JPH02255995A - Character segmenting method - Google Patents

Abstract

PURPOSE:To exactly segment a character from a character picture in an undetermined form by finding out a segmenting position from the value of peripheral distribution only when a candidate segmenting position does not exist at the time of character extraction. CONSTITUTION:A CPU 101 segments the character one by one from the character picture, which is read by a picture input device 102, in a RAM 105. At first, the standard character height of the character and the tickness of a line are estimated from the peripheral distribution in a row direction and next, a standard character interval, word interval and standard character width are estimated and the extraction of a word is executed. Then, the candidate character segmenting position is estimated. For the segmentation of the character in the extracted word, the like component of the character is extracted and simultaneously, the height and width of the character are extracted. When the width of the character exceeds an allowable size, the extraction of a contour is executed again. When the candidate character segmenting position does not exist, a candidate character segmenting range is determined according to the peripheral distribution in a direction vertical to the row direction and the extraction of the contour is executed again. Then, only the inside of the area is extracted and the character is segmented. Thus, the character picture in the undetermined form such as the character picture, in which the peripheral distribution is overlapped, etc., can be exactly segmented.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention provides a character recognition device that searches for a character area from a character image by inputting characters written on a sheet of paper as an image, and converts it into a code number. Concerning how to cut out characters.

[Conventional technology]

In recent years, rapid advances in character recognition devices have made it possible to automatically extract character areas from various document images, cut out and recognize individual characters, and automatically create character files. Various methods have been devised for cutting out characters.

For example, one commonly used method is to count the vertical peripheral distribution of extracted character lines.

For example, when the vertical peripheral distribution of extracted character lines as shown in FIG. 4(a) is counted, a distribution as shown in FIG. 4(b) is obtained. (FIG. 4(b) only shows whether or not a character image exists.) The value of the distribution makes it possible to know the location of a character, and it is possible to extract one character by one character. In addition, for characters whose marginal distributions overlap as shown in Figure 4 (c), if the extracted character lines are determined to be multiple characters, the character break is estimated based on the character pitch and the character is forcibly extracted. I was going.

[Problem to be solved by the invention]

However, when cutting out characters using this method, it is possible to accurately cut out characters if the target character image is a character image with a fixed pitch as shown in Figure 4(a), Character images as shown in (c), Figure 5 (a)
), it is impossible to accurately cut out the characters in a character image as shown in FIG. 11(a).

The character image in FIG. 4(c), the character image in FIG. 5(a), and the character image in FIG. 11(a) are all proportional characters with an indefinite character pitch. Therefore, the character position and character pitch cannot be estimated simply from the vertical peripheral distribution as in the conventional method.

A case will be described in which an attempt is made to cut out characters using the conventional method with reference to FIG. 4(c). Figure 4(C) shows the fourth
This is a character image with the same font as in Figure (a). Figure 4 (a
) has a constant character pitch, whereas FIG. 4(C) is a proportional character with an undefined character pitch. If we focus on the vertical peripheral distribution of the existing character image in Figure 4(d), it is almost the same as the peripheral distribution in Figure 4(a) in Figure 4(b) in most parts, but the marginal distribution of YO is different. The corresponding parts are different.

In FIG. 4(d), the marginal distribution of Y and the marginal distribution of O overlap. Therefore, from the peripheral distribution, Y and O are judged to be one character and are separated, or they are separated at a location different from the actual location. Now, if we judge that there are two characters and try to forcefully separate them into two characters, if we focus on the character pitch, we will separate the character image at 1201 shown in Figure 12, and if we focus on the shape of the peripheral distribution, we will get the following: In step 1202, the character images are separated. 1201.1202 No matter where the character image is separated, the character Y or O will be cut off in the middle and extracted together with another character, making it impossible to extract the character accurately.

