Four-dimensional vectors to the input layer _{(W Cmean, H Cmean, W} i, H i) gave, may be performed learning gives Ti as teacher data. At this time, the structure of the neural network is as follows.
It is also possible to have a structure with 5 neurons in the hidden layer, 1 neuron in the output layer, and freely change the number of neurons from 2 to 1000 in the hidden layer. It is also possible. Of course, a two-dimensional vector (W _i / W _Cmean , H _i / H _Cmean ) can be given to the input. In this case, the input layer becomes two neurons. That is, since the contact detection feature may be provided to the input layer of the neural network, the same number of neurons as the number of dimensions of the contact detection feature may be provided to the input layer.

Further, when using the discriminant analysis, considered T _i are _{category, (W Cmean, H Cmean,} W i, H i) it is also possible to perform the discrimination analysis using, (W _i / W _Cmean , H _i / H
_Cmean ) can also be used. In other words, the contact detection feature,
Discriminant analysis may be performed by giving a category corresponding to the feature.

As described above, it is determined using the learned neural network, the regression coefficient, and the discriminant axis that the connected component having a value larger than the predetermined threshold value is in contact with the connected component. Is possible. That is, when using a trained neural network, (W
_Cmean , H _Cmean , W _i , H _i )
When the output is 0.5 or more, it can be determined that contact has occurred, and a similar determination can be made by multiple regression analysis.

Similarly, a two-dimensional vector of (W _i / W _Cmean , H _i / H _Cmean ) is input, and if the output is 0.5 or more, it can be determined that contact has occurred. A similar determination can be made in multiple regression analysis. At this time, the threshold value is not limited to 0.5, and a value from 0 to 1 can be freely set. In the case of discriminant analysis, (W _Cmean , H _C
It is possible to use a discrimination score for _mean , W _i , H _i ), and it is possible to determine that contact has occurred when a discrimination score equal to or greater than a predetermined threshold is obtained.
Of course, (W _i / W _Cmean , H _i / H _Cmean ) can be used for the determination.

As another method, for example, (W _Cmean , H _Cmean , W
_i , H _i ), the four-dimensional space where the data exists is F, and considering the event of Ti = 1 and the event of Ti = 0, the probability density functions p (F | T _i = 1) and p
(F | T _i = 0), and the value of this probability density function p (F | T _i = 1) gives a value equal to or greater than a certain threshold (W _Cmean , H _Cmean , W _i ,
H _i ), it is possible to determine that the i-th connected component is in contact, and the prior probability P (T _i = 1) of the event of T _i = 1 and the event of T _i = 0 ) and obtained in advance the P (T _i = 0), I asked the posterior probability of an event of T _i = 1 of the event and T _i = 0 from Bayes' theorem, the larger the posterior probability of T _i = 1 event It is possible to determine that the i-th connected component is in contact, and T _i = 1
The posterior probability of the event is a predetermined threshold (0 to 1
If the value is larger than (a value set during), it is possible to determine that the i-th connected component is in contact. Of course, when this can be used _{(W i / W Cmean, H} i / H Cme an). Here, a four-dimensional space or a two-dimensional space has been described.
As F, the above-described method can be used.

At this time, the probability density function can be obtained by, for example, an EM algorithm using a mixture normal distribution, and the prior probability of each event is determined in advance by the ratio of how much the contact pattern is included in the learning data. Can be easily obtained by obtaining.

Further, as a contact detection feature, (W _Cmean , H
_Cmean, W _i, H _i) and _{(W i / W Cmean, H} i / H Cmea n) as well as image information of the connected components also it is considered to use added.
In the case of using the four-dimensional data or the two-dimensional data described so far, the connected component is treated as rectangular information, and it is difficult to know what pattern the image pattern inside is. By using image information, it is possible to know the shape pattern of the connected component itself. As this image information, for example, a 16 × 16
, And can be used as it is as a 16 × 16-dimensional vector. Using the image information created in this way, the contact detection feature is 4D + 16 × 16
It is a two-dimensional vector or a two-dimensional + 16 × 16-dimensional vector.

Also, the principal component analysis is performed by normalizing the images of the connected components as many learning samples to a size of 16 × 16, and ten eigenvectors are used in descending order of the eigenvalues. The value of the inner product can also be used. In this case, the image information is a 10-dimensional vector because the inner product of the image information with 10 eigenvectors is used. Here, it has been described that ten eigenvectors are used in the descending order of eigenvalues, but the number is not limited to ten and can be set freely from one to 256. In the above description, the image is normalized to a size of 16 × 16. However, the image is not limited to this size, and may be normalized to an arbitrary size as N × M as necessary. It is possible. At this time, N and M are 1 to 2 respectively.
It can be set freely up to 56.

For example, when a character is described in the horizontal direction, when the vertical direction is set to N = 1 and the horizontal direction M is set to the same value as the horizontal width of the connected component, this vector becomes black in the writing direction of the character string. This is equivalent to the case where a projection histogram of pixels is used. In other words, it can be said that the present invention is a generalization of a conventional method using a projection histogram of black pixels. However, if M in the horizontal direction is set to be the same as the width of each connected component, the number of dimensions of the contact detection feature calculated from each connected component will be different one by one.
For example, by setting N = 1, M = 64, etc., statistical learning becomes possible.

It is also possible to divide the original image area into small areas, obtain the content ratio of black pixels while overlapping the areas little by little, and use this as image information. The content ratio is a ratio of black pixels occupying in the region. For example, if the size of the small region is 10 × 10 and there are 30 black pixels in the small region, the content ratio of the black pixels is 30/10 × 10 = 0.3. How to divide the small area will be described with reference to FIG.
Divided into small areas. FIG. 5 is an explanatory diagram corresponding to CC1 in the example of FIG. Although CC1 of the circumscribed rectangle is the width and height (W _1, H _1), making small areas by separating this small rectangular of _{(W 1/2, H 1} /2). At this time, this small area slightly overlaps with an adjacent small area,
As shown in the figure, 9 areas from Area1 to Area9 are created.
By counting the number of black pixels in this area and dividing the area by the area of the area, the black pixel occupancy for each small area can be obtained, which can be used as image information. Using the image information created in this way, the contact detection feature is 4D +
It is a 9-dimensional vector or a 2-dimensional + 9-dimensional vector.

As described above, when the image information for the i-th connected component is Img _i , (W
_{Cmean, H Cmean, W i,} H i, it is possible to give the Img _i). Also
_{(W i / W Cmean, H} i / H Cmean, Img i) can be given.

When the connected components that have come into contact are detected in this way, it is necessary to cut these connected components next. Several methods are conceivable for cutting this connected component at the contact point. For example, “Ishidera et al .:“ A method of character segmentation based on stroke extraction ”(Digital Society of Japan, D-348, 1996) And the method proposed in
The method described in JP-A-9-318638 can also be used.

FIG. 6 is a flowchart illustrating this process. In ST11, the connected components of the input image are input, and the average value (W _Cmean , H _Cmean ) of all the connected components is _obtained . S
At T12, attention is paid to the connected components one by one, and a contact detection feature for the i-th connected component is created. In ST13, it is determined whether or not contact has occurred statistically. The connected component that has caused contact is subjected to a cutting process in ST14, and is output as a cutout minimum unit in ST15. In addition, the connected component determined to have not caused a contact in ST13 moves directly to ST15 without performing the processing of ST14,
It is output as the minimum cutout unit. This process is performed for all connected components.

