JP2768289B2 - Character segmentation device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to character string recognition, and more particularly to an apparatus for extracting characters from a character string image.

[0002]

2. Description of the Related Art When extracting characters from a handwritten character string,
Since the variation in the feature amount such as the size and width of the character and the character spacing is large, it is difficult to cut out one character at a time using only such a feature amount. In order to solve this problem, there is disclosed a method in which a character string is first divided into the finest patterns, and all possible combinations of the finest patterns are obtained and used as a plurality of character cutout candidates (Japanese Patent Laid-Open No. Sho 62).
-190574).

As a method of extracting the finest pattern, a method of using a connected region of black pixels as a basic pattern is known. As another method for extracting the finest pattern, a patent specification (Japanese Patent Application Laid-Open No. 4-363780) filed by the same applicant as the present application is provided so that a divided pattern can be extracted even when characters are in contact with each other. As described in Japanese Patent Application Laid-Open No. H10-157, there is known a method in which a valley of a projection function obtained by counting black pixels in a character string image in a direction perpendicular to the character string direction is cut out as a candidate line.

[0004]

However, the method of extracting the finest pattern described above has a problem that a contacted pattern cannot be separated. Further, the method described in Japanese Patent Application Laid-Open No. 4-363780 has a problem that the number of finest patterns is greatly increased, and a problem that a cutout position is inaccurate. Further, the method described in Japanese Patent Application Laid-Open No. 62-190574 has a problem that the number of character cutout candidates becomes very large, and the number of times of performing character recognition processing becomes large, resulting in a very slow processing speed.

Therefore, the present invention can determine the correct cutout position even if there is a contact, and reduce the number of character cutout candidates while maintaining the rate at which the output character cutout candidates include correct characters. It is an object of the present invention to provide a character cutout device capable of shortening the processing time.

[0006]

According to a first aspect of the present invention, there is provided a character extracting apparatus for inputting a character string image, and a character element block for extracting a character element from the character string image input by the image input section. An extracting unit, a standard size estimating unit that estimates a standard block size based on the character element block extracted by the character element block extracting unit, and a standard block size calculated by the standard size estimating unit. A normalization block calculation unit that calculates a normalization block by normalizing the character element block; a contact data storage unit that stores typical data of the contacted normalization block; The typical data of the normalized block calculated by the calculation unit and the typical data of the normalized block that caused the contact stored in the contact data storage unit are compared. A contact block detecting unit that detects the normalized block that has caused the contact, and cutting by cutting the character element block corresponding to the normalized block that has caused the contact detected by the contact block detecting unit. Contact block cutout unit that creates a completed block, a primitive segment creation unit that creates a primitive segment consisting of the character element block extracted by the character element block extraction unit and the cut block created by the contact block extraction unit, A standard data estimating unit for estimating standard data from the primitive segments created by the primitive segment creating unit; and a normalization set from the primitive data created by the primitive segment creating unit and the standard data calculated by the standard data estimating unit. A normalized combination segment creating unit for creating a matching segment, a character-likeness data storage unit storing character-likeness data, a normalized combination segment created by the normalized combination segment creating unit, and the character-likeness data storage An output candidate determining unit for detecting a character-like normalized combination segment by comparing character-likeness data stored in the unit, and a combination corresponding to the normalized combination segment detected as character-like by the output candidate determining unit. It comprises a character candidate output unit for creating a segment from the primitive segment and outputting the segment as a character candidate.

According to a second aspect of the present invention, in the first aspect of the present invention, the contact block extracting section is arranged such that, when the contact block is extracted, the image of the block to be cut is perpendicular to the character string direction. A projection function obtained by counting black pixels in the direction is subjected to smoothing and sharpening at the same time, and a minimum value of the projection function is set as a cutout position.

According to a third aspect of the present invention, in the first aspect of the present invention, the contact block detecting section detects a contact between the start of the character string and the end of the character string when detecting the contact block. It is characterized in that the condition of the raised block is changed.