Next, a case will be described in which characters are extracted by a conventional method from the character image 15 shown in FIG. 5(a). FIG. 5(a) shows the word m1llion, which is composed of seven letters. When the vertical peripheral distribution of the word is counted, as shown in FIG. 5(b), m, i, 1.
1. Marginal distribution of 5 characters of i 501.502.
503, 504, and 505 are connected to form one cluster of marginal distributions.

Therefore, it is impossible to easily estimate the end of a character from the marginal distribution. If you try to cut and paste characters based on the character pitch, characters such as i and 1 are half the standard character width, so there is a possibility that the two characters will be judged as one character and cut out together. big. Furthermore, if characters such as i and 1 are separated based on the shape of the peripheral distribution, it may be possible to successfully separate characters such as i, 1, etc., but there is a high possibility that the character m will be decomposed, and the reliability of character extraction is extremely low.

Also, in the case of the character image shown in FIG. 11(a), the fifth
As in the case shown in the figure, it is impossible to accurately cut out characters only from the shape of the peripheral distribution.

Therefore, the present invention is intended to solve the above-mentioned problems, and its purpose is to select one of character images in which adjacent characters touch each other, character images in which the peripheral distributions overlap, and character images in which the character pitch is not constant. The purpose is to provide a method for accurately cutting out a single character.

Means for Solving the Problems] In a character recognition device that reads a character image written on paper etc. using the optical image input means of the present invention and converts the characters in the input image data into a recognized single code number. The character extraction method is as follows: ■ Estimate the mark character height and line thickness of the characters from the peripheral distribution in the line direction, ■ Estimate the standard character spacing,
Word spacing and standard character width are estimated, and words are extracted. ■ Character extraction candidate positions are estimated from the peripheral distribution. ■ Characters in the extracted words are extracted by simultaneously extracting the contours of connected components of the characters. The character height and width are extracted, ■ If the character width exceeds the allowable size from the standard character width, the outline is extracted again within the range of the character extraction candidate position, ■ The character extraction candidate position If the character does not exist, determine the range of character extraction based on the peripheral distribution in the direction perpendicular to the line direction, extract the outline again, and extract the character by extracting only the inside of the area surrounded by the outline. It is characterized by

The standard character height of characters is determined by focusing on the shape of the peripheral distribution in the line direction.
It is characterized in that the width of the portion where the peripheral distribution suddenly changes and becomes larger is defined as a standard character height, and the minimum line width of a character line is estimated from the size of the standard character height.

It is characterized by estimating standard character spacing and word spacing by taking statistics on the size of portions where no characters exist in the peripheral distribution in the direction perpendicular to the line direction.

The present invention is characterized in that the standard character width is estimated by taking statistics on the size of the portion of the peripheral distribution in the direction perpendicular to the line direction where the value of the peripheral distribution is larger than the minimum line width.

The standard character width is estimated to be 75% of the standard character height based on the above statistics.
Characteristically, in an area larger than , the maximum value of the block closest to the standard character height is used.

The classification of clusters in the statistics is characterized in that if the distance between clusters is less than a certain value proportional to the standard character height, the clusters are considered to be the same cluster.

The present invention is characterized in that word positions are extracted by comparing the standard character spacing and word spacing with the peripheral distribution in the direction perpendicular to the line direction.

The present invention is characterized in that the center portion of each portion of the peripheral distribution in the direction perpendicular to the line direction where the value of the peripheral distribution is smaller than the minimum line width is set as a character extraction candidate position.

If it is determined that the character is a connected character based on the extraction character width, the character extraction candidate position is preferentially extracted.

When estimating the character extraction position from the value of the marginal distribution, 1
It is characterized by finding the smallest value of the marginal distribution near the width of 72 characters and around the width of 1 character.

To extract the area surrounded by the outline of the character, create an image area of the same size as the original WI image, draw the outline of the character in the image area, fill in the inside of the outline, and then take the common part with the original image. By doing so, only the target characters are extracted.

〔Example〕

The present invention will be described in detail below based on examples.

As shown in the block diagram of FIG.
Image human power device 102, recognized character display device 103, ROM
104 and RAM 105.

