JP2728085B2 - Character extraction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character extracting method used in a character recognition apparatus for searching a character area from a character image and converting it into a code number by inputting a text written on paper as an image. .

[0002]

2. Description of the Related Art In recent years, with the rapid progress of a character recognition device, a character area can be automatically extracted from various document images, and each character can be cut out and recognized, and a character file can be automatically created. As a result, various methods for extracting characters have been devised.

[0003] For example, a method that is generally used is a method of counting the marginal distribution of an extracted character line in a direction perpendicular to the line direction.

For example, a peripheral distribution in a direction perpendicular to the line direction of an extracted character line as shown in FIG. 4A is counted. In this counting, if only the presence or absence of the count value, that is, whether or not a character image exists, is shown in the figure, a peripheral distribution as shown in FIG. 4B is obtained. , It is possible to know the position of the character, and the character can be cut out one by one. However,
As shown in FIG. 4C, in an extracted character line, in a character whose peripheral distribution overlaps, the character width of the overlapping character portion becomes larger, such as “Y” and “o”. Therefore, when a plurality of characters are determined, character cuts are estimated based on the character pitch, and characters are forcibly cut out.

[0005]

However, when a character is cut out by such a method, if the target character image is a character image having a constant pitch as shown in FIG. Is possible, but FIG.
In the character image as shown in FIG. 5C, the character image as shown in FIG. 5A, and the character image as shown in FIG. FIG.
(C) Character image, FIG. 5 (A) character image, FIG.
The character images in (A) are all proportional characters with an unfixed character pitch. Therefore, the character position and the character pitch cannot be estimated only by the marginal distribution in the direction perpendicular to the line direction as in the conventional method.

A case in which a character as shown in FIG. 4C is cut out by a conventional method will be described.
FIG. 4C is a character image in the same font as FIG. 4A. FIG. 4A shows a proportional character having a fixed character pitch, while FIG. 4C shows a proportional character having a variable character pitch. Paying attention to FIG. 4D, which shows the peripheral distribution in the direction perpendicular to the line direction of the character image in FIG. 4C, FIG. 4B, which is the peripheral distribution of FIG. Although they are almost the same, the part corresponding to the peripheral distribution of “Yo” is different. In FIG. 4D, the peripheral distribution of “Y” and the peripheral distribution of “o” overlap. Therefore, from such a peripheral distribution, "Y" and "o" are determined as one character, or separated at a place different from the actual one. Now, if it is supposed that two characters are determined and forcibly divided into two characters, focusing on the character pitch, the character image is separated at 1201 shown in FIG. Will separate the character image. Regardless of the position of the character image at 1201 or 1202, the character image is cut off in the middle of the character "Y" or "o" and is extracted together with another character. is there.

Next, from the character image shown in FIG.
A case where characters are cut out by a conventional method will be described. FIG. 5A is a word composed of seven characters “million”. When the peripheral distribution of the word in the direction perpendicular to the row direction is counted and the count value is illustrated, as shown in FIG. 5B, “m”, “i”, “l”,
Marginal distributions 501 and 50 of five characters “l” and “i”
2, 503, 504, and 505 are connected to form one peripheral distribution block. Therefore, it is impossible to easily estimate a character break from this peripheral distribution.
If the characters are separated based on the character pitch, characters such as "i" and "l" are half the standard character width, so that two characters are determined to be one character. There is a high possibility of being cut out. Also, if the character is separated based on the shape of the marginal distribution, characters such as “i” and “l” may be separated well, but the possibility of decomposing “m” is large. Very small.

[0008] Also, in the case of the character image shown in FIG. 11A, it is impossible to cut out the character accurately only from the shape of the marginal distribution as in the case of FIG.

An object of the present invention is to solve the above-mentioned problems, and an object of the present invention is to provide a method for accurately cutting out individual characters from a character image in which adjacent characters are in contact with each other. is there.

[0010]

According to a first aspect of the present invention, there is provided a character recognition apparatus for reading a Western character image written on paper or the like by an optical image input means and recognizing a character in the input image data. In the character segmentation method, the standard character height of the character is estimated from the peripheral distribution in the line direction, and the value of the line width to be ignored is determined based on the standard character height, and the marginal distribution in the direction perpendicular to the line direction is determined. The standard character width is estimated by taking statistics of a portion where the value of the marginal distribution is larger than the value of the line width to be ignored.

According to a second aspect of the present invention, in the character segmenting method according to the first aspect, the standard character width is estimated in the area where the value of the marginal distribution is substantially larger than the half of the standard character height by the statistics. It is characterized by doing.

