JPH11338972A - Device and method for segmenting character string - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make it possible to accurately segment a character string even when an overhang and a separation stroke are present. SOLUTION: A character string extraction means 16 extracts a partial pattern of characters based on the connection information of the character string. A character size calculation means 18 calculates the histogram of the longitudinal or lateral character size of a circumscribed rectangle circumscribing about the extracted partial pattern and calculates an average character size and calculates the distribution value based on the result. A character pitch calculation means 21 calculates the histogram of a pitch between the characters and calculates an average character pitch and calculates the distribution value based on the result. An integration means 22 integrates the characters and also changes the integration condition of the characters corresponding to the distribution value of the average character size and the distribution value of the pitch. A small part integration means 26 integrates the characters by discriminating a small part pattern in the partial pattern in the character string based on the information of the average value and distribution value of the average character size and the pitch.

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 and a character extracting device for extracting a character string.

[0002]

2. Description of the Related Art In recent years, as input peripheral devices for handwriting,
There is an increasing demand for handwritten character recognition devices. In this handwritten character recognition device, each character in a character string is cut out and the character is recognized. In this case, in order to accurately recognize the character string, the character string cutout processing is important.

Conventionally, in a document such as a form in which a character writing position is specified in advance, characters are clearly written in a specified range (character frame) and the character is cut out.
The maximum rectangle in the character string or the simple average of the rectangles in the character string is calculated as the character size for character recognition. In addition, the pitch between characters is calculated to extract characters and integrate characters.

Here, there are the following known techniques for calculating the pitch. For example, JP-A-4-09847
Technique 7 performs one-dimensional projection on character patterns,
The pitch is calculated based on the obtained interval between the white pixel and the black pixel. Japanese Patent Application Laid-Open No. 60-173885 discloses a technique for performing one-dimensional Fourier transform on document image data to estimate a pitch. The technique disclosed in Japanese Patent Application Laid-Open No. Sho 62-195893 is to determine an estimated range of character cutout based on height information of a character string and to search for a cutout position within the estimated range.

Japanese Patent Laid-Open No. 4-17086 discloses a technique for unifying characters by extracting a circumscribed rectangle of connected black pixels and integrating the circumscribed rectangle based on the character size and pitch. As described above, in the conventional character cutout method, each character is written neatly in a character frame at equal intervals and the cutout method is used, so that characters can be recognized with considerable accuracy. By the way, it is troublesome for a writer to write each character neatly in a character frame, so characters are written at a free pitch on paper without a character frame. In the characters written at this free pitch, the size of the characters, the overlap between the characters, the contact, the overhang where the characters approach each other, the separation stroke in which one number is separated into a plurality of patterns, and the like occur.

[0006]

However, when a conventional character segmentation method is used for a character string written at a free pitch, there are the following problems. First, when the largest rectangle in a character string is used as a reference size for a character string having a variable character size, if there is at least one extremely large character, it is not appropriate as the size when the characters are integrated. Did not.

Further, if the average character size is simply calculated for a character string having a large number of small part patterns such as a cloud point, the average character size is calculated to be smaller than the actual size because the average character size is affected by a small size rectangle such as a cloud point. Would.

Further, when the pitch is calculated by the one-dimensional projection for a character string having an overhang, a boundary between a white pixel and a black pixel does not appear. For this reason, there is a problem that the character recognition rate is reduced.

In the method of determining the estimated range of character cutout based on the height information of a character string, there is no problem with a highly square printed character or a handwritten kanji character, but a large variation in character size such as a handwritten numeral. With a character string, there is a problem that the cutout estimation range is not clearly defined.

On the other hand, conventionally, when a character string is extracted, a character on a document image is defined as a black pixel and a background is defined as a white pixel, which is projected in the horizontal (vertical) direction to obtain a histogram. There is a method to extract a character string from Also,
Enlargement and reduction operations are performed on black pixels multiple times, and the resulting group is converted into a single character string (Transactions of the Institute of Electronics, Information and Communication Engineers, '83 / 4 Vol.J66-D No 437-444) and concatenation. A method of obtaining a circumscribed rectangle of a pattern, performing projection with the area of the rectangle weighted to the center coordinates, and obtaining a marginal distribution of the obtained projection values (Transactions of the Institute of Electronics and Communication Engineers '85 / 12 Vol.J6
8-D No 12 2123-2131).