Hereinafter, the character cutting method of the present invention for cutting out each character from a character image read into the RAMI 05 by the image input device 102 will be explained in detail based on the flowchart shown in FIG.

The character image read into the RAMIO 3 by the image input device 102 is shown in FIG. 3(a). In the present invention, first, the marginal distribution in the row direction is counted. When the peripheral distribution in the row direction of the character image in FIG. 3(a) is counted, a peripheral distribution in the shape of 301 in FIG. 3(b) is obtained. Western characters are the third
It is composed of three types of characters: a character with a large character height corresponding to d in Figure (a), a character with a small character height equivalent to e, and a character with a large character height but a low character position corresponding to y. Therefore, if we count the peripheral distribution in the row direction of a character image made up of these three types of characters, it is usually shown in Figure 3(b).
The peripheral distribution has a shape like 301 shown in FIG. The peripheral distribution 301 is made up of three areas 311, 312, and 313, and depending on the character image, the area 312 or the area 313 may not exist. However, an area of 311 always exists, and the area has a width of 304
corresponds to the height of a small character (hereinafter referred to as standard character height). Therefore, it is possible to know the standard character height 304 by paying attention to the shape of the peripheral distribution. In addition, in printed European and American type, the following relationship usually holds between the standard character height and the thickness of the character line: standard character height:character line thickness≧16=1. Therefore, when focusing on the vertical peripheral distribution of a character image, the value of the peripheral distribution is 1/1 of the standard character height.
It can be determined that a portion with a value lower than 16 is a whisker portion of a character, or a portion where characters are in contact. Here, a value of 1/16 of the standard character height, which serves as a criterion for judgment, is determined as the minimum line width.

As the next step, standard character spacing, word spacing, and standard character width are determined from the vertical peripheral distribution of character lines. First, data is obtained as to whether or not the character mR exists when projected in a direction perpendicular to the character line. The character image is shown in Figure 4 (
In the case of a), the data indicating whether or not a character image exists, that is, the data projected in the direction perpendicular to the character image, is shown in FIG. 4(b). Area 401 is a portion where a character image exists,
A region 402 is a portion where no character image exists, that is, a portion corresponding to the character spacing. Part 4 where the character does not exist
If we take the statistics of 02, we get the result as shown in Figure 6(a). Similarly,
When the statistics of the projected data of the character image in FIG. 4(c) and the portion 402 where the character image does not exist in FIG. 4(d) are taken, the result is as shown in FIG. 6(b). The data indicated by the solid lines in Figures 6(a) and 6(b) indicate the data in Figures 4(b) and 4(d), respectively, but generally the character spacing histogram is The tendency is shown by the dotted lines shown in FIG. 6(a) and FIG. 6(b). The two 17 stogranos 1 can each be divided into two chunks.

It can be determined that one of the clusters is a cluster of character spacings, and the other is a cluster of word spacings. Therefore,
Character spacing technique 1 makes it possible to estimate standard character spacing and word spacing. Generally, a histogram of character spacing is a histogram like that shown in FIG. 6, but sometimes a histogram like that shown in FIG. 8 is obtained. There are many clusters of data in the histogram. When there are many lumps like this, classify the lumps using the following method. Printed Western characters are usually regular 1. <They are printed side by side.

]7 Therefore, the character spacing should be approximately equal, but sometimes the character spacing differs depending on the shape of the character. However, the spacing does not spread beyond a certain value proportional to the standard character height. L, Therefore, in the present invention, 1/16 of the standard character height is set to 1 of the distance between clusters.
．． The threshold value was set to 805. The threshold value 805 and the mass 8
11.812.813.814.815 interval 801.
Comparing 802.80 and 804 (from , -, mass and 2), it can be determined that 11 and 812 are the same mass, and that 813, 814, and 815 are the same mass. , standard character spacing and word spacing can be estimated.