If an example of the operation when the connected component information of FIG. 4 corresponding to the example of FIG. 3 is input is shown, at ST11 (W
_Cmean , H _Cmean ) = (96.3, 107.9), and the contact determination feature is obtained for each connected component in ST12. For the serial numbers 1 to 3 (CC1 to CC3) of the connected components, the outputs of the neural networks are all 0.1 or less, and it is determined that no contact has occurred. However, the output of the neural network for the connected component (CC4) with the serial number 4 is 0.9 or more, the connected component with the serial number 4 is determined to be in contact, and the cutting process is performed. The output of the neural network for the connected components (CC5 to CC8) with serial numbers 5 to 8 becomes 0.1 or less again, and it is determined that no contact has occurred.

It is determined that the connected components CC1, CC2, CC3 and CC5, CC6, CC7, CC8 are not in contact, and the connected component information is output as the minimum unit information as it is. On the other hand, the connected component CC4 is divided into four minimum units PS02, PS05, PS06, and PS08 by the cutting process in ST14 as shown in FIG. At this time, it is necessary to create minimum unit information having the same data format as the connected component information for the four minimum units to be newly generated. In this process, the same process as that in the connected component extraction unit may be performed again using the image obtained by cutting the CC4. FIG. 8 shows an example of the minimum unit information created by the minimum unit creating unit in this way.
In FIG. 8, the minimum unit information PS01, PS03, and PS04
PS07, PS09, PS10, and PS11 are connected component information CC1 and C, respectively.
It corresponds to C2, CC3, CC5, CC6, CC7 and CC8, and the contents are exactly the same. Therefore, the minimum unit information newly created by the cutting process of the connected component CC4 is PS02 and PS05.
PS06 and PS08.

Next, the two-dimensional connection relation creating unit 14 creates a connection relation between the minimum units in the two-dimensional space. The connection relationship and the adjacency relationship are similar, but here, the adjacency relationship simply means that they are graphically adjacent to each other. The connection relationship will be described later in detail.

First, the significance and significance of the adjacency will be described. The conventionally known method of creating a character segmentation candidate is based on the premise that the input character string has already been segmented line by line, and by sorting the minimum unit on the axis in the writing direction of the character string, the adjacency relationship is obtained. It was easy to create. That is, if the minimum units are sorted by the description position, the order is already determined, and it is obvious that the minimum units adjacent to each other can be considered to be adjacent to each other. If the adjacency relationship is known, it is not necessary to separately process apparently unnecessary character cutout candidates, and it is also possible to avoid erroneously outputting apparently unnecessary character cutout candidates.

This will be described with reference to the example of FIG. This example is an example in which a condition is given in advance that the pattern is a horizontally written pattern, that only one line is written, and that the contact point has already been cut. Given a character string pattern and the minimum unit from PS21 to PS27 was obtained by sorting the center position of the circumscribed rectangle in order from the left, for example, skipping PS22 and creating a combination of PS21 and PS23 Can be said to be very rare to represent the correct character. Similarly skip PS22 and PS23 and PS21 and PS
It is very rare that even if you make 24 combinations, this represents the correct one character. As shown in FIG. 9, when a combination is created by skipping adjacent minimum units, the pattern is not considered to represent a correct character. . As described above, it is important to determine the adjacency in advance so as not to create an unnecessary combination.

However, an object of the present invention is to create a character segmentation candidate in a situation where there is no assumption that the input pattern has only one line. That is, it is unknown how many lines the input pattern consists of, and in the present invention, the minimum unit obtained at the time when the processing of the minimum unit creating unit ends is merely a state scattered on the two-dimensional space. Absent.

Now, when trying to define this adjacency in two dimensions, it is impossible to use a simple sort. There are several methods for defining the minimum unit adjacency scattered on two dimensions. For example,
Using Voronoi division, the minimum units sharing Voronoi sides can be defined as having an adjacent relationship.

More specifically, if a chain code is obtained by tracing the outline of each minimum unit, the coordinate values of points on the outline can be easily calculated by using the chain code. This series of coordinate values is called a coordinate value chain code.
A sampling point can be sampled every other and used as a generating point. FIG. 10B shows the result of generating the generating points for the pattern shown in FIG. 10A, and the black circles indicate the generating points on the contour.

These generating points can be generated for all the minimum units, and Voronoi division can be performed on the generating points. Next, Voronoi sides generated from generating points belonging to the same minimum unit are deleted, and finally a Voronoi division result of the minimum units is obtained. This Voronoi division result is an approximation of the Voronoi division of the minimum unit.In order to improve the approximation accuracy, instead of sampling every fifth coordinate value with a coordinate value chain code and using it as a mother point, a coordinate value chain code is used. It is possible to increase the accuracy by reducing the sampling interval between every third and every two samples, and if you want to speed up the processing time even if you lower the accuracy a little, you can increase the sampling interval further It is possible. FIG.
(C) shows an example of the result of Voronoi division of the minimum unit.
In (c) of FIG. 10, Voronoi sides made from the same minimum unit are shown by dotted lines, and Voronoi sides made from different minimum units are shown by thick lines. In addition, for the pattern of the character string image in FIG. 2A, the extraction of the connected components and the cutting of the contact pattern are completed, and Voronoi division is performed from a state in which 11 minimum units (PS01 to PS11) are obtained. The results are shown in FIG. Voronoi sides are indicated by dotted lines.

FIG. 12 is a flowchart for explaining the processing for obtaining the adjacency. In ST21, a generating point is generated on the contour of all minimum units, and Voronoi division is performed by using all generating points generated in ST22 to perform Voronoi division. ST
In 23, focusing on one Voronoi side, if it is determined in ST24 that the Voronoi side is obtained between the generating points of the same minimum unit, the Voronoi side is deleted in ST25, otherwise Moves to ST26. In ST26, if attention is paid to all Voronoi sides, the process proceeds to ST27; otherwise, the process returns to ST23. In ST27, by using the remaining Voronoi sides, the minimum unit pair sharing the same Voronoi side is determined to have an adjacent relation, and the adjacent relation information is output. FIG.
FIG. 13 shows an output example of the adjacency information obtained when the pattern (a) is input.

The adjacency information is a matrix,
When a certain minimum unit i and a minimum unit j are adjacent to each other, the value at the position of the i-th row and the j-th column (same for the j-th row and the i-th column) is 1, and when they are not adjacent, the value is 0.

In general, the Voronoi division often assumes that the vertical and horizontal axes are equidistant, but points on an ellipse can be equidistant as shown in FIG. 14 (a). At this time, as shown in FIG. 14 (b), the value of the abscissa of the generating point is given by multiplying by 0.5, and the value of the ordinate is given as it is, and ordinary Voronoi division (the vertical and horizontal axes are equal) 14A, it is also possible to perform Voronoi division such that the elliptical portions are equidistant as shown in FIG. 14A. Here, the value of the abscissa is multiplied by 0.5 and the value of the ordinate is unchanged (1.
However, this ratio can be freely set, and the horizontal axis may be freely set to x times and the vertical axis to y times.

Next, a determination is made as to whether connection is possible based on the minimum unit adjacent relationship. This connection availability determination focuses on an arbitrary minimum unit pair that is adjacent to each other, and determines whether the pair may be connected.

Here, the significance of the connection relationship will be described. For example, when a pattern as shown in FIG. 15A is given, the minimum units PS101 to PS103 share Voronoi sides and are adjacent to each other. Here, the Voronoi sides shared by the minimum units PS101 and PS103 are two (B1 and B2)
It is important that there is. That is, when a combination is created using a pattern that shares two Voronoi sides like PS101 and PS103 to create a character segmentation candidate, FIG. 15 (b)
, The character segmentation candidate has a Voronoi region closed inside. It is very rare that such a pattern is obtained as a correct character, so the PS101
And PS103 are adjacent, but there is no point in creating a combination of the two. That is, it is necessary not to create a character cutout candidate consisting of only PS101 and PS103. Create a connection to avoid this. That is, if it is determined that patterns sharing two Voronoi sides like PS101 and PS103 are not connected,
It is possible to control not to create a character cutout candidate consisting of only PS101 and PS103.