[0009]

According to the method in which the connected area of black pixels is used as a basic pattern, character elements can be completely separated if there is no contact between characters, but the contacted pattern cannot be separated.

Further, according to the method using the projection of black pixels, a contacted pattern can be separated, but there is a problem that the cutout position becomes very large, and the cutout position is not as accurate as the former method.

Therefore, first, the connected area of the black pixels is used as a basic pattern, and only the contacted pattern is separated using the projection of the black pixel. Therefore, it is possible to solve the problem of extracting a pattern so that the cutout position is accurate and the number of cutouts is as small as possible.

According to this method, the disadvantages of each other can be compensated for. That is, the cut-out position can be obtained very accurately except for the contacted pattern, the contacted pattern can also be cut out, and an increase in the number of cut-outs can be suppressed. However, when this method is adopted, the problem cannot be solved unless a contacting pattern can be detected.

For this purpose, the present invention includes a contact block detecting section for detecting a contacted pattern.

When a connected region of black pixels is extracted as a basic pattern, a contacted pattern and a non-contacted pattern are distributed as shown in FIG. In FIG. 2, the basic pattern is normalized by dividing by the average block width of the entire character string. Even if the basic pattern data extracted from various character string images is plotted in the space of the block width and height, a coherent distribution is obtained because the size of characters written in different images is different. However, by performing the normalization process on the basic pattern, it becomes possible to evaluate different character string images in the same manner, and the distribution becomes unified.

Since the distribution of the block data of the contacted pattern can be known from FIG. 2, if the unknown data in the distribution area is the contacted pattern,
The contact pattern can be detected.

Further, if it is possible to select only possible characters from all possible combinations of the finest patterns, the number of character extraction candidates can be reduced while maintaining the rate at which the output character extraction candidates include correct answers. Can be solved.

For this reason, the present invention includes an output candidate determining unit that selects only characters that are more likely to be characters from all possible combinations of the finest patterns.

When considering the finest pattern combination (segment), the width, height and gap of the segment are considered as evaluation parameters as shown in FIG. This 3
FIG. 4 shows the data distribution of a segment that is a character and a segment that is not a character (redundant) in one parameter space. The data in FIG. 4 is also normalized by the average data of the entire character string as in the case of FIG. Since the data distribution of the character segment can be known from FIG. 4, if the unknown data included in the distribution area is a character pattern, it is possible to detect whether the data is an extra combination.

[0019]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of an embodiment of the first invention.

The configuration of the embodiment shown in FIG. 1 includes an image input unit 101 for inputting a character string image, a character element block extracting unit 102 for extracting a character element from the character string image, and a character element block extracted from the extracted character element block. A standard size estimating unit 103 for calculating a standard block size of one character; a normalizing block calculating unit 104 for normalizing a character element block using an average block size; Contact data storage unit 105 that stores temporary data, and a contact block detection unit 10 that detects a contacted block by comparing the data stored in the contact data storage unit 105 with a normalized block.
6, a contact block cutout unit 107 for cutting a contact block, a primitive segment creation unit 108 for creating a segment cut out in the finest state (hereinafter, referred to as a primitive segment), and a standard of one character from the primitive segment. Data estimating unit 109 for calculating a typical block size and a standard gap between primitive segments, a normalized combination segment creating unit 110 for creating a normalized combined segment from primitive segments and standard data, and storing character-like data Character likeness data storage unit 111
And an output candidate determining unit 112 for detecting a character-like normalized combination segment by comparing the character-likeness data stored in the character-likeness data storage unit with the normalized combination segment, and a character candidate output for outputting a character candidate. And a unit 113.

Next, the operation of this embodiment will be described.