Next, the standard character width of the character image is estimated. If the standard character width can be estimated, it will be a clue for determining whether the character is a connected character at time 17 of character extraction and determining the extraction candidate position Ii8. For the standard character width 111 constant, statistics for the jJ character area are taken, similar to the estimation of the standard character spacing and word spacing. If the character images are as shown in Figures 4(a) and 4(c), the character width statistics will be as shown by the solid lines in Figures 7(a) and 7(b),
−・Obtain data on trends shown by dotted lines for boat fishing. Figure 7 (
Since the data shown in a) has one block, the maximum value of the block can be set as the standard character width. Figure 7 (b
), in the data shown in cluster 702 and cluster 7
There are two clusters of 03. Since the standard character width of normally printed European and American characters is very close to the standard character height, it is possible to set the maximum value of the cluster 702 near the standard character height 701- as the standard character width. Also, as in the case of estimating character spacing, sometimes when the histogram consists of many clusters, in Figure 9, the distance between clusters is 901.9.
02.903 and 2 threshold value 904, it is possible to determine that the lumps 911, 912, and 913 are the same lump.

Furthermore, even if many lumps are classified using the threshold value between the lumps, three lumps as shown in FIG. 10 may be formed. The reason why the statistical prefectures are as shown in Figure 10 is because
This is because characters such as i and 1 are included in the character image. Therefore, in the present invention, in order not to judge the maximum value of 1, 1, etc.]001 as the standard character width, 75
Find the maximum value of the lump in the area larger than the % position 1004. By doing this, we were able to exclude 1's and clusters of 1's that were about half the size of the standard character width, making it possible to accurately estimate the standard character width. As described above, the character width of a normal character image can be estimated with ease, but there are character images as shown in FIG. 5(a) among the character images. The character image is composed of seven characters: m, i, 1゜1゜i, o, n, among which m, i, l,
1. The letter 5 of i is touching. Therefore, even if we count the vertical marginal distribution j2 to estimate the standard character width,
Estimation of standard character width is not possible. Therefore, in the present invention,
In printed European and American characters, the character line width is proportional to the standard character height1
．． Noting that the whisker part at the bottom seen in Figure 5(a) is thinner than the certain value,19, the peripheral portion in the vertical direction is i-th. 5(1)), a process is provided in which the marginal distribution is truncated with a value of 1j binary 508 proportional to the standard character height. Through this process, only the portion where the edge of the peripheral distribution is larger than a certain value 508 is determined to be a character area.1. When the character Rj area is displayed, it becomes 1, 511.512.51 as shown in Figure 5(e).
3.51.4.515.516.517 are m, respectively
i, I, 1. i, O.

6.51-1 to 517 indicating the character position and width of the mouth
Although the seven pieces of data ' are slightly smaller than the actual character width, they always contain the information necessary for characters, and it is possible to judge these areas as character areas. Therefore, by taking statistics on the character area of the data shown in FIG. 5(c), it is possible to estimate the character width of the mark.

Next, words are extracted from the character image. Words are extracted by comparing the standard character spacing and word spacing with the vertical peripheral distribution of the character line (if the character image is shown in Figure 4(a), the peripheral distribution is shown in Figure 4(b)). , it becomes possible to find the word interval from the size of the part where the character does not exist, and it becomes possible to extract the word from the character image.

Next, each character is extracted from the extracted word by extracting connected components, but there is another problem here. Assume now that the extracted word is a character image as shown in FIG. 5(a). In this case, when extracting connected components, m
The five characters i l l i are extracted. The extracted character must be determined to be a concatenated character based on its character width, and must be forcibly separated in the middle. The separation position is usually performed at a location where the value of the peripheral distribution is small near the standard character width. However, in the case of FIG. 5(a), characters such as m, i, and l are included. In such a character, i
Since the width of the letters `` and ``l'' is half of the standard character width, when two letters are consecutive, the width becomes one character and the two letters are cut out together. Since the value of is quite small, there is a very high possibility that it will be judged as a break between characters and be cut off midway. Therefore, the present invention calculates midpoints 521 to 526 between characters at character positions 511 to 517 in Figure 5(c) in a certain 1ii508 or higher area proportional to the standard character height in the peripheral distribution in Figure 5(b). . Use this as a candidate position for cutting out. As is clear from FIG. 5, the cutout candidate positions 521 to 526 represent boundaries between characters. Therefore, when cutting out a character, if the size of the character is larger than the allowable size from the standard character width, it is possible to cut out the character by using this position, and the number of characters cut out at the wrong position is greatly reduced. do.