However, assuming that PS101 and PS102 are both adjacent and connected, and if PS102 and PS103 are also adjacent and connected, PS101 and PS1 are similarly connected.
A combination of 02 and PS103 can be created. In other words, PS101 and PS103 cannot be directly connected, but can be indirectly connected by sharing PS102 when creating a combination.

As described above, when arbitrary pairs of adjacent minimum units share two Voronoi sides, it can be determined that there is no connection. In other words, it is possible to determine that connection should not be made in the connection availability determination.

In order to improve the accuracy of connection determination,
It is possible to use several other methods together
You. As an example of the method, for example, as shown in FIG.
Given the minimum unit of, the circumscribed rectangle information of the minimum unit i
The coordinates of the center point (X_i, Y_i) And outside
The width and height of the tangent rectangle (W_i, H_i) And extract the minimum unit j
Also the center point coordinates (X_j, Y_j) And the width and height (W_j, H_j)
New circumscribed rectangles created by extracting and combining these
P_ijCenter point coordinates (X_ij, Y_ij) And the bounding rectangle width and height
(W_ij, H_ij), And the average width of all the smallest units and
Average height (W_PSmean, H_{PSm ean}), And using these values
Connection judgment feature (X_i, Y_i, W_i, H_i, X_j, Y_j, W_j, H_j, X_ij, Y _ij,
W_ij, H_ij, W_PSmean, H_PSmean) Statistically determines whether connection is possible
Settings can be made. This is a 14-dimensional vector
It is expressed as Or (X_i/ W_PSmean, Y_i/ H_{PS mean}, W_i/ W
_PSmean, H_i/ H_PSmean, X_j/ W_PSmean, Y_j/ H_PSmean, W_j/
W_PSmean, H_j/ H_PSmean, X _ij/ W_PSmean, Y_ij/ H_PSmean, W_ij/ W
_PSmean, H_ij/ H_PSmean) Using a 12-dimensional vector
It is also possible.

For example, a character string image for learning is prepared, and connection determination features are obtained in advance from all pairs of adjacent minimum units so that the minimum unit i and the minimum unit j are included in one character. If it is included, the training data is set to T _ij = 1, otherwise (i.e., if the minimum unit i and j are included in different characters), the training data is set to T _ij = 0 and learning is performed. it can.

When performing, for example, regression analysis as statistical learning, it can be considered as a problem of estimating teacher data _Tij using the connection determination feature as an explanatory variable. When a multilayer perceptron-type neural network that learns by back propagation is used, the learning may be performed by providing a connection determination feature to the input layer and providing _Tij as teacher data.
The structure of the neural network at this time is such that the number of neurons in the input layer is the same as the number of dimensions of the contact determination feature, with five neurons in the intermediate layer and one neuron in the output layer. The number of neurons can be freely changed from 2 to 1000 in the intermediate layer. When discriminant analysis is used, _Tij is considered to be a category, and discriminant analysis can be performed using the connection judgment feature. It is also conceivable to use principal component analysis, Mahalanobis distance, or the like.

As described above, it is determined that a connection can be made to a minimum unit pair having a value larger than a predetermined threshold value using a learned neural network, a regression coefficient, and a discriminant axis. It is possible. In other words, when a learned neural network is used, a connection determination feature vector is input, and if the output is 0.5 or more, it can be determined that the pair is connectable, and the same applies to multiple regression analysis. Can be determined. In the case of discriminant analysis, it is possible to use a discrimination score for the connection determination feature, and it is possible to determine that contact has occurred when a discrimination score equal to or greater than a predetermined threshold is obtained.

The significance of performing this processing will be described. For example,
When pairs of adjacent minimum units, such as PS05 and PS07 in FIG. 11, are positioned diagonally far from each other, it is not possible that these pairs constitute the same character as each other. It is rare. On the other hand, PS03 and PS05 in FIG.
In addition, PS05 and PS06 are located in the horizontal and vertical directions close to each other, and it can be said that in such a pair, there is a high possibility that the same character is constituted by each other.

When the adjacency relation is defined by Voronoi division as in the present invention, the number of adjacent pairs may increase. Therefore, it is necessary to obtain a more precise connection relation by the above method. Is high. That is, it can be said that this processing is necessary in order not to create any extra character cutout candidates.

Further, as a connection determination feature for learning, not only a 14-dimensional vector and a 12-dimensional vector but also image information of a minimum unit may be added and used. When 14-dimensional data or 12-dimensional data described above is used, the minimum unit is treated as rectangular information, and it is difficult to know what kind of image pattern is inside the rectangular information. By using image information, it becomes possible to know the shape pattern of the minimum unit itself. As this image information, for example, the image pattern of the minimum unit i is normalized to a size of 16 × 16 to obtain a 16 × 16
It is also possible to obtain a dimensional vector as Img _i , similarly obtain a 16 × 16 dimensional vector as Img _j from the pattern of the minimum unit j, and use Img _i and Img _j as image information. In this case, the connection determination feature is 14 dimensions + (16 × 16)
× 2 dimensional vector or 12 dimensional + (16 × 16) × 2 dimensional vector.

Further, the principal unit analysis is performed by normalizing the image of the minimum unit as many learning samples to a size of 16 × 16, and ten eigenvectors are used in descending order of the eigenvalue, and the eigenvector and the input pattern are compared. Dot product can also be used. In this case, the image information is a 10-dimensional vector, and the connection determination feature is a vector whose dimension is increased by 20 dimensions. Here, it has been described that ten eigenvectors are used in the descending order of eigenvalues, but the number is not limited to ten and can be set freely from one to 256.

In the above description, the size is 16 × 16. However, the size is not limited to this, and it is possible to normalize the size to N × M as needed. is there. At this time, N and M can be freely set in the range of 1 to 256, respectively. It is also possible to divide the original image area of the minimum unit of interest into small areas, obtain the content ratio of black pixels while overlapping the areas little by little, and use this as image information. The content ratio is a ratio of black pixels occupying in the region. For example, if the size of the small region is 10 × 10 and there are 30 black pixels in the small region, the content ratio of the black pixels is 30 /
10 × 10 = 0.3. Explaining how to divide a small area with reference to FIG. 5, an image of a connected component is divided into nine small areas. This small area slightly overlaps with the adjacent small area, but the number of black pixels in this area is counted and divided by the area of the area to obtain the black pixel occupancy for each small area. , Which can be used as image information.

As described above, the image information of the minimum unit i is
Img_i, Img the image information of the minimum unit j_jAnd connection
As a fixed feature, (X_i, Y_i, W_i, H_i, X_j, Y_j, W_j, H_j, X_ij, Y_ij,
W_ij, H_ij, W _PSmean, H_PSmean, Img_i, Img_j) Can be used
Wear. Or (X_i/ W_PSmean, Y_i/ H_{P Smean}, W_i/ W_PSmean, H_i
/ H_PSmean, X_j/ W_PSmean, Y_j/ H_PSmean, W_j/ W_PSmean, H_j/ H
_PSmean, X_ij/ W_PSmean, Y_ij/ H_PSmean, W_ij/ W_PSmean, H_ij/ H
_PSmean, Img_i, Img_j) Can also be used.

The significance of using the minimum unit image pattern will be described. For example, PS08 in FIG.
Paying attention to the relationship between PS09 and PS09, there is a return in the part below the vertical stroke in PS08. In kanji, this return often means the end of a character, and when such a pattern is written, it is not very often that a correct character is represented even when combined with the smallest unit below it. Detecting such a pattern and using it for determining the connection relationship has the effect of suppressing the creation of any extra character cutout candidates, thereby preventing the explosion of combinations.