Here, when the vertical (left and right) intervals are small in a handwritten character string, some of the characters of a certain character string may enter between the characters of an adjacent character string as shown in FIG. Likely to happen. In this case, the method of obtaining the histogram makes it difficult to find a valley portion from the projected histogram, and the shape of the histogram depends on the complexity of the character.

Further, in the method of performing the operation of enlarging and reducing the black pixel, the upper and lower character strings are all united. In the method of weighting and projecting the coordinates of the center of the rectangle based on the circumscribed rectangle of the connection pattern, character size information cannot be read from the projection value because only the center of the rectangle is weighted. For this reason, there was a problem that it was difficult to specify the range of the peripheral distribution. In addition, when the size of the height is almost constant but the size of the width fluctuates, such as numbers, the character area is weighted and applied to the center coordinates of the rectangle. , There is a problem that the character is susceptible to horizontally long characters.

A first object of the present invention is to provide a character extracting method and a character extracting apparatus which can accurately extract characters even if there is an overhang or a separation stroke.

A second object of the present invention is to provide a method for extracting a character which can easily extract a character string even when the character string has a small space between the character strings and there is a contact between the character strings. And a device for cutting out characters.

[0015]

Means for Solving the Problems The present invention has the following constitution in order to solve the above-mentioned problems and achieve the object. FIG. 1 is a principle diagram of a character cutting device according to the present invention. The character string cutout device cuts out a character in order to recognize each character of the character string.
6, character size calculating means 18, character pitch calculating means 2
1, an integrating means 22 and a small part integrating means 26 are provided.

The character string extracting means 16 extracts a character partial pattern based on character string connection information. The character size calculation means 18 calculates a vertical or horizontal character size histogram of a circumscribed rectangle circumscribing the extracted partial pattern, and calculates an average character size and its variance based on the result.

The character pitch calculating means 21 calculates a histogram of the pitch between characters, and calculates the average character pitch and the variance thereof based on the result. The integrating means 22 calculates the average character size, average character pitch, size variance value,
Character integration is performed while changing the character integration condition according to the pitch variance value.

The small part integrating means 26 integrates the characters by determining the small part pattern among the partial patterns in the character string based on the average character size. The character size calculating means 18 includes calculating means 32, 34, 38 and a determining means 36. The calculating means 32 calculates a histogram of the length in the vertical or horizontal direction of the rectangle circumscribing the partial pattern of the character in the character string. The calculating means 34 calculates a provisional average character size based on the entire histogram in the character string. The determining means 36 determines a character size calculation area based on the provisional average character size. Calculation means 38 calculates an average character size within the determined character size area.

Further, an area ratio or a height ratio between the average character size and the character size of the circumscribed rectangle is obtained by using the calculated average character size. A small separation stroke extracting unit 20 for extracting a small separation stroke is provided.

When calculating the pitch between characters, the character pitch calculating means 21 uses a distance between circumscribed rectangles as a pitch for a partial pattern other than a small separation stroke, and calculates a histogram of the pitch. Means 21 for calculating provisional average character pitch based on histogram
b, means 21c for determining a character pitch calculation area based on the provisional average character pitch, and means 21d for calculating an average character pitch within the determined character pitch area.

Further, when integrating characters including small separation strokes, the integrating means 22 sets the average character size,
The character integration condition is changed according to the value of the evaluation function including the average character pitch, the size variance, and the pitch variance.

When the pitch width is equal to or less than a certain threshold and the evaluation function value is within a certain range, the integrating means 22
It is preferable to include a certainty factor integration unit 24 that calculates the distance between the extracted small separation stroke and the patterns located on the left and right sides thereof and integrates them based on the distance ratio between them.

The partial integration means 26 includes a linear density calculator 4
2. It includes a tilt calculator 44 and a determiner 46. The linear density calculation unit 42 calculates the linear density for the small separation stroke, the partial patterns located on the left and right sides of the small separation stroke, and the partial pattern obtained by integrating them.