Next, extract each character one by one. In the present invention, characters are extracted by extracting connected components. Most European and American characters have one connected component, and even for characters that do not have one connected component, most characters can be determined based only on the main connected component, so this method is effective. To extract connected components, first, the contours of connected components are extracted. By extracting the outline, information on the position and size of the character can be obtained. When the outline is extracted, the extracted character width is compared with the already determined standard character width.

If the extracted character width is larger than the standard character width, one extracted character and one I11 cut character are extracted. If the extracted character width is larger than the standard character width, first, it is checked whether or not there is a cutting candidate position in the extraction area. If a cutout candidate position exists, it is most appropriate to cut out the character at that position, so the outline of the character of the connected component is extracted again within the range of this position. In addition, the character image is shown in Figure 11 (a
), (In this character image, there are only five midpoints, 1121 to 1125, of areas where the peripheral distribution is smaller than the constant size 1110 for the standard character height. However, (There are nine characters.) When the outline of the character image 1152 is extracted, it is determined that the character width is unacceptable compared to the standard character width.

Therefore, the existence of the extraction candidate position is confirmed. However, the peripheral distribution 1142 of the character image 1152 has a minimum line width of 1110
Since there is no value smaller than , there is no candidate cutting position in the middle of the character.

In this case, attention is paid to values near the standard character width of the peripheral distribution 1142 of extracted characters. As a result, it becomes possible to search for the contact part 1131 between the letters p and ", and it becomes possible to cut out one character after another. In the same way, cutting out is repeated, and in FIG. In addition to the positions, it is possible to extract three cutout positions from 1131 to 1133, and to accurately cut out nine characters.Furthermore, the present invention extracts the contours of the connected components of the characters, and then extracts the extracted character width. Since this method extracts characters when the width is one character, it is possible to extract characters without any problem even in a character image such as that shown in FIG. 12(a). However, in the marginal distribution diagram 12(b), the marginal distributions of Y and O overlap and form one marginal distribution 120.
3 is formed. Here, as in the conventional case, the marginal distribution 120
If you try to forcibly separate the character from the shape of 3, it will be 120
If you try to cut it off due to the size of the peripheral distribution, you will end up cutting it off at the position 1201. Regardless of the position, the result is that the characters are cut off at an inappropriate position, but this does not cause any problems in the present invention.

After the contours of the connected components of the characters have been extracted at an appropriate size, the connected components of the characters are extracted. In the present invention,
Another area with the same size as the character image is provided. 1st
To extract the character P from the character image 1301 in FIG. 3(a), first the outline of the character P in FIG. 13(a) is drawn in another area 1302. Next, the area surrounded by the outline is filled in to obtain FIG. 13(b). After this, by extracting the common part between the original image in FIG. 13(a) and the developed image in FIG. 13(b) of the extracted character area, FIG.
) to obtain the image shown in ). As can be seen from the image in FIG. 13(C), from the images of P and e that overlap in position, P
It is now possible to clearly extract only the images.

As described above, according to the present invention, the mark character height and minimum line width are determined from the peripheral distribution in the line direction, and standard character spacing, word spacing, and mark character length are determined with reference to these values, so accurate Find standard character spacing, word spacing, and standard character length. As a result, accurate word extraction becomes possible. Furthermore, when cutting out characters, characters are cut out with reference to the cutting candidate position and the peripheral portion m while comparing with the standard character width, making it possible to cut out characters at accurate positions.