According to a third aspect of the present invention, in the character segmenting method according to the second aspect, the distribution of the blocks in the statistics is such that a distance between the blocks is equal to or less than a predetermined value proportional to the standard character height. In this case, they are regarded as the same block.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below in detail based on embodiments.

An example of the outline of the whole process of character extraction is a character recognition in which a character image written on paper or the like is read by an optical image input means, and characters in the input image data are recognized and replaced with a code number. It is used in equipment, and estimates the standard character height of characters from the marginal distribution in the line direction and determines the thickness of lines to be ignored. Estimate the standard character spacing, word spacing and standard character width from the peripheral distribution in the direction perpendicular to the line direction, extract words, estimate character cutout candidate positions from the peripheral distribution, The cutout extracts the outline of the connected component of the character, and also extracts the character height and the character width. If the character width exceeds the allowable size from the standard character width, the character is cut again within the range of the character cutout candidate position. If the contour is extracted and the character extraction candidate position does not exist,
The character extraction range is determined based on the peripheral distribution in the direction perpendicular to the line direction, and the outline is extracted again. The character is extracted by extracting only the inside of the area surrounded by the outline.

Focusing on the shape of the peripheral distribution in the line direction, the standard character height of a character is defined as the standard character height, where the width of the portion where the peripheral distribution changes rapidly and becomes large is defined as the standard character height. The minimum line width of the character line can be estimated.

The standard character spacing and word spacing can be estimated by taking statistics on the size of the portion where no character exists in the marginal distribution in the direction perpendicular to the line direction.

The standard character width can be estimated by taking statistics on the size of the marginal distribution in the direction perpendicular to the line direction and the value of the marginal distribution larger than the minimum line width.

The standard character width can be determined to be the maximum value of the block closest to the standard character height in an area larger than 75% of the standard character height in the aforementioned statistics.

As for the classification of the blocks in the statistics, if the distance between the blocks is equal to or less than a certain value proportional to the standard character height, the blocks can be regarded as the same block.

The position of a word can be extracted by comparing the standard character spacing and word spacing with the marginal distribution in the direction perpendicular to the line direction.

In the peripheral distribution in the direction perpendicular to the row direction, each central portion of the portion where the value of the peripheral distribution is smaller than the minimum line width can be set as a character extraction candidate position.

When the character is determined to be a concatenated character from the extracted character width,
The character extraction candidate position can be extracted with priority.

When estimating the character cut-out position from the value of the marginal distribution, it is possible to find the part where the marginal distribution value is the smallest near the half character width and around the l character width.

To extract an area surrounded by the outline of a character, an image area having the same size as the original image is provided, the outline of the character is drawn in the image area, and the inside of the outline is painted. , Only the target character can be extracted.

As shown in the block diagram of FIG. 1, a character recognition device in which the character extracting means is a constituent
1, image input device 102, recognized character display device 103, R
The OM 104 and the RAM 105 are arranged horizontally.

In the following, a RAM 1
A character extracting method for extracting individual characters from the character image read in 05 will be described in detail with reference to a flowchart shown in FIG.

FIG. 3A shows a character image read into the RAM 105 by the image input device 102. First, the peripheral distribution in the row direction is counted. Character image figure 3 (A)
When the peripheral distribution in the row direction is counted, 301 in FIG.
Is obtained. Western characters are shown in Figure 3
A character with a large character height corresponding to "d" in (A), "e"
, And three types of characters having a large character height corresponding to “y” but a low character position. Therefore, when the peripheral distribution in the row direction of a character image composed of these three types of characters is counted,
Normally, a peripheral distribution having a shape like 301 shown in FIG. The marginal distribution 301 is formed by combining three regions 311, 312, and 313, and depending on the character image,
There are cases where the area 312 or the area 313 does not exist. However, a region 311 always exists, and a width 304 between the upper limit 302 and the lower limit 303 of this region corresponds to the height of a character having a small character height (hereinafter, referred to as “standard character height”). Therefore, it is possible to know the standard character height by focusing on the shape of the peripheral distribution in the row direction.

Further, for printed European and American prints, the relationship of standard character height: character line thickness ≧ 16: 1 is usually established between the standard character height and the character line thickness. Therefore, when focusing on the vertical distribution of the character image, a portion where the value of the peripheral distribution is lower than 1/16 of the standard character height is a whisker portion of the character,
Alternatively, it can be determined that the part is in contact with the character. Here, a value of 1/16 of the standard character height serving as a criterion for determination is determined as a line width to be ignored.