The inclination calculator 44 calculates the inclination of the small separation stroke. The determination unit 46 determines whether the small separation stroke should be integrated with any of the left and right partial patterns based on the line density and the inclination.

When calculating the linear density, the linear density calculator 42 divides the circumscribed rectangle of the partial pattern into n equal parts, and takes the maximum value of the linear density counted from the m-th line to the mn-th line. It is good to

Further, for the straight line segment candidates extracted in the line density calculation, the line density is calculated in one direction with a pattern, and then the line density is calculated in the other direction. It is preferable to include a multi-direction linear density calculator 52 that calculates the sum of the linear densities in a plurality of directions.

The inclination calculating section 44 divides the partial pattern at equal intervals in the direction in which the size of the vertical and horizontal sizes of the circumscribed rectangle is longer, and calculates the inclination based on the intersection between the partial pattern and each of the bisectors. To

A handwritten character recognition device, a printed character recognition device,
The present invention is applicable to a handwritten character cutout device such as a character cutout device in drawing recognition. FIG. 2 is a principle diagram of the second invention. The second invention includes a pattern extracting unit 14, a weighting projecting unit 64, a character string axis determining unit 66, and a character string extracting unit 68.

The pattern extracting means 14 extracts a character partial pattern based on character string connection information. The weighting projection unit 64 obtains a projection histogram by performing weighting projection on a vertical or horizontal line segment of a circumscribed rectangle circumscribing the extracted partial pattern.

The character string axis determining means 66 determines the character string axis based on the projection histogram. Character string extraction means 68
Extracts a character string based on the character string axis. Further, the weighting projection unit 64 performs weighting projection according to the distance from the center of the vertical or horizontal line segment of the circumscribed rectangle of the pattern as a peak. The character string axis determining means 66
Determines the central axis of the character string from the peak value of the projection histogram. The character string extracting means 68 extracts a character string to which a pattern belongs based on the distance between the central axis and the center of each circumscribed rectangle. For example, these means are used when characters do not touch between the upper, lower, left, and right character strings.

Furthermore, a calculating means 71 for calculating the average of the heights of the characters with respect to the circumscribed rectangle and a removing means 72 for removing a circumscribed rectangle having a size equal to or more than a predetermined multiple of the average character height as a contact character block are provided. Good to do. For example, these means are used when a character touches between upper, lower, left, and right character strings and it is known in advance that the character size in the input data is substantially uniform.

There is provided a removing means 73 for removing a circumscribed rectangle crossed by a plurality of character string axes as a contact character string chunk, and the weighting projection means 64 provides peaks at upper and lower ends of vertical or horizontal line segments of the circumscribed rectangle of the pattern. And the distance from the top,
Weighted projection is performed according to the distance from the lower end. The character string axis determining means 66 determines the central axis of the character string from the peak value at the upper end and the peak value at the lower end of the projection histogram.
The character string extracting means 68 extracts a character string to which a pattern belongs from a circumscribed rectangle excluding the contact character string block, based on the distance between the center axis and the center of each circumscribed rectangle.

The weighting projection means 64 performs weighting projection on the upper and lower ends of the circumscribed rectangle of the pattern.
The character string axis determining means 66 determines a candidate position of an upper end and a candidate position of a lower end of the projection histogram, and determines a central axis of the character string from the candidate positions of the upper end and the lower end. The character string extracting means 68 extracts a character string to which a pattern belongs based on the distance between the central axis and the center of each circumscribed rectangle.

Further, a circumscribed rectangle integrating unit 61 for integrating the circumscribed rectangles when the circumscribed rectangles overlap each other is provided. Calculating means 74 for calculating an average character size for the result of the integration; removing means 75 for removing a circumscribed rectangle having a predetermined size or more of the average character size as a contact character block
Be prepared to have. For example, these means are used when a character touches between upper, lower, left, and right character strings and it is known in advance that the character size in the input data is substantially uniform.