The image input unit 101 is means for inputting a binarized character string image. Character element block extraction unit 10
Reference numeral 2 denotes a unit for extracting a character component from the image of the character string as a block. There are several methods for extracting character components as blocks. For example, a black pixel connection region of a character string image may be extracted as a block, and thereafter, a block obtained by integrating blocks overlapping in a direction perpendicular to the character string direction may be used as a character component block.

Alternatively, a method described in “Ariyoshi et al .:“ Examination of handwritten character segmentation using marginal distribution ”, Shoshin Sogaku Sozenzen, 60” may be used. An example is shown, and when a character string image as shown in Fig. 5A is input, a character element block is extracted as shown in Fig. 5B. This is a means for estimating the standard character size of one character of a character string from a block.There are several methods for extracting the standard character size here. The average width and height of the text components can be taken to a standard size, or, alternatively, once the average and variance of the width and height of all text components of the string can be calculated. The data that deviates from the average by more than a certain variance Removing the data as abnormal data may be a standard size by take heart again average.

The normalized block calculating unit 104 normalizes the blocks of all character components by dividing the block size by a standard size, and calculates a normalized block.

The contact block detecting unit 106 compares the normalized data of all the character constituent blocks with the typical data of the contacted normalized block stored in the parameter storage unit 105 to generate the contacted block. Is detected. There are several methods for detecting the contact block here, but for example, the contact block is detected by comparing each of the width, aspect ratio, and area of the normalized block with a predetermined value. The width of the target normalized block is W and the height is H, and the conditions of Expressions (1), (2), and (3) are calculated, and W> a1 (1) H <b1 × W + c1 (2) H> d1 / W (3) A normalized block that satisfies all the conditions in equations (1), (2), and (3) is a block in which contact occurs. Can be realized. The shaded area in FIG. 6 is an area that satisfies this condition. Here, a1, b1, c1, and d1 are predetermined areas stored in the contact data storage unit.

As another example, the distribution function of the contact block and the non-contact block is estimated in the space spanned by W and H, and the discriminant function that can best separate the two distribution functions is used. Contact block detection can also be performed. Further, the distribution function of the contact block and the non-contact block is estimated, and the probability of the contact block is obtained. If the probability of the contact block is larger than a predetermined value, the contact block can be determined as the contact block.

The contact block extracting section 107 forcibly cuts a character component block corresponding to the normalized block determined to be in contact with the contact block detecting section to create a cut block. It is. There are several methods for forcibly cutting the contacted block. For example, it can be easily realized by a method described in a patent specification (JP-A-4-363780) by the same applicant as the present application.

The primitive segment creation unit 108
By reading the character element block and the cut block that has been subjected to forcible cutting, create the finest block, and remove the very fine blocks from the finest block as noise, and the character components Create a primitive segment that is the smallest unit of.

FIG. 7 shows an example of creating a primitive segment. When a character string image as shown in FIG. 7A is input, a primitive segment is created as shown in FIG. 7B. The standard data estimating unit 109 estimates a standard width and height from all primitive segments in the same manner as the standard size estimating unit, and also estimates a standard gap. There are several methods for extracting the standard gap here. For example, a standard gap may be obtained by removing data in a portion where the value of the gap becomes negative as abnormal data and then calculating an average value.
Another method is to remove the data where the gap value is negative as abnormal data, then calculate the average and variance. And a standard gap may be obtained by calculating the average again.

Normalized combination segment creating section 110
Creates a segment that combines several primitive segments (see FIG. 3), the height of the segment,
A normalized combination segment is created by normalizing the width and gap using the average data. At this time, the gap value of the normalized combination segment created from only one primitive segment is zero,
In the case of a normalized combination segment having a plurality of gap values, the maximum value of the plurality of gap values is used.