In the next step, the standard character spacing, word spacing and standard character width are determined from the vertical marginal distribution of the character line. First, when the image is projected in a direction perpendicular to the character line, data as to whether or not a character image exists is obtained. When the character image is shown in FIG. 4A, data indicating whether or not a character image exists, that is, data projected in a direction perpendicular to the character image is shown in FIG. 4B. An area 401 is a part where a character image exists, and an area 402 is a part where a character image does not exist, that is, a part corresponding to a character interval. FIG. 6A shows the statistics of the portion 402 where no character exists. Similarly, the projected data when the character image is shown in FIG. 4C. Portion 402 where the character image in FIG. 4D does not exist.
6B is obtained as shown in FIG. FIG.
(A) and the data shown by the solid line in FIG.
4B shows the data of FIG. 4D, and the histogram of the character spacing tends to be generally indicated by the dotted lines shown in FIGS. 6A and 6B. These two histograms can each be divided into two blocks. One of the blocks can be determined to be a block of character spacing, and the other can be a block of word spacing. Therefore, the standard character spacing and word spacing can be estimated from the character spacing statistics. Generally, the histogram of the character spacing is a histogram as shown in FIG. 6, but sometimes a histogram as shown in FIG. 8 is obtained. A histogram contains a number of chunks of data. When there are many such clusters, the clusters are classified by the following method. Printing Western characters are usually printed side by side.

Therefore, the character spacing should be almost equal, but sometimes the character spacing differs depending on the shape of the character. However, the interval is not scattered beyond a certain value proportional to the standard character height. Therefore, 1/16 of the standard character height is set as the threshold 805 of the distance between blocks. Threshold 805 and masses 811, 81
2,801,814,815 interval 801,802,8
By comparing the data with 03,804, the masses 811 and 812 are the same, and the masses 813, 814, and 815 are the same.
Can be determined to be the same block, and the standard character interval 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 becomes a clue to determine whether a character is a single character or a concatenated character at the time of character extraction, and to estimate a candidate extraction position. The estimation of the standard character width takes the statistics of the character area in the same manner as the estimation of the standard character interval and the word interval. If the character image is as shown in FIGS. 4 (A) and 4 (C), the statistics of the character width are as shown by the solid lines in FIGS. 7 (A) and 7 (B), and generally indicated by dotted lines. Get Day Evening. Since the data shown in FIG. 7A has one block, the maximum value of the block can be used as the standard character width. In the data shown in FIG. 7B, there are two blocks, a block 702 and a block 703. Normally, for printed Western characters, the standard character width is very close to the standard character height.
02 becomes the standard character width. Also, as in the case of estimating the character spacing, if the histogram sometimes consists of many clusters, FIG. 9 shows the distances 901, 902, 903 between clusters and the threshold 90.
By comparing 4 with lump 9 lll and lump 9
12 and the mass 913 can be determined to be the same mass. Further, even if many clusters are classified based on the threshold value between clusters, three clusters as shown in FIG. 10 may occur. The statistical result is as shown in FIG. 10 because characters such as "i" and "l" are included in the character image.

Therefore, the lump 100 such as "i" or "l"
In order to prevent the maximum value of 1 from being determined as the standard character width, the maximum value of the block is searched for in an area larger than the position 1004 at 75% of the standard character width. By doing so, it is possible to exclude a cluster of “i” and “l” which are about half the width of the standard character width, and to accurately estimate the standard character width. As described above, the character width of a normal character image can be estimated, but a character image as shown in FIG. 5A exists in the character image. The character images are "m", "i",
It is composed of seven characters “l”, “l”, “i”, “o”, and “n”, of which “m”, “i”,
The five characters “l”, “l”, and “i” are in contact. Therefore, even if the peripheral distribution in the direction perpendicular to the line direction is counted for estimating the standard character width, the standard character width cannot be estimated.

Therefore, in the present invention, in printed European and American characters, the character line width has a thickness greater than or equal to a certain value in proportion to the standard character height, and the bottom beard portion shown in FIG.
Focusing on the line being thinner than a certain value, counting the marginal distribution diagram 5 (B) in the direction perpendicular to the line direction, and determining the line width to ignore with a certain value 508 proportional to the standard character height , And a process of cutting off the marginal distribution. According to this process, only a portion where the value of the peripheral distribution becomes larger than a certain value 508 is determined as a character region, and the character region is displayed as shown in FIG. 5C. 511, 512, 513,
514,515,5l6,5l7 are "m",
The character positions and character widths of “i”, “l”, “l”, “i”, “o”, and “n” are shown. The seven data 511 to 517 are slightly smaller than the actual character width, but always include information necessary for characters, and these areas can be determined as character areas. Therefore, it is possible to estimate the standard character width by collecting the statistics of the character area shown in FIG. 5C.