The weighted projection means 64 performs weighted projection that attenuates from one end to the other end of the circumscribed rectangle and weighted projection that attenuates from the other end to one end of the circumscribed rectangle. Ask for. The center position of one character and the existing area of one character are estimated from the respective peak values. For example, these means are used for an online handwritten character string.

A character extracting method for extracting a character in order to recognize each character in a character string,
An extraction step of extracting a partial pattern of a character based on the concatenation information of the character string, and calculating a vertical or horizontal character size histogram of a circumscribed rectangle circumscribing the extracted partial pattern and calculating an average character size based on the result And calculating a variance thereof, calculating a histogram of the pitch between characters, calculating an average character pitch based on the result, and calculating the variance thereof, and calculating the average character size and average character pitch. An integration step of integrating characters while changing the integration condition of the characters according to the size variance value and the pitch variance value, and by discriminating a small partial pattern from the partial patterns in the character string based on the average character size. A small part integration step of integrating characters.

In the calculation step, a distance between the circumscribed rectangles is set as a pitch, a histogram of the pitch is calculated, and a provisional average character pitch is calculated based on the histogram. A character pitch calculation area is determined based on the provisional average character pitch, and an average character pitch is calculated within the determined character pitch area.

Further, an extracting step of extracting a partial pattern of a character based on the connection information of a character string, and performing weighted projection on a vertical or horizontal line segment of a circumscribed rectangle circumscribing the extracted partial pattern. And a determination step of determining a character string axis based on the projection histogram, and an extraction step of extracting a character string based on the character string axis.

[0039]

According to the present invention, first, the character string extracting means 16 extracts a partial pattern of a character based on the connection information of the character strings. Then, the character size calculating means 18 calculates a vertical or horizontal character size histogram of a circumscribed rectangle circumscribing the extracted partial pattern, and calculates an average character size and its variance based on the result. Further, the character pitch calculating means 21 calculates a histogram of the pitch between characters, and calculates the average character pitch and the variance thereof based on the result.

Next, the integrating means 22 integrates the characters while changing the character integration conditions according to the average character size, average character pitch, size variance value, and pitch variance value. Then, the small part integrating means 26 integrates the characters by determining the small part pattern among the partial patterns in the character string based on the average character size.

That is, the average size and pitch of characters are strictly calculated for a character string having an indefinite pitch and a variation in character size, and the integration conditions are adaptively adjusted according to their average and variance values during integration. , The character can be cut out with high precision. In addition, since the integration of characters is performed by focusing on the small part pattern, accurate and high-speed processing can be performed.

Further, the pattern extracting means 14 extracts a character partial pattern based on the character string concatenation information, and the weighting projecting means 64 generates a vertical or horizontal line segment of a circumscribed rectangle circumscribing the extracted partial pattern. A projection histogram is obtained by performing weighted projection on the projection histogram.

Then, the character string axis determining means 66 determines the character string axis on the basis of the projection histogram.
Since the character string is extracted based on the character string axis, even if the character strings are close to each other and a part of a certain character belongs to another character string, the character string can be extracted with high accuracy and high speed. .

[0044]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing a character extracting method and apparatus according to the present invention; FIG. 3 is a configuration block diagram of a first embodiment of a character cutout apparatus to which the character cutout method is applied.

<First Embodiment> In the first embodiment, for example, a character is cut out from a handwritten character string having an overhang, and one character is cut out from a handwritten character having an irregular pitch and a change in character size.

In FIG. 3, the input pattern section 12 has a character string pattern including a separation stroke and an overhang. The connection pattern extraction unit 14
Then, only a connected pattern is extracted by inputting a character string input pattern and performing labeling. A character string extraction unit 16 is connected to the connection pattern extraction unit 14.

The character string extracting section 16 extracts a labeled character string. An average character size calculation unit 18 is connected to the character string extraction unit 16. The average character size calculation unit 18 calculates the average character size based on the extracted character string, and includes a histogram calculation unit 32, a provisional average character size calculation unit 34, and a character size calculation area determination unit 3.
6. An average character size calculation unit 38.