FIG. 18 is a flowchart for explaining the processing for obtaining the connection relationship. At ST31, focusing on the i-th minimum unit, a connection determination feature calculated between the i-th minimum unit and the j-th minimum unit adjacent to the i-th minimum unit is obtained with reference to the adjacent relation information at ST32. At this time, it is created by adding information as to whether the two Voronoi sides are shared by the i-th minimum unit and the j-th minimum unit. In ST33, if the i-th minimum unit and the j-th minimum unit share two Voronoi sides, it is immediately determined that they do not have a connection relationship, and if not, statistical analysis is performed using the connection determination feature. It is determined whether or not the connection is established. In ST34, when attention is paid to all the minimum units j adjacent to the minimum unit i, the process proceeds to ST35, and otherwise, the process returns to ST32. In ST35, when attention is paid to all minimum units, the process proceeds to ST36,
Otherwise, return to ST31. In ST36, the connection relation determined in ST33 is output as connection relation information.

FIG. 19 shows an example of connection relation information created when the pattern shown in FIG. 2A is input. The connection relation information has the same format as the adjacent relation information. For example
Although PS01 and PS03 are determined to be adjacent by Voronoi division, they are determined to be out of connection because the distance between the centers of the rectangles is large. Also PS02 and PS03
Are also determined to be adjacent by Voronoi division,
Since both have a rectangular center-to-center distance and are obliquely adjacent to each other, it is determined that there is no connection relationship.
The same applies to PS05 and PS07.

As described above, whether connection is possible or not is determined.
After the connection relation is given, the character segmentation candidate output unit 1
In step 5, a combination pattern is created from some of the minimum units, and it is determined whether or not the combination may be used as a character cutout candidate, that is, whether output is possible.

In creating a combination pattern, it is necessary to determine in advance how many minimum units are to be combined.

Here, when a certain combination pattern is composed of one minimum unit, the number of combinations is defined as one. Further, the number of combinations when a certain combination pattern is composed of two minimum units is defined as 2, and the number of combinations when a certain combination pattern is composed of N minimum units is N.

When creating a combination pattern, for example, a combination pattern having 1 to 10 combinations may be created. Here, the lower limit of the number of combinations is 1 and the upper limit is 10, but the upper limit of the number of combinations can be set freely, and a value from 2 to 100 can be set.

When a pattern having the number of combinations of 1 is created, one minimum unit may be a combination pattern. When creating a pattern having two combinations, for example, a pattern can be created by focusing on a certain minimum unit i and combining it with a minimum unit j determined to be connectable to the minimum unit i. At this time, i ≠ j. When creating a pattern having three combinations, for example, a certain minimum unit
Focusing on i, created by combining the minimum unit j determined to be connectable to the minimum unit i and the minimum unit k determined to be connectable to either the minimum unit i or the minimum unit j It is possible. At this time, i ≠ j ≠ k. Similarly, it is possible to create a combination pattern of the number N of combinations.

The character cutout candidate output section 15 determines whether or not the combination pattern created in this way may be output as a character cutout candidate.

For this determination, several methods are conceivable. For example, when a combination of some minimum units is created, a pattern having a Voronoi region closed inside should be output as a character extraction candidate. Can not be determined.

More specifically, (a) and (b) of FIG.
Each pattern has three minimum units, but (a)
In the example, if only a combination of PSa1 and PSa3 is created, a closed Voronoi region will be created inside. Similarly, (b)
In the above example, even with a combination of PSb1 and PSb2, a Voronoi region closed inside is created. In general, characters in a character string have the property that one closed area is created when the minimum unit is Voronoi-divided, and the property that character patterns contrary to this are very rare is used.
It is possible to perform control so that a pattern having a Voronoi region closed inside is not output as a character extraction candidate.

Further, the determination accuracy can be further improved by determining the output propriety of the combination pattern using different determination methods. For example, paying attention to a certain combination pattern, the height and width of the circumscribed rectangle of the combination pattern are set to (H _C , W _C ), and it is possible to determine whether or not to output statistically as a character cutout candidate. . Here, the subscript C means the C-th combination pattern.

For example, as a statistical determination method, a large number of combination patterns that are correct character candidates are prepared in advance for learning, and the width and height of the circumscribed rectangle are set to (W
_L, when the H _L), W and (W on a two-dimensional space spanned by H _L, plots the H _L), for example, separated in a 50 × 50 mesh, correct character candidate in Mesh Is found, and the frequency can be used. Here, the suffix L means the L-th learning data. At this time, the space spanned by W and H originally extends from zero to infinity, but the range in which the learning data (W _L , H _L ) exists is determined in advance, and the range may be divided into 50 × 50 meshes. . It is also possible to set the frequency of a portion that deviates from a preset mesh range to zero.

When a certain combination pattern is considered, it can be determined that a character cutout candidate can be output when the frequency value is equal to or greater than a certain threshold value. By this determination, it is possible to eliminate apparently abnormal patterns (patterns whose width, height, area is too large, the aspect ratio is crushed abnormally or vertically long), and extra character cutout candidates are Output can be suppressed.

Also, it has been described that the width and height of the circumscribed rectangle of the combination pattern are used as (W _C , H _C ) to determine whether or not it is possible to output as a character cutout candidate. The average value of the width and height of all the minimum units obtained from is obtained as (W _PSmean , H _PSmean ), and the width and height of the circumscribed rectangle of the combination pattern C that is the correct answer are normalized by the average value as ( W _C / H _PSmean , H _C / W _PSmean ) can be obtained and used.

Normally, the width and height of characters obtained from different character string images differ greatly in size, but by using quantities normalized by the minimum unit width and height of each image,
It is possible to reduce the size variation.

In addition to this determination, a further determination method may be added. For example, assuming that the combination pattern C of interest is made up of the minimum units i, j, k, consider another minimum unit l that is connected to one of the minimum units i, j, k. i ≠ j ≠ k. The center coordinates (X _C , Y _C ) of the circumscribed rectangle of the combination pattern and the width and height (W _C , H _C ) of the circumscribed rectangle are obtained.
The center point coordinates (X _Cl , Y _Cl ) and the width and height (W _Cl , H _Cl ) of the circumscribed rectangle are also obtained from the minimum unit l connected to the center, and the center of the new circumscribed rectangle P created when these are combined Create the point coordinates (X _P , Y _P ) and the width and height (W _P , H _P ) of the circumscribed rectangle, and further calculate the average width and average height (W _PSmean , H _PSmean ) of all minimum units, By using these values, it is possible to statistically determine whether or not to output as character extraction candidates.

That is, (X_C,
Y_C, W_C, H_C, X_Cl, Y_Cl, W_Cl, H_Cl, X_P, Y_P, W_P, H_P, W_PSmean, H
_PSmean) Can be used to statistically determine whether output is possible.
Wear. Also, (X_C/ W_PSmean, Y_C/ H_PSmean, W_C/ W_PSmean, H_C/ H
_PSmean, X_Cl/ W_PSmean, Y_Cl/ H_PSmean, W _Cl/ W_PSmean, H_Cl/ H
_PSmean, X_P/ W_PSmean, Y_P/ H_PSmean, W_P/ W_PSmean, H_P/
H_PSmean) Can also be used.

For example, a character string image for learning is prepared.
From 1 to N combinations based on the adjacency
Combination pattern CP_CCreate Where C is a serial number
And CP_CMeans the i-th combination pattern.
Next, the combination pattern CP _CAnd other adjacent minimum units
Set PS_ClAsk for. Where the subscript C is CP_CAdjacent to
L means CP_CNumber l adjacent to
It is the smallest unit of the eye. Where all CPs
_CAnd PS_ClThe output feasibility judgment feature is obtained from the pair of
Combination pattern CP_CRepresents the correct character
Teacher data T_Cl= 1, otherwise (ie combination
Pattern CP_C(If is not a correct character)
_Cl0 can be set for learning.