Next, words are extracted from the character images.
In extracting words, the standard character spacing and word spacing are compared with the vertical marginal distribution of character lines. When the character image is as shown in FIG. 4A, since the peripheral distribution is as shown in FIG. 4B, it is possible to find the word interval from the size of the portion where the character does not exist. Word extraction becomes possible.

Next, the extraction of each character from the extracted word is performed by extracting the connected component, but there is another problem here. Now, assume that the extracted word is a character image as shown in FIG. In this case, when the connected component is extracted, the five characters “milli” are extracted as one character. This extracted character is determined to be a concatenated character based on its character width, and must be forcibly cut off halfway. The separation position is usually set at a position where the value of the peripheral distribution near the standard character width is small. However, FIG.
In the case of (A), characters such as “m”, “i”, and “l” are included. In such a character, the characters "i" and "l" have half the width of the standard character width. Therefore, if two characters continue, the width becomes one character and two characters are cut out together. In the case of "m", the value of the marginal distribution once becomes very small in the vicinity of the standard character width, so that it is determined that the character is a break, and it is very likely that the character will be cut off in the middle.

Therefore, in the peripheral distribution of FIG. 5B, an area in which the value is equal to or more than a certain value 508 in proportion to the standard character height, the characters at character positions 511 to 517 in FIG.
6 is determined as a candidate clipping position. The cutout candidate positions 521 to 526 represent boundaries between characters as is clear from FIG. Therefore, when cutting out a character, if the character size is larger than the allowable size from the standard character width, it is possible to cut out the character by using such a cutout position, and to cut out the character at an incorrect position. Is extremely diminished.

Next, characters are extracted one by one. Character extraction is performed by extracting connected components. This method is effective because most of the Western characters have one connected component, and even if the character has no connected component, most characters can be determined only by the main connected component. In the extraction of the connected component, first, the outline of the connected component is extracted. By extracting the contour, information on the position and size of the character can be obtained. When the outline is extracted, the standard character width already obtained is compared with the extracted character width.

If the extracted character width is larger than the standard character width, the extracted character is determined to be one character and one character is extracted. If the extracted character width is unacceptable from the standard character width, first, it is checked whether or not there is a cutout 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 the position.

In the case where the character image is a character image as shown in FIG. 11A (in this character image, the middle point of the area whose peripheral distribution is smaller than a fixed size 110 with respect to the standard character height is There are only five positions 1121 to 1125. However, there are nine characters.)
When the outline of the character image 1152 is extracted, the character width is determined to be unacceptable from the standard character width. Therefore, the existence of the extraction candidate position is confirmed. However, since the marginal distribution 1142 of the character image 1152 does not have a value smaller than the ignored line width 1110 in the middle of the character, there is no cutout candidate position. In this case, the marginal distribution 114 of the extracted character
Note the value near the standard character width of 2. As a result, it is possible to search for the contact portion 1131 between the characters “a” and “r”, and it is possible to cut out characters one by one. The cutout is repeated in the same manner, and in FIG.
In addition to the five extraction candidate positions of 1-125, 11
It is possible to extract three cutout positions 31 to 1133 and to cut out nine characters accurately.

After extracting the outline of the connected component of the character, if the extracted character width is one character width, the character is extracted. For example, a character image as shown in FIG. Can extract characters without any problem. This character image is composed of three characters "Y", "o", and "u".
Contactless characters. However, marginal distribution diagram 1
In FIG. 2B, the peripheral distributions of “Y” and “o” overlap to form one peripheral distribution 1203. Here, if the character is forcibly separated from the shape of the peripheral distribution 1203 as in the related art, the character is separated at the position 1202. In addition, if an attempt is made to separate according to the size of the peripheral distribution, the separation is performed at the position of l201. Either position will result in the characters being cut off at inappropriate positions, but this method does not pose any problems.

After the outline of the connected component of the character is extracted with an appropriate size, the connected component of the character is extracted. Another area of the same size as the character image is provided. FIG.
In extracting the character “P” from the character image 1301 of FIG. 3A, first, the outline of the character “P” in FIG. Next, the area surrounded by the outline is painted out to obtain FIG. Thereafter, a common part between the original image FIG. 13 (A) and the extracted character area image FIG. 13 (B) is extracted, whereby FIG.
The image shown in FIG. 3 (C) is obtained. As can be seen from FIG. 13C, it is possible to cleanly extract only the “P” image from the “P” and “e” images that overlap in position.