Furthermore, character extraction I7 extracts an image of an area surrounded by the outline of connected components of characters, so it is possible to accurately extract only one character from character images whose marginal distribution overlaps and the extraction position is unknown. It became.

As described above, the present invention has made it possible to more accurately cut out individual characters, which had been considered difficult up until now, and has greatly improved the reliability of character recognition devices that use this method as a component. ”It became possible to do this.

Furthermore, since the present invention uses standard character height and extraction of connected components of characters, it is particularly suitable for printing European and American characters.

〔Effect of the invention〕

As described below, the present invention has many of the following effects and greatly improves the reliability of character recognition devices.

By counting the peripheral distribution in the line direction (2. By estimating the standard character height and minimum line width from the shape of the peripheral distribution, it is possible to more accurately determine the character area in the direction perpendicular to the line direction, and it is possible to cut out characters. This makes it possible to find candidate positions.

It also made it possible to accurately determine standard character spacing, word spacing, and standard character length.

By taking statistics on the areas where no characters are present in the peripheral distribution in the direction perpendicular to the line direction, accurate character spacing and word spacing can be determined, making it possible to extract accurate words.

In the peripheral distribution in the direction perpendicular to the line direction, the accurate standard character length can be determined by taking statistics on the area below a certain size proportional to the standard character height, and by setting the center of the area as the candidate cutting position. , it is possible to separate characters at precise positions.

When estimating the standard character width, by finding the maximum value of the area larger than 75% of the standard character height,
This eliminates the possibility of accidentally selecting a block of characters with a character width of 215 minutes, and if the interval between the blocks is proportional to the standard character height and is less than a certain value, it is considered to be the same block, so it is accurate. Clusters can be classified, and as a result, standard character width, standard character spacing, and word spacing can be accurately estimated.

Character extraction [2] extracts only the area inside the outline of the character, so it is possible to accurately extract only the target character even in a character image where the extraction position is unknown in the vertical peripheral distribution. It is.

When extracting characters, if the extracted character width exceeds the allowable standard character width, the candidate position for extraction is first checked, and only if there is no candidate position for extraction, the extraction position is determined from the value of the peripheral distribution. Therefore, it is now possible to extract characters more accurately even in character images that would be separated at the wrong position using only the marginal distribution values.

As described above, the present invention makes it possible to accurately extract individual characters from character images where adjacent characters touch, from character images where the peripheral distributions overlap, and from character images where the character pitch is not constant. did. As a result, it has the effect of greatly improving the reliability of a character recognition device that uses the character segmentation method as a component.

[Brief explanation of drawings]

FIG. 1 shows a block diagram of a character recognition device in which the character cutting means of the present invention is used. FIG. 2 shows a flowchart of the character cutting means of the present invention. FIGS. 3(a) and 3(b) are diagrams illustrating the standard character height extraction method of the present invention. Figures 4 (a) to (d), Figures 5 (a) to (c), and 6
Figures (a) and (b), Figures 7 (a) and (b), Figures 8, 9, and 10 show how the mark character spacing, word spacing, standard character length, and extraction candidate positions are estimated. The illustrated diagram is shown. Figure 11 (a) (b), Figure 12 (a) (b),
FIGS. 13(a) to 13(C) are diagrams illustrating how characters are extracted. 101 ・ 102 Φ 103 ・ 104 ・ 105 ・ 301 ・ 304 ・ 401 ・ 402 ・ ・ CPU ・ Image input device ・ Recognized character display device ・ ROM φRAM ・ Line direction peripheral distribution ・ Standard character height ・ Character area ・ Character spacing 501 to 507 511-517 521-526 508・ ・ ・ ・ 701 Salary ・ ・ ・ 801-804 805φ ・ Φ ・ 901-903 904◆ ・ ・ ・ 1004 Fist ・ ・ 1110 Fight φ ・ 1131-11 ・Character peripheral distribution・Character area・Character cutting candidate position, minimum character line width, standard character height, cluster spacing, cluster spacing threshold, cluster spacing, cluster spacing threshold, 75% of standard character height, minimum character line width, Character cutting position/Character cutting position and above Applicant: Seiko Epson Co., Ltd. Agent Patent attorney Masayoshi Kamiyanagi (and 1 other person) Figure 2 Figure 4 (α) Figure 4 Figure 4 (d) 3+11 Figure 3 ( Figure 3 (G) [Su customers [1%] Figure 6 (^ Figure 6 (G) Figure 3 and Figure 8 [YJF size] Figure 7 (α2 [YJF size]) Figure 7 Figure (Tonneau Figure 9 190ψ Figure 10 ↑ Cow 7 = / 10 Yu ↑ ↑ ↑