When performing, for example, regression analysis as statistical learning, the output data determination feature is used as an explanatory variable and the teacher data T
It can be considered as a problem of estimating _Cl . When a multilayer perceptron-type neural network that learns by back propagation is used, an input layer is provided with an output availability determination feature, and learning is performed by providing T _Cl as teacher data. The structure of the neural network at this time is as follows:
The number of neurons in the input layer is given the same number as the number of dimensions of the contact determination feature, and it is possible to have a structure having five neurons in the intermediate layer and one neuron in the output layer, The number of neurons in the intermediate layer can be freely changed from two to 1,000. Also, when using discriminant analysis, T _Cl is considered a category,
It is also possible to perform discriminant analysis using the output availability determination feature. It is also conceivable to use principal component analysis, Mahalanobis distance, or the like.

As described above, using the learned neural network, the regression coefficient and the discriminant axis, the combination pattern CP having a value larger than the predetermined threshold value is obtained.
It is possible to output _C as a character extraction candidate. In other words, when using a trained neural network, input a vector of the output availability determination feature and output 0.
If the number is 5 or more, it is highly likely that the combination pattern represents a correct character, the character pattern can be output as a character segmentation candidate, and a similar determination can be made by multiple regression analysis. Also, in the case of discriminant analysis, it is possible to use a discriminant score for an output propriety judgment feature, and if a discriminant score equal to or greater than a predetermined threshold is obtained, it is possible that the combination pattern represents a correct character. If it is high, it can be output as a character segmentation candidate.

Also, generally, one combination pattern CP
_{For C} , a plurality of adjacent minimum units PS _Cl are obtained.
In this case, the output determination of CP _C is the minimum unit adjacent PS
_This is performed as many times as the number of _Cl . At this time, if at least one of the l determinations is determined to be equal to or less than the threshold, the CP _C
Can be rejected as having a low possibility of representing a correct character, or the ratio of l times that the correct character is determined to represent the correct character, i.e., if it is determined that m represents the correct character, m When the value of / l is larger than a predetermined threshold, it is possible to output a character extraction candidate as a possibility that it represents a correct character.

Here, a description will be given with reference to FIG. FIG.
From (a) (e) illustrates an example of CP ₅ from combination pattern CP _1. If described in example (a) of FIG. 21, CP ₁
When evaluating, first performs output determination of CP ₁ from the relation CP ₁ and PS11, to output determination of CP ₁ then the relationship CP ₁ and PS12.

In the example of FIG. 21A, the relationship between CP ₁ and PS ₁₁
FIG. 22 shows a state in which the output permission / inhibition determination feature is created in the case of performing the output permission / inhibition determination of ₁ .

Here, the meaning of making such output availability determination is as follows.
I will explain justice. For example, the pattern in FIG.
Each is a character cutout candidate representing a correct character.
(a) CP ₁However, character extraction that is appropriate in terms of size
As a candidate, and the minimum unit PS around it₁₁And PS
₁₂Against CP₁What spatial relationship has
If you evaluate, PS₁₁And PS₁₂Vertically positioned for both
You can see that it is. Character strings can be written vertically or horizontally.
Because it is often listed, CP₁Is the smallest unit around
It can be said that the arrangement is appropriate. The same applies to the case (c).

On the other hand, in the pattern (e), the two minimum units PS ₅₁ and PS ₅₂ which are assumed to be diagonally combined with the internal minimum units forming the combination pattern and are not the same character are also diagonal. , The pattern is clearly not valid as a character segmentation candidate.

That is, the size of the combination pattern itself
If you try to evaluate the validity of a character using only (width and height), once the pattern as shown in (e) is judged to be a valid size, the pattern of the arrangement with the surroundings is irregular. Even so, there is a problem that it cannot be excluded.

FIGS. 21 (b) and (d) show an example in which the two patterns (a) and (c) are in the middle of the pattern (e). It can be expected that the pattern of (d) is more likely to be determined to be unsuitable as a character segmentation candidate and eliminated than the example of ()). It is also significant to use statistical judgment in order to make such a delicate judgment with high accuracy.

[0096] Further, as the output determination features for learning, be used also added image information of the minimum unit PS _{C l} adjacent to the well vector of 14-dimensional vectors and 12-D, a combination pattern CP _C it Is also conceivable.
When 14-dimensional data or 12-dimensional data described above is used, the combination pattern and the minimum unit are treated as rectangular information, and it is difficult to know what kind of image pattern is inside. However, using the image information makes it possible to know the combination pattern and the shape pattern of the minimum unit itself. As this image information, for example, an image pattern of the combination pattern CPC is normalized to a size of 16 × 16 to obtain a 16 × 16-dimensional vector Img _C , and a 16 × 16-dimensional vector is similarly obtained from the pattern of the minimum unit PS _Cl . The vector can be obtained as Img _Cl , and Img _C and Img _Cl can be used as image information. In this case, the feature amount is a vector whose dimension is increased by (16 × 16) × 2 dimensions.

Also, the combination pattern as a large number of learning samples and the image of the minimum unit are each normalized to a size of 16 × 16, and the principal component analysis is performed. For each image, ten eigenvectors are used in the order of large eigenvalue. Thus, the inner product of the eigenvector and the input pattern can be used. In this case, each piece of image information is a 10-dimensional vector, and the feature amount is a vector whose dimension is increased by 20 dimensions. In this case, 10
Although it has been described that the number of eigenvectors is used, the number is not limited to ten, but can be set freely from one to 256.

In the above description, the size is 16 × 16. However, the size is not limited to this, and it is possible to normalize the size to N × M as needed. is there. At this time, N and M can be freely set in the range of 1 to 256, respectively. It is also possible to divide the combination image of interest and the original image area of the minimum unit into small areas, obtain the content ratio of black pixels while overlapping the areas little by little, and use this as image information. The content ratio is a ratio of black pixels occupying in the region. For example, if the size of the small region is 10 × 10 and there are 30 black pixels in the small region, the content ratio of the black pixels is 30/10 × 10 = 0.3. Explaining how to divide a small area with reference to FIG. 5, an image of a connected component is divided into nine small areas. This small area slightly overlaps with the adjacent small area, but the number of black pixels in this area is counted and divided by the area of the area to obtain the black pixel occupancy for each small area. , Which can be used as image information.

As described above, image information is also added as an output availability determination feature (X _C , Y _C , W _C , H _C , X _Cl , Y _Cl , W _Cl , H _Cl , X
_P , Y _P , W _P , H _P , W _Psmean , H _Psmean , Img _C , Img _Cl ) can be used to statistically determine whether output is possible. Also, (X _C / W
_PSmean , Y _C / H _PSmean , W _C / W _PSmean , H _C / H _PSmean , X _Cl / W
_PSmean , Y _Cl / H _PSmean , W _Cl / W _PSmean , H _Cl / H _PSmean , X _P / W
_{PSmean, Y P / H PSmean,} W P / W PSmean, H P / H PSmean, Img C, Im
g _Cl ) can also be used.