As described above, according to the present invention, the standard character height and the line width to be ignored are obtained from the marginal distribution in the line direction, and the standard character spacing, word spacing, and standard character width are calculated with reference to these values. The exact standard character spacing, word spacing, and standard character width are determined. As a result, accurate word segmentation becomes possible. Also, at the time of character extraction, characters are extracted with reference to the extraction candidate position and the surrounding distribution while comparing with the standard character width, so that it is possible to extract characters at accurate positions. Furthermore, since character extraction extracts an image of a region surrounded by the outline of a connected component of a character, it is possible to accurately extract only one character even from a character image where the extraction position is not known due to overlap in the peripheral distribution. became.

As described above, according to the present invention, it has become possible to more accurately cut out characters one by one, which has been difficult until now. It became possible to greatly improve.

Further, since the present invention uses the extraction of the standard character height and the connected components of the characters, the present invention is particularly suitable for printed Western characters.

[0045]

As described above, according to the present invention, the peripheral distribution in the line direction is counted, the standard character height is estimated from the shape of the peripheral distribution, and the line width to be ignored is determined. In the marginal distribution in the direction perpendicular to the direction, by taking statistics of a part where the value of the marginal distribution is larger than the value of the line width to be ignored, characters connected at the whiskers are separated and the standard character width is estimated. be able to.

According to the statistics, in a region where the value of the marginal distribution is larger than almost half of the standard character height,
By estimating the standard character width, it is possible to separate characters at more accurate positions.

Further, the distribution of blocks in the statistics can be more reliably cut out by regarding the blocks as being the same block if the distance between the blocks is equal to or less than a predetermined value proportional to the standard character height.

[Brief description of the drawings]

FIG. 1 is a block diagram of a character recognition device using a character cutout unit of the present invention.

FIG. 2 shows a flowchart of a character segmenting means of the present invention.

FIGS. 3A and 3B are diagrams illustrating a method for extracting a standard character height according to the present invention.

FIGS. 4A to 4D are diagrams illustrating a state of estimating a standard character interval and a word interval, a standard character width, and a cutout candidate position according to the present invention.

FIGS. 5A to 5C are diagrams illustrating a state of estimating a standard character interval and a word interval, a standard character width, and a cutout candidate position according to the present invention.

FIGS. 6A and 6B are diagrams illustrating a state of estimating a standard character interval and a word interval, a standard character width, and a cutout candidate position according to the present invention.

FIGS. 7A and 7B are diagrams illustrating a state of estimating a standard character interval and a word interval, a standard character width, and a cutout candidate position according to the present invention.

FIG. 8 is a diagram illustrating a state of estimating a standard character interval and a word interval, a standard character width, and a cutout candidate position according to the present invention.

FIG. 9 is a diagram illustrating a state of estimating a standard character interval and a word interval, a standard character width, and a cutout candidate position according to the present invention.

FIG. 10 is a diagram illustrating a state of estimating a standard character interval and a word interval, a standard character width, and a cutout candidate position according to the present invention.

FIGS. 11A and 11B are diagrams illustrating a state of character extraction according to the present invention.

FIGS. 12A and 12B are diagrams illustrating a state of character extraction according to the present invention.

FIGS. 13A to 13C are diagrams for explaining the character extraction according to the present invention.

[Explanation of symbols]

101: CPU, 102: image human input device, 103: recognition character display device, 104: ROM, 105: RAM, 30
1 ... line direction peripheral distribution, 304 ... standard character height, 401 ... character area, 402 ... character spacing, 50l to 507 ... character peripheral distribution, 511 to 517 ... character area, 521 to 526 ... character cutout candidate position, 508 ... lowest Character line width, 70 l ...
Standard character height, 80l to 804 ... Lump spacing, 805 ...
Lump interval threshold, 90l to 903: lump interval, 904 ... lump threshold, 1004: 75% of standard character height, 1110: minimum character line width, 1131
1133: character extraction position, 1201 to 1202: character extraction position.

Claims (3)

(57) [Claims] 1. A character cutout method for a character recognition device which reads a Western character image written on paper or the like by an optical image input means and recognizes characters in input image data. The standard character height is estimated, and the value of the line width to be ignored is determined based on the standard character height.In the peripheral distribution in the direction perpendicular to the line direction, the value of the peripheral distribution is the value of the line width to be ignored. A character extraction method characterized by estimating a standard character width by collecting statistics of a portion larger than a value. 2. The character cutout method according to claim 1, wherein a standard character width is estimated in an area where the value of the marginal distribution is substantially larger than half of the standard character height by the statistics. 3. The distribution of chunks in the statistics is regarded as the same chunk if the distance between chunks is equal to or less than a predetermined value proportional to the standard character height. Character extraction method.