The histogram calculator 32 calculates a histogram of the length of each rectangle in the vertical (horizontal) direction. The provisional average character size calculation unit 34 calculates the vertical (horizontal) average character size based on the histogram, and sets this size as the provisional average character size.

The character size calculation area determination unit 36 calculates the character size at which the histogram has the maximum frequency in the area to the right of the tentative average character size, for example. A region for taking the above is determined. The average character size calculation unit 38 calculates the average character size again in the area. The relationship between the maximum frequency value and the histogram is as shown in FIG. 9, and the details will be described in paragraphs 66 and thereafter.

The small character stroke extracting unit 20 is connected to the average character size calculating unit 38. The small separation stroke extraction unit 20 extracts a small separation stroke using the average character size, the area condition, and the height condition. A character pitch calculator 21 is connected to the small separation stroke extractor 20.

The character pitch calculation unit 21 calculates an average character pitch for the strokes that are not determined to be small separation strokes, using the distance between the circumscribed rectangles as the pitch (see P in FIG. 10). Originally, this pitch should be called a gap. However, in the present specification, since the correction of the phrase may cause confusion, the interval shown by P in FIG. 10 is called "pitch".

The character pitch calculator 21 includes a histogram calculator 21a, a provisional average character pitch calculator 21b, a character pitch calculation area determiner 21c, and an average character pitch calculator 21d. The histogram calculation unit 21a
A histogram of the pitch between each rectangle is calculated. The provisional average character pitch calculation unit 21b calculates an average character pitch based on the histogram and sets the pitch as a provisional average character pitch. The character pitch calculation area determination unit 21c
For example, in a region to the right of the provisional average character pitch, a character pitch at which the histogram takes the maximum frequency value (MAX value) is calculated, and an area in which the histogram takes the maximum frequency value (MAX value) / 2 or more in the horizontal direction of the character pitch. To determine. The average character pitch calculator 21d calculates the average character pitch again in the area.

The average character size / pitch integration unit 22 is connected to the small separation stroke extraction unit 20 and the character pitch calculation unit 21, and the extracted small separation stroke and the average character pitch, average character size, size variance value, pitch Characters are integrated based on the information on the variance value. This average font size
The confidence integrating unit 24 is connected to the pitch integrating unit 22.

The certainty degree integration unit 24 calculates the distance between the extracted small separation stroke and the character pattern located on the left and right sides of the stroke and quantifies the distance ratio between them as the certainty degree of integration. Performs integration. A simple recognition processing unit 26 is connected to the certainty degree integration unit 24.

The simple recognition processing section 26 performs processing for handwritten numerals having an overhang. The simple recognition processing section 26 performs linear density processing on small separation strokes, patterns located on the left and right thereof, and patterns obtained by integrating them. , Tilt,
Character size is easily identified and characters are integrated.
The simple recognition processing unit 26 includes a small separation stroke linear density calculation unit 42, an inclination calculation unit 44, character size determination units 46 and 50,
Right adjacent rectangular line density calculator 48, multiple direction line density calculator 5
2

The small separation stroke linear density calculation unit 42 calculates the line density for the small separation stroke in order to determine whether the character is a small written character. An inclination calculator 44 is connected to the small separation stroke linear density calculator 42.

The inclination calculator 44 calculates the X of the small separation stroke.
The inclination of the direction or the inclination in the Y direction is calculated, and it is determined whether the inclination is a separation stroke of 5 or 7 based on the calculated inclination. A character size determination unit 46 and a right adjacent rectangular line density calculation unit 48 are connected to the inclination calculation unit 44.

The character size discriminating unit 46 determines whether the distance to the left rectangle is smaller than a certain threshold value (for example, 1.5 times) the distance to the right rectangle with respect to the angle reliably calculated as 5. Determine whether or not.

The right adjacent rectangular line density calculator 48 calculates the line density of the right stroke of the small separation stroke, and determines whether or not the right stroke is the right part of 7. The right adjacent rectangular line density calculation unit 48 is connected to a character size determination unit 50 and a multi-direction line density calculation unit 52.