The output candidate determining unit 112 detects a character-like normalized combination segment by comparing the normalized combination segment with the character-likeness data stored in the character-likeness data storage unit 111. Here, several methods are conceivable for detecting a character-like normalized combination segment. As an example, by detecting each of the normalized combination segment width, aspect ratio, area, maximum gap length, segment width and gap ratio, with a predetermined value, detection of a character-like normalized combination segment can be performed. Assuming that the width of the target normalized combination segment is W, the height is H, and the gap is G, the equations (4), (5),
The conditions (6), (7) and (8) are compared, and W> a2 (4) H <b2 × W + c2 (5) H> d2 / W (6) G> e2 (7) G> f2 × W + g2 (8) A normalized combination segment that satisfies either equation (4) or (5) and satisfies (6), and (4), (7) ),
This can be realized by deleting the normalized combination segment that satisfies any one of the conditions of (8) as being unlikely to be a character, and setting the remaining normalized combination segment that has not been deleted to be a character. 8A shows a region that satisfies either (4) or (5) and also satisfies (6), and the hatched portion in FIG. 8B shows (4), (7), and (8). ) Indicates an area that satisfies one of the conditions. Where a2, b
2, c2, d2, e2, f2, and g2 are predetermined constants stored in the character-likeness data storage unit.

In another example, the distribution functions of a normalized combination segment that is a character and a normalized combination segment that is not a character in a space spanned by W, H, and G are estimated, and the two distribution functions are most estimated. It is also possible to detect a character-like normalized combination segment using a discriminant function that can be well separated. Furthermore, the distribution function of the normalized combination segment that is a character and the normalized combination segment that is not a character is estimated, the probability that the character is a character is obtained, and if the probability is larger than a predetermined value, the character-based normalized combination segment .

The character candidate output unit 112 is means for outputting, as character candidates, blocks corresponding to the normalized combination segments determined to be character-like by the character candidate determination unit.

In the second invention, the projection function obtained by counting black pixels in the direction perpendicular to the character string direction for the image of the block to be cut in the contact block cutout unit 107 in FIG. The minimum value of the projection function that has been subjected to smoothing and sharpening at the same time is set as a cutout position.

This processing can be realized, for example, as follows. Let f (x) be a projection function obtained by counting black pixels in the direction perpendicular to the character string direction for the image of the block to be cut. The function f (x) is subjected to the calculation of the equation (10) using the Gaussian function g (x) expressed by the equation (9), and the projection function h (x) in which smoothing and sharpening are performed simultaneously Is calculated (see FIG. 9).

G (x) = (1 / σ√2π) exp (−x ² / 2σ ² ) (9) h (x) = {f (xy) (αg (y) + β} ² g (x)) dy (10) where α and β are predetermined constants. Also ² g
(X) is the second derivative of g (x).

In the third aspect of the present invention, the condition of the contacted block is changed between the start of the character string and the end of the character string in the contact block detecting section of FIG.
In other words, a process of gradually expanding the area (hatched area) serving as the condition of the contact block in FIG.

This processing is performed by, for example, equations (1), (2),
With the values of a1, b1, c1, and d1 in equation (3) as initial values, initial values are given at the start of the character string as shown in FIG. 10A, and a1 and d1 are set by the end of the character string. Is gradually changed to a small value, and b1 and c1 are gradually changed to a large value as shown in FIG.

[0039]

As described above, according to the character extracting apparatus of the first invention, a character element block that has made contact is detected from character element blocks that have been cut out in a state that includes contact, and contact has been made. Since the cutout is performed using the projection of black pixels only for the character element block, there is no contact pattern, the cutout position is accurate, the number of cutouts is small and primitive segments can be extracted, and the combination of primitive segments is created. Extra output can be reduced by detecting character-like segments.

According to the character extracting device of the second invention,
When cutting out a contact block, a projection function that performs smoothing and sharpening simultaneously on the projection function obtained by counting black pixels in the direction perpendicular to the character string direction for the image of the block to be cut Is used as the cutout position, the cutout position can be accurately obtained even when no minimum value appears in the projection function.