The operation of the character segmentation candidate output unit will be described with reference to a flowchart shown in FIG. In ST41, a combination pattern in which the number of combinations is 1 to N is created, and in ST42, the generated C-th combination pattern CP _C
Pay attention to the CP and refer to the connection-related information in ST43.
_Focus on the l-th minimum unit PS _Cl connected to _C. In ST44, an output availability determination feature obtained between CP _C and PS _Cl is obtained. At this time, the Voronoi region created by the CP _C in addition to the output determination features, is also added information whether there is a Voronoi region closed making different minimum unit. In ST45,
Outputs but proceeds to next step ST46 is determined to not be statistically output determination rows by using output determination characterized if not so, when containing the Voronoi region closed CP _C Will be In ST46, when attention is paid to all l, the process proceeds to ST47, and otherwise, the process returns to ST43. In ST47, when all the Cs are focused on, the process proceeds to ST48, otherwise, S is performed.
Return to T42. In ST48, the combination pattern determined to be output in ST45 is output as a character cutout candidate.

FIG. 24 shows an example of a character cutout candidate output when the upper limit of the number of combinations is set to 4 when a pattern as shown in FIG. 2A is input. From the left column vertically separated, the number of combinations is 1, and the rightmost column shows character cutout candidates with the number of combinations of 4. In the example of FIG. 2A, an example in which only one line is described is shown. However, even if such a pattern is subjected to a line cutout process, the line may be cut out as one line, or may be cut out as two lines. In some cases, once line segmentation is erroneous, it becomes difficult to create a correct character segmentation candidate.
(a) explained that normally only one line is described, but if this character string image is something like a cipher, the first line is `` Tsukitsukicho '' and the second line is `` Na Hita Even if it is written with the intention of "3", according to this method, a corresponding character segmentation candidate is generated. In other words, according to the present method, it is possible to generate a character segmentation candidate so as to be able to cope with either case without knowing whether it is one line or two lines.

As in the present invention, by creating character extraction candidates without assuming line extraction processing, character extraction candidates can be correctly obtained even for character string patterns for which line extraction is difficult. At this time, the character cutout candidate is obtained as data of the image pattern obtained by combining the minimum units, the width and height of the circumscribed rectangle, and the coordinates of the center of the circumscribed rectangle.

The data corresponding to FIG. 24 is held as character cutout candidate information in a format as shown in FIG.
That is, a character cutout candidate is its serial number
ID, the number of combinations of how many minimum units are combined, the width and height (Wi, Hi) of the circumscribed rectangle, the coordinates (Xi, Yi) of the center of the circumscribed rectangle, and the ID of the minimum unit used in the combination It is output as character cutout candidate information. Here, the subscript i indicates the i-th character cutout candidate, and corresponds to the ID number. Further, the image pattern of the character cutout candidate can be easily created because the image pattern can be extracted by referring to the minimum unit information using the IDs of all the minimum units used in the combination.

As described above, according to the present invention, the connection relationship between the minimum units in character segmentation is defined two-dimensionally, and a combination pattern is created by following the connection relationship. It is possible to create a character segmentation candidate, and to correctly obtain a character segmentation candidate even for a character string pattern for which line extraction is difficult. In addition, when outputting a combination pattern as a character segmentation candidate, the arrangement relationship between the minimum units around the combination pattern is evaluated, and the image-like pattern of the combination pattern is also evaluated. It is possible to suppress that it is done, and it is possible to prevent the explosion of the combination.

The character cutout determining section 16 performs recognition processing, linguistic knowledge processing, and layout analysis based on the character cutout candidates obtained by the character cutout candidate output section 15 and outputs a character cutout result. The character cutout confirmation section is described in Japanese Patent Application
This can be easily realized by using the method described in JP-A-203148.

The character cutout determining unit 16 using the method described in Japanese Patent Application No. 7-203148 starts processing from a state where character recognition results are scattered two-dimensionally. To create multiple character string candidates,
This method calculates the layout cost for these character string candidates and reads the character string using the character recognition cost, linguistic cost, and layout cost at the same time. Is possible.

When the character cutout candidate as shown in FIG. 24 is obtained, the character cutout determining section 16 reads the input of FIG. 2A using the knowledge of the address as the linguistic knowledge. Finally, the character cutout result is shown in FIG.
become that way. At this time, as shown in FIG. 27, character extraction result information is output in a data format having a character extraction candidate ID. When using data of the character extraction result after character extraction processing, such as displaying the processing result,
It can be easily handled by referring to the minimum unit information and the character segmentation candidate information.

In the above embodiment, the character cutout candidate generating apparatus and the character cutout candidate generating method have been described. However, the present invention provides a program for realizing the above character cutout candidate generating processing by a computer, and furthermore, this program Is also included, and the effects described in the above embodiments can be obtained by loading and realizing the program on a computer.

[0109]

As described above, according to the present invention, the connection relationship between the minimum units in character segmentation is defined two-dimensionally, and a combination pattern is created by following the connection relationship. It is possible to create a character segmentation candidate without removing the character segment, and it is possible to obtain a character segmentation candidate corresponding to both cases even for a character string pattern in which it is difficult to extract a line and it is not clear whether it is one line or two lines. In addition, when outputting a combination pattern as a character segmentation candidate, the arrangement relationship between the minimum units around the combination pattern is evaluated, and the image-like pattern of the combination pattern is also evaluated. It is possible to suppress being done. Then, since the character string is read using the character recognition cost, the linguistic cost, and the layout cost at the same time, even if multiple lines are input at the same time, without assuming that line extraction is performed, the combination does not explode, The extracted character pattern can also be correctly cut, and correct character cutout can be performed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a character segmentation candidate creation device according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of a character string image.

FIG. 3 is a diagram showing an example in which connected components are extracted from the example of FIG. 2 (a).

FIG. 4 is a diagram illustrating an example of connected component information corresponding to the example of FIG.

FIG. 5 is a diagram illustrating an operation when creating image information.

FIG. 6 is a flowchart illustrating a procedure in a minimum unit creating unit.

FIG. 7 is a diagram illustrating an example in which a minimum unit is created for the example of FIG.

FIG. 8 is a diagram illustrating an example of minimum unit information corresponding to the example of FIG.

FIG. 9 is a diagram for explaining the significance of defining an adjacency relationship.

FIG. 10 is a diagram for explaining Voronoi division processing.

FIG. 11 performs Voronoi division on the example of FIG.
It is a figure showing the example which extracted Voronoi side.

FIG. 12 is a flowchart illustrating a procedure for creating an adjacent relationship.

FIG. 13 is a diagram illustrating an example of adjacent relation information corresponding to the example of FIG.

FIG. 14 is a diagram illustrating an operation example when Voronoi division is performed by changing the distance scale between the vertical axis and the horizontal axis.

FIG. 15 is a diagram for explaining the significance of creating a connection relationship.

FIG. 16 is a diagram illustrating an operation when creating a connection determination feature.

FIG. 17 is a diagram illustrating the significance of adding image information to a connection determination feature.

FIG. 18 is a flowchart illustrating a procedure for creating a connection relationship.

FIG. 19 is a diagram illustrating an example of connection relation information corresponding to the example of FIG.

FIG. 20 is a diagram illustrating an operation of determining whether to output a combination pattern.

FIG. 21 is a diagram illustrating a relationship between a combination pattern and a minimum unit around the combination pattern.

FIG. 22 is a diagram illustrating an operation when creating an output possibility determination feature.

FIG. 23 is a flowchart illustrating a procedure in a character segmentation candidate creation unit.

FIG. 24 is a diagram illustrating an example of a character cutout candidate corresponding to the example of FIG.

FIG. 25 is a diagram illustrating an example of character cutout candidate information corresponding to the example of FIG. 24;

FIG. 26 is a diagram illustrating an example of a character cutout result corresponding to the example of FIG.