The character size discriminating unit 50 determines that the line density of the right stroke from the right adjacent rectangular line density calculating unit 48 is equal to or longer than the threshold value (for example, 1.8 times) of the average horizontal size as the distance from the left stroke in the vertical direction. It is determined whether the separation stroke is small and the inclination of the separation stroke is within a predetermined range.

The multi-direction linear density calculation unit 52 calculates the multi-directional linear density when the linear density of the right stroke is 1 in the vertical direction and 1 in the horizontal direction. <Processing of First Embodiment> Next, a description will be given of a character extracting method realized by the character extracting apparatus of the first embodiment. FIG. 4 is a diagram showing a processing flow of a character cutout method according to the first embodiment, FIG. 5 is a flowchart showing an average character size calculation processing flow, FIG. 30 is a flowchart showing an average character pitch calculation processing flow, and FIG. is there.

First, the pattern output from the input pattern section 12 is a binary image that has been subjected to extreme processing such as inclination correction and rotation correction to remove noise, and has been subjected to pre-processing such as fading. Further, it is assumed that the character string pattern includes overhangs and separation strokes between characters, and does not include overlapping, contact, or continuous characters between characters.

Then, a character string pattern is input from the input pattern section 12 to the connection pattern extraction section 14 (step 101). Next, in order to distinguish individual strokes, for example, as shown in FIG. 7, the connection pattern extraction unit 14 performs labeling by assigning a labeling number to a pattern connected by eight connections (step 102). ).

At this time, since the size of the partial pattern obtained by the labeling becomes a problem later, the connected pattern extracting unit 14 simultaneously performs the labeling and the circumscribed rectangular coordinate values of the partial pattern as shown in FIG. (The upper left and lower right black dot portions) are also calculated.

Then, the character string extracting section 16 extracts the labeled character string (step 103). When there is no character contact or the like, the partial pattern obtained by labeling is either a single character by itself or a partial pattern constituting one character. These are discriminated, and attention is paid to the average size of characters to extract only partial patterns.

Next, the average character size calculating means 18 calculates the average character size based on the circumscribed rectangle of each character of the character string from the character string extraction unit 16 (step 104).
FIG. 5 shows a flow of the process of calculating the average character size. As shown in FIG. 5, first, the histogram calculation unit 32 calculates a histogram of the length of each rectangle in the vertical (horizontal) direction (step 151).

Then, the provisional average character size calculating section 34
Calculates the vertical (horizontal) average character size based on the histogram, and sets this size as the provisional average character size (step 152). At this time, if the character is a kana character, as shown in FIG. 9A, the histogram becomes bimodal due to a small separation stroke generated from a turbid point, ha, ri, ku, and the like. In addition, the histogram becomes bimodal due to small separation strokes generated from 5 and 7 when it is a number, or small separation strokes generated from A and E when it is an alphabet.

Therefore, the calculated provisional average character size is calculated to be smaller than the average character size. Therefore, the character size calculation area determination unit 36 calculates the character size in which the histogram has the maximum frequency value (MAX value) in the area to the right of the provisional average character size, and the histogram has the maximum frequency value (left and right) of the character size. An area having a value of (MAX value) / 2 or more is determined (step 153).

As shown in FIG. 9B, when there is a bias in the peak of the histogram, the histogram at the provisional character size is set to the maximum frequency (MAX value), and the histogram is set to the maximum frequency (MAX value). / 2 or more is determined.

Then, the average character size calculation unit 38 calculates the average character size again in the area (step 1).
54). According to this method, it is possible to calculate the average character size without being affected by the small separation stroke such as the turbid point and without depending on the distribution of the histogram as shown in FIG.

Next, the small separation stroke extracting section 20 uses the circumscribed rectangle already extracted for each stroke to determine whether or not the area of the circumscribed rectangle is equal to or less than 1/2 of the area of the average character size. The height of the rectangle is 4 of the height of the average character size
It is determined whether it is / 5 or less (step 105).

Then, the small separation stroke extracting unit 20
If the conditions of the area ratio and the height ratio are satisfied, the partial pattern of the circumscribed rectangle is extracted as a small separation stroke (step 106).