According to the character extracting device of the third invention,
In the detection of the contact block, by changing the conditions of the block that caused the contact between the start of the character string and the end of the character string, the writing of the character string was written in large characters, and at the end of writing Even with a character string in which characters are written very small, contact can be cut out with high accuracy.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a schematic configuration of an embodiment of the present invention.

FIG. 2 is a diagram for explaining the data distribution of a character element block that has made contact and a character element block that has not made contact.

FIG. 3 is a view for explaining processing of a normalized combination segment creating unit in the block diagram shown in FIG. 1;

FIG. 4 is a diagram for explaining data distribution of a character segment and a non-character segment.

FIG. 5 is a view for explaining processing of a character element block extraction unit in the block diagram shown in FIG. 1;

FIG. 6 is a view for explaining processing of a contact block detection unit in the block diagram shown in FIG. 1;

FIG. 7 is a view for explaining processing of a primitive segment creation unit in the block diagram shown in FIG. 1;

FIG. 8 is a view for explaining processing of an output candidate determination unit in the block diagram shown in FIG. 1;

FIG. 9 is a view for explaining processing of a contact block cutout unit in the block diagram shown in FIG. 1;

FIG. 10 is a view for explaining processing of a contact block detection unit in the block diagram shown in FIG. 1;

[Explanation of symbols]

 Reference Signs List 101 Image input unit 102 Character element block extraction unit 103 Standard size estimation unit 104 Normalized block calculation unit 105 Contact data storage unit 106 Contact block detection unit 107 Contact block extraction unit 108 Primitive segment creation unit 109 Standard data estimation unit 110 Normalization combination Segment creation unit 111 Characteristic data storage unit 112 Output candidate determination unit 113 Character candidate output unit

Claims (3)

(57) [Claims] 1. A character string image to be read is obtained,
An image input unit for inputting a character string image in a character segmentation device that divides an acquired character string image into primitive segments, which are the finest character elements, and extracts characters composed of the primitive segments; and A character element block extraction unit that extracts a character element from the character string image input in, a standard size estimation unit that estimates a standard block size based on the character element block extracted by the character element block extraction unit, A normalized block calculation unit that calculates a normalized block by normalizing the character element block based on the standard block size calculated by the standard size estimation unit;
A contact data storage unit that stores typical data of the contacted normalized block, and a contact block that generates a contact stored in the normalized block calculated by the normalized block calculation unit and the contact data storage unit. A contact block detection unit that detects the normalized block that has caused a contact by comparing typical data of the normalized block, and the normalized block that has caused the contact detected by the contact block detection unit. A contact block cutout unit that creates a cut block by cutting the corresponding character element block, and a cutoff block created by the character element block extracted by the character element block extraction unit and the contact block cutout unit. A primitive segment creation unit for creating a primitive segment,
A standard data estimating unit for estimating standard data from the primitive segments created by the primitive segment creating unit; and a normalized combination from the primitive segments created by the primitive segment creating unit and the standard data calculated by the standard data estimating unit. A normalized combination segment creating unit for creating a segment, a character-likeness data storage unit storing character-likeness data, a normalized combination segment created by the normalized combination segment creating unit, and the character-likeness data storage unit An output candidate determining unit for detecting a character-like normalized combination segment by comparing the character-likeness data stored in the output candidate determining unit; and a combination set corresponding to the normalized combination segment detected as character-like by the output candidate determining unit. Character segmentation apparatus comprising: a character candidate output unit for outputting a character candidate by creating instruments from the primitive segment. 2. The contact block extracting unit according to claim 1, wherein, when cutting out the contact block, a projection function obtained by counting black pixels in a direction perpendicular to a character string direction with respect to an image of a block to be cut is provided. 2. The character extracting apparatus according to claim 1, wherein a minimum value of a projection function subjected to smoothing and sharpening at the same time is used as an extraction position. 3. The contact block detecting section, when detecting a contact block, changes a condition of a block that has caused a contact between a character string start time and a character string end time. The character segmentation device according to claim 1.