FIG. 27 is a diagram illustrating an example of character cutout result information corresponding to the example of FIG. 26;

[Explanation of symbols]

 Reference Signs List 11 character string image storage unit 12 connected component extraction unit 13 minimum unit creation unit 14 two-dimensional connection relationship creation unit 15 character cutout candidate output unit 16 character cutout determination unit 17 computer 18 character string image reading device 19 display means 20 recording medium

Claims (37)

[Claims] 1. A character segmentation apparatus for generating a character segmentation candidate from a character string image, comprising: a character string image storage unit for inputting and storing the character string image; and binarizing the character string image to extract a connected component The connected component extraction unit performs a determination as to whether or not all the connected components are in contact with each other. When it is determined that the connected components are not in contact with each other, the connected component itself is used as a minimum unit, and the connected components are contacted. A minimum unit creating unit that cuts the connected component to create a new minimum unit when it is determined that the minimum unit is formed, and a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space. A two-dimensional connection relationship creating unit for obtaining a combination pattern, and creating a combination pattern by combining a plurality of the minimum units having a connection relationship with each other, and extracting a combination pattern having a high possibility of being a correct character. A character cutout candidate output unit for outputting as a candidate. 2. A character extracting apparatus for generating a character extracting candidate from a character string image, comprising: a character string image storage unit for inputting and storing the character string image; and binarizing the character string image to extract a connected component. The connected component extraction unit performs a determination as to whether or not all the connected components are in contact, and when it is determined that the connected components are not in contact, the connected component itself is used as a minimum unit, and the connected components are A minimum unit creating unit that cuts the connected component to create a new minimum unit when it is determined that the minimum unit is formed, and a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space. A two-dimensional connection relationship creating unit for obtaining a combination pattern, and creating a combination pattern by combining a plurality of the minimum units having a connection relationship with each other, and extracting a combination pattern having a high possibility of being a correct character. A character extraction candidate output unit that outputs as a candidate, and a character extraction determination unit that performs recognition processing, linguistic knowledge processing, and layout analysis based on the character extraction candidate, and outputs a character extraction result, is provided. Character extraction device. 3. The minimum unit creation unit refers to the average value of the width and height of all connected components extracted from the character string image and the width and height of the connected component of interest to generate contact. 3. The character segmenting device according to claim 1, wherein whether the component is a connected component is statistically determined. 4. The method according to claim 1, wherein the minimum unit creation unit further statistically determines whether or not the connected component is in contact with reference to an image pattern representing the shape of the focused connected component. The character segmentation device according to claim 3, wherein 5. The two-dimensional connection relation creation unit determines that adjacent minimum units that share Voronoi sides by Voronoi division are adjacent when determining the adjacent relation of the minimum unit extracted from the character string image. Further, when the minimum units having the adjacent relation share two Voronoi sides, it is determined that the minimum units are not in the adjacent relation. A character cutout device according to one of the preceding claims. 6. The two-dimensional connection relationship creating unit creates a combination of the width, height, and center coordinate value of each of the plurality of minimum units determined to be in an adjacent relationship, and the plurality of minimum units. With reference to the width, height, and center coordinate values of the circumscribed rectangle and the average value of the width and height of all the minimum units extracted from the character string image, the connection relationship is determined by statistical evaluation. 6. The character segmenting device according to claim 5, wherein: 7. The character according to claim 6, wherein the two-dimensional connection relation creation unit further obtains a connection relation with reference to an image pattern representing a shape of each of the plurality of minimum units. Cutting device. 8. The character cutout candidate output section creates a plurality of minimum unit combination patterns by tracing the connection relationship obtained by the two-dimensional connection relationship creation section when creating the combination pattern. The character cutout device according to any one of claims 1 to 7, wherein: 9. The character cutout candidate output unit refers to the width, height, and center coordinate value of the combination pattern and the average value of the width and height of all minimum units extracted from the character string image. 9. The character cutout device according to claim 8, wherein whether to output the combination pattern as a character cutout candidate is determined by statistical evaluation. 10. The character segmentation candidate output unit,
Width, height and center coordinate value of the combination pattern,
The width, height and center coordinate values of the minimum unit adjacent to the combination pattern, the width, height and center coordinate values of a circumscribed rectangle created by combining the combination pattern and the minimum unit, and the character string image With reference to the average value of the width and height of all the minimum units extracted from the above, statistical evaluation is performed for the same number of times as the number of the minimum units adjacent to the combination pattern, and the combination pattern is a character cutout candidate. 9. The character segmenting device according to claim 8, wherein whether to output the character is determined. 11. The character segmentation candidate output unit,
Furthermore, with reference to the image pattern representing the shape of the combination pattern and the image pattern representing the shape of the minimum unit adjacent to the combination pattern, it is determined whether to output the combination pattern as a character cutout candidate. The character cutout device according to claim 9 or 10, wherein: 12. A character extracting method for generating a character extracting candidate from a character string image, comprising: a character string image storing step of inputting and storing the character string image; and binarizing the character string image to extract a connected component. Performing a connected component extraction step, and determining whether or not all the connected components are in contact.When it is determined that the connected components are not in contact, the connected component itself is used as a minimum unit, and the connected components are A minimum unit creation step of cutting the connected component to create a new minimum unit when it is determined that the minimum unit has been established; and a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space. A two-dimensional connection relationship creating step of obtaining a combination pattern, and creating a combination pattern by combining a plurality of the minimum units having a connection relationship with each other, and a combination having a high possibility of being a correct character. A character segmentation candidate output step of outputting a matching pattern as a character segmentation candidate. 13. A character clipping method for generating a character clipping candidate from a character string image, comprising: a character string image storing step of inputting and storing the character string image; and binarizing the character string image to extract a connected component. Performing a connected component extraction step, and determining whether or not all the connected components are in contact.When it is determined that the connected components are not in contact, the connected component itself is used as a minimum unit, and the connected components are A minimum unit creating step of creating a new minimum unit by cutting the connected component when it is determined that the minimum unit has been connected; and a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space. A two-dimensional connection relationship creating step of obtaining a combination pattern, and creating a combination pattern by combining a plurality of the minimum units having a connection relationship with each other, and a combination having a high possibility of being a correct character. A character cutout candidate output step of outputting a matching pattern as a character cutout candidate, and a character cutout determining step of performing recognition processing, linguistic knowledge processing, and layout analysis based on the character cutout candidate, and outputting a character cutout result. A character extraction method characterized by having at least a character. 14. In the minimum unit creation step, contact is generated by referring to the average value of the width and height of all connected components extracted from the character string image and the width and height of the connected component of interest. 14. The character segmentation method according to claim 12, wherein it is statistically determined whether the component is a connected component. 15. The method according to claim 15, wherein in the minimum unit creating step, it is further determined statistically whether or not the connected component is in contact with reference to an image pattern representing the shape of the focused connected component. The character segmentation method according to claim 14. 16. In the two-dimensional connection relation creation step, when determining the adjacent relation of the minimum unit extracted from the character string image, it is determined that the minimum units sharing the Voronoi sides by Voronoi division are adjacent. And then