The reason why the horizontal size (width) is not taken into consideration is that there are some small separation strokes such as the small separation stroke of 5, which are not different in size from the average size.

Next, in the character pitch calculating section 21, for those not determined as the small separation strokes in the processing of step 105 (they are regarded as one character by themselves), as shown in FIG. A pitch histogram is calculated using the distance p as the pitch, and the average character pitch and its variance are calculated based on the histogram (step 107).

The calculation of the average character pitch will be described with reference to FIG. First, the histogram calculator 21a calculates a histogram of the pitch between the respective rectangles (Step 161). The provisional average character pitch calculation unit 21b calculates an average character pitch based on the histogram, and sets the pitch as a provisional average character pitch (step 162). The character pitch calculation area determination unit 21c calculates, for example, a character pitch where the histogram takes a MAX value in an area to the right of the provisional average character pitch, and determines an area where the histogram takes MAX / 2 or more in the horizontal direction of the character pitch. (Step 163). The average character pitch calculation unit 21d calculates the average character pitch again in that area (step 16).
4).

Next, the average character size / pitch integration unit 22
Determines whether or not to perform integration based on the following evaluation function F. F = MP / MW- (α × VP + β) It is determined whether or not F is equal to or greater than zero in the above equation (step 108).
The integration of the partial patterns using the size / pitch dispersion is completed (step 109).

Here, MP is a pitch average, and MW is a size average. VP is pitch variance, α is 1.
6, and β is 0.5. The values of these parameters are examples.

That is, the integration judgment according to the value of the pitch variance with the degree of blank space between characters (pitch average / size average) is completed. Next, in step 108, if F is smaller than zero, the certainty factor integration unit 24 calculates the distance between the extracted small separation stroke and the pattern located on the left and right thereof and integrates the distance ratio between them. Is quantified as a certainty factor, and integration is performed using the quantified value. A specific example of the integration based on the certainty factor will be described.
For example, using distances a, b, c, and d as shown in FIGS. 11B and 11C, b is greater than 2.6 times a and c is 2.6 times d. If it is larger than twice (step 110), the partial patterns are integrated (step 111). In this case, the aforementioned quantified value is 2.6.

Next, in step 110, if b is smaller than 2.6 times a and c is smaller than 2.6 times d, small separation stroke integration processing by simple recognition described below is performed. (Step 112) is executed.

The small separation stroke integration processing based on the simple recognition is performed according to the flow shown in FIG. In this processing, as shown in the figure, first, it is determined whether or not the horizontal / vertical ratio of the small separation stroke is greater than 2.6. Here, a pattern matching method is not used for the small separation stroke. To determine whether it is a single character or a character partial pattern. As described above, since there is no need to perform complicated processing, there is an advantage that processing can be performed at high speed.

The small stroke separation integration process based on simple recognition will be described below in detail with reference to FIG. First, when the horizontal / vertical ratio of the small separation stroke is smaller than 2.6 (step 12),
1), the small separation strokes are integrated as 5 (step 122). The horizontal / vertical ratio of the small separation stroke is 2.
If it is larger than 6, it is determined whether the aspect ratio of the small separation stroke is smaller than 1/3 (step 12).
3). If the horizontal / vertical ratio of the small separation stroke is smaller than 1/3, the routine proceeds to the routine of 7. If the horizontal / vertical ratio of the small separation stroke is greater than 1/3, the linear density is calculated.

In the case of numbers, those extracted as small separation strokes are limited to small written characters or 5 or 7 small separation strokes. For this reason, first, the small separation stroke linear density calculation unit 42 calculates the line density for the small separation stroke in order to discriminate one character written small and the separation stroke of 5 or 7.

As a method of calculating the linearity, as shown in FIG. 12, the small separation stroke linearity calculating unit 42 checks whether the circumscribed rectangle is horizontally long or vertically long. For the time being, linearity is calculated for two lines other than the center (FIG. 12 (a). If the line is horizontally long, the circumscribed rectangle is vertically divided into four equal parts and the same processing is performed (FIG. 12 (b)). In addition,
As another method of calculating the linearity, the circumscribed rectangle is divided into n equal parts.
The maximum value of the linearity counted from the nth line to the nmth line may be taken as the linearity.

Here, when the line density is calculated in the horizontal direction for a horizontally long stroke as shown in FIG. 13, an incorrect line density is calculated. Therefore, the line density calculation method is changed to a circumscribed rectangle. Change accordingly.

According to this, an accurate linearity can be calculated without being affected by the unevenness of the pattern. Then, the small separation stroke linearity calculation unit 42 determines whether the vertical direction of the linearity is 2 or less or the horizontal direction is 1 or less (step 1).
24). If neither of these two determination conditions is satisfied, it is determined that the stroke is not a small separation stroke, and a REJECT is performed (step 125).

On the other hand, if the condition is satisfied, the small separation stroke linear density calculation unit 42 determines whether the aspect ratio of the small separation stroke is 1 or more (step 126). If this condition is satisfied, the inclination calculation unit 44 calculates the inclination of the small separation stroke in the X direction (step 127).
If the above condition is not satisfied, the inclination calculation unit 44 calculates the inclination of the small separation stroke in the Y direction (Step 1).
28).

FIG. 14A shows the method of calculating the inclination.
(B) As shown in (c), the circumscribed rectangle is divided into four equal parts, and
The inclination between two intersections between the first and third lines and the stroke is calculated. In practice, the intersection is not a point but has a certain width, so the midpoint is selected.

Similar to the method of calculating the linear density, the method of calculating the inclination is also distinguished depending on whether the circumscribed rectangle is horizontally long or vertically long. This is because an error may occur when the inclination is calculated in the horizontal direction for a horizontally long rectangle.

As described above, an appropriate inclination can be calculated by changing the calculation direction of the inclination depending on whether the circumscribed rectangle is vertically long or horizontally long. Next, based on the calculated inclination, the inclination calculation unit 44 determines the angle range of the separation stroke having the inclination of 5 (−
40 ° to 28 °) or the angle range of the separation stroke of 7 (step 129). This is because, as shown in FIGS. 15 (a) and 15 (b), the angles of the separation strokes when handwriting 5 and 7 are written by hand are almost in an inversion relationship.

Here, the angular distribution of the small separation stroke of 7 is wider than the angular distribution of the small separation stroke of 5.
Therefore, when the character size discriminating unit 46 discriminates between 5 and 7, the distance to the left rectangle is smaller than 1.5 times the distance to the right rectangle, although the angle is definitely calculated as 5. It is determined whether or not this is the case (step 130), and if this condition is satisfied, it is integrated as 5 (step 13)
1). If this condition is not satisfied, the routine proceeds to the routine of step 132-7.

On the other hand, in step 129, the object rejected in the character size and the object calculated as the angle of 7 are subjected to the following processing. First, the right adjacent rectangular line density calculation unit 48 calculates the line density of the right stroke of the small separation stroke, and determines whether the right stroke is the right part of 7. As shown in FIG. 16, the line density is calculated by examining the line density in both the vertical and horizontal directions in order to distinguish the right part of 7 from 2 and 9.

Then, it is determined whether the vertical direction of the line density of the right stroke is 2 or less and the horizontal direction is 1 or less (step 1).
32). If the condition is not satisfied as a result of the calculation of the line density, the character size determination unit 50 determines that the distance from the left stroke is smaller than 1.8 times the average horizontal size, and the inclination of the separation stroke is −80 ° to 51. It is determined whether it is 6 ° (step 133). If this condition is satisfied, it is integrated as 5 (step 131). If the condition is not satisfied, R
EJECT is performed (step 134).

On the other hand, as a result of the calculation of the linear density, if the vertical direction is 2 and the horizontal direction is 1, (step 132,
135) Assuming that there is a possibility of 7, the character size when integrated with the small separation stroke is examined. If the character size is not more than a certain threshold value of the average character size, it is integrated as 7 (step 136).

On the other hand, when the linear density is calculated by this method,
In step 135, the multi-direction linear density calculation unit 52 checks whether the line density calculated in the vertical direction is 1 and the horizontal direction