Claims (11)

[Claims] (1) The character cutting method in a character recognition device that reads a character image written on paper etc. using an optical image input means, recognizes the characters in the input image data, and replaces them with a code number is as follows: [1] In the line direction. Estimate the standard character height and line thickness of the characters from the marginal distribution, and [2] Estimate the standard character spacing, word spacing, and standard character width from the peripheral distribution in the direction perpendicular to the line direction, and extract words. , [3] Estimate character extraction candidate positions from the marginal distribution; [4] Extract characters in the extracted word by extracting outlines of connected components of characters, and at the same time extracting character height and character width; [5] ] If the character width exceeds the allowable size from the standard character width, the outline is extracted again within the range of the character extraction candidate position. [6] If the character extraction candidate position does not exist, the outline is extracted in the line direction. A character is characterized in that the range of character extraction is determined based on the peripheral distribution in the direction perpendicular to the character, and the outline is extracted again, and [7] the character is extracted by extracting only the inside of the area surrounded by the outline. Cutting method. (2) The standard character height of a character is determined by focusing on the shape of the peripheral distribution in the line direction, and determining the width of the part where the peripheral distribution changes rapidly and becomes larger as the standard character height, and from the size of the standard character height, 2. The character cutting method according to claim 1, further comprising estimating a minimum line width of a character line. (3) The character segmentation method according to claim 1, characterized in that the standard character spacing and word spacing are estimated by taking statistics on the size of the portion where no characters are present in the peripheral distribution in the direction perpendicular to the line direction. . (4) The standard character width is estimated by taking statistics on the size of a portion of the peripheral distribution in the direction perpendicular to the line direction where the value of the peripheral distribution is larger than the minimum line width. How to cut out characters as described. (5) Character cutting according to claim 1, characterized in that the standard character width is estimated by using the maximum value of the cluster closest to the standard character height in an area larger than 75% of the standard character height according to the statistics. Method. (6) The character segmentation method according to claim 1, wherein in the classification of clusters in the statistics, if the distance between the clusters is less than or equal to a certain value proportional to the standard character height, the clusters are considered to be the same cluster. . (7) The character segmentation method according to claim 1, wherein the word position is extracted by comparing the standard character spacing and word spacing with a peripheral distribution in a direction perpendicular to the line direction. (8) The character according to claim 1, wherein the center portion of each portion of the peripheral distribution in the direction perpendicular to the line direction where the value of the peripheral distribution is smaller than the minimum line width is set as a character extraction candidate position. Cutting method. (9) The character extraction method according to claim 1, wherein when it is determined that the extracted character is a connected character based on the width of the extracted character, the character extraction candidate position is cut out preferentially. (10) Character cutting according to claim 1, characterized in that when estimating the character cutting position from the value of the marginal distribution, a point with the smallest value of the marginal distribution is found near 1/2 character width and near 1 character width. Method. (11) To extract the area surrounded by the outline of a character, create an image area of the same size as the original image, draw the outline of the character in the image area, fill in the inside of the outline, and then extract the common area with the original image. 2. The character extraction method according to claim 1, wherein only target characters are extracted by taking .