The character according to any one of claims 12 to 15, wherein when the minimum units having the adjacent relation share two Voronoi sides, it is determined that the minimum units are not in the adjacent relation. Cutting method. 17. The method for determining the adjacency relation of the minimum unit includes the steps of: (a) creating a mother point on the contour of all of the minimum units; and (b) Voronoi using all the created mother points. Dividing and obtaining a Voronoi edge; and (c) focusing on a certain Voronoi edge in the Voronoi edge and, if the Voronoi edge is obtained between the generating points of the same minimum unit, (D) determining whether all Voronoi sides have been focused on, and returning to (c) if there are Voronoi sides that have not been focused on; and (e) determining the same Voronoi sides. 17. The method according to claim 16, further comprising: determining that the minimum units to be shared have an adjacent relationship. 18. The two-dimensional connection relationship creating step, wherein the width, height, and center coordinate value of each of the plurality of minimum units determined to be in an adjacent relationship is combined with the plurality of minimum units. Statistically evaluating the width, height, and center coordinate value of the circumscribed rectangle and the average value of the width and height of all the minimum units extracted from the character string image to obtain a connection relationship. 18. The character segmentation method according to claim 16, wherein: 19. The character segmentation according to claim 18, wherein in the two-dimensional connection relation creating step, a connection relation is further obtained by referring to an image pattern representing the shape of each of the plurality of minimum units. Method. 20. In the character cutout candidate output step, when the combination pattern is created, by following the connection relationship obtained in the two-dimensional connection relationship creation step, a combination pattern of a plurality of minimum units is created. The character cutout method according to any one of claims 12 to 19, characterized in that: 21. In the character cutout candidate outputting step, the width, height, and center coordinate value of the combination pattern and the average value of the width and height of all the minimum units extracted from the character string image are referred to. 21. The character cutout method according to claim 20, wherein the character pattern is statistically evaluated to determine whether to output the combination pattern as a character cutout candidate. 22. In the character segmentation candidate output step, the width, height and center coordinate value of the combination pattern, the width, height and center coordinate value of the minimum unit adjacent to the combination pattern, and the combination pattern The width, height, and center coordinate value of a circumscribed rectangle created by combining the minimum unit and the average value of the width and height of all minimum units extracted from the character string image, 21. The character segmentation method according to claim 20, wherein whether to output the combination pattern as a character segmentation candidate is determined by statistically evaluating the same number of times as the number of the minimum units adjacent to the pattern. 23. The character segmentation candidate output step, further comprising: referring to an image pattern representing the shape of the combination pattern and an image pattern representing the shape of the minimum unit adjacent to the combination pattern. 23. The character extracting method according to claim 21, wherein it is determined whether or not to output as a character extracting candidate. 24. A determination as to whether or not to output as a character segmentation candidate includes the steps of: (a) creating a combination pattern having 1 to N combinations; and (b) the C-th combination pattern from the combination patterns. a step of focusing on the combination pattern CPc, (c) is connected to the CP _C l th the minimum unit PS
A step of focusing on _Cl, and performing an output acceptability decision obtained between the PS _Cl (d) and the CP _C, it is determined whether or not focusing on all l (e), are focused (F) determining whether or not all Cs have been focused, and returning to (b) if there are Cs not focused; 24. The character extracting method according to claim 20, further comprising: g) outputting the combination pattern determined to be output-possible in (d) as a character extracting candidate. 25. A character string image storing process for inputting and storing a character string image, a connected component extracting process for extracting a connected component by binarizing the character string image, and contacting all the connected components. Determine whether or not, when the connected component is determined not to be in contact with the connected component itself as the minimum unit, cut off the connected component when it is determined that the connected component is in contact A minimum unit creation process for creating a new minimum unit, a two-dimensional connection relationship creation process for obtaining a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space, Character combination candidate output processing for creating a combination pattern by combining the minimum units with each other and outputting a combination pattern having a high possibility of being a correct character as a character extraction candidate. Character segmentation program to be executed. 26. A character string image storing process for inputting and storing a character string image, a connected component extracting process for extracting a connected component by binarizing the character string image, and contacting all the connected components. Determine whether or not, when the connected component is determined not to be in contact with the connected component itself as the minimum unit, cut off the connected component when it is determined that the connected component is in contact A minimum unit creation process for creating a new minimum unit, a two-dimensional connection relationship creation process for obtaining a two-dimensional connection relationship between the minimum units arranged in a two-dimensional space, A character segmentation candidate output process for creating a combination pattern by combining the minimum units with each other and outputting a combination pattern likely to be a correct character as a character segmentation candidate; Complement based on, performs the recognition processing and language knowledge processing and layout analysis, character segmentation program for executing a character segmentation decision process of outputting the character extraction result to the computer. 27. The minimum unit creation processing includes the steps of: averaging the width and height of all connected components extracted from the character string image;
27. The character segmentation program according to claim 25, further comprising a process of referring to the width and height of the connected component of interest to statistically determine whether the connected component is in contact. 28. The minimum unit creation processing further includes the step of referring to an image pattern representing the shape of the target connected component,
28. The character segmenting program according to claim 27, further comprising a process of statistically determining whether or not the connected component is in contact. 29. In the two-dimensional connection relationship creation processing, when obtaining the adjacent relationship of the minimum unit extracted from the character string image, it is determined that the minimum units sharing the Voronoi side by Voronoi division have an adjacent relationship. 29. The method according to claim 25, further comprising a step of determining that the minimum units are not in an adjacent relationship when the adjacent minimum units share two Voronoi sides. Character extraction program. 30. A process for judging the adjacency relationship of the minimum unit, comprising: (a) a process of creating a mother point on all the contours of the minimum unit; and (b) a process of using all the created mother points. (C) paying attention to a certain Voronoi edge in the Voronoi edge, and if the Voronoi edge is obtained from the same minimum unit between the generating points, the Voronoi edge (D) determining whether all Voronoi sides have been focused on, and returning to (c) when there is a Voronoi side that has not been focused on; (e) the same Voronoi side 30. The computer-readable storage medium according to claim 29, further comprising: a process of determining that the minimum units sharing the same are adjacent to each other. 31. The two-dimensional connection relationship creation process, wherein the width of each of the plurality of minimum units determined to be adjacent to each other,
Height and center coordinate value, and width, height and center coordinate value of a circumscribed rectangle created by combining the plurality of minimum units,
31. The character segmentation program according to claim 29, further comprising a process of statistically evaluating a connection relationship by referring to an average value of widths and heights of all minimum units extracted from the character string image. 32. The two-dimensional connection relationship creation processing further includes:
32. The character cutout program according to claim 31, further comprising a process of obtaining a connection relationship by referring to an image pattern representing each shape of the plurality of minimum units. 33. A process for creating a plurality of minimum unit combination patterns by following the connection relationship obtained in the two-dimensional connection relationship creation step in the character cutout candidate output process when creating the combination pattern. 33. The character segmentation program according to claim 25, comprising: 34. The character cutout candidate output process refers to the width, height, and center coordinate value of the combination pattern and the average value of the width and height of all minimum units extracted from the character string image. 34. The character cutout program according to claim 33, further comprising a process of statistically evaluating to determine whether to output the combination pattern as a character cutout candidate. 35. In the character cutout candidate output processing, the width, height and center coordinate value of the combination pattern, the width, height and center coordinate value of the minimum unit adjacent to the combination pattern, and the combination pattern The width, height, and center coordinate value of a circumscribed rectangle created by combining the minimum unit and the average value of the width and height of all minimum units extracted from the character string image, The character cutout program according to claim 33, further comprising a process of statistically evaluating the same number of times as the number of the minimum units adjacent to the pattern to determine whether to output the combination pattern as a character cutout candidate. 36. In the character cutout candidate output processing, the combination pattern is further referred to by referring to an image pattern representing the shape of the combination pattern and an image pattern representing the shape of the minimum unit adjacent to the combination pattern. 36. The character extraction program according to claim 34, further comprising a process of determining whether or not to output a character as a character extraction candidate. 37. A process for determining whether or not to output as a character segmentation candidate, comprising: (a) a process of creating a combination pattern having a combination number of 1 to N; a process of focusing on the combination pattern CPc of, (c) is connected to the CP _C l th the minimum unit PS
A process for focusing on the _Cl, the process for outputting determination obtained between the PS _Cl (d) and the CP _C, it is determined whether or not focusing on all l (e), are focused When there is no l, the process returns to (c). (F) It is determined whether all Cs have been focused on, and when there is a C that has not been focused on, the process returns to (b). 37. The character cutout program according to any one of claims 34 to 36, further comprising: g) outputting the combination pattern determined to be outputtable in (d) as a character cutout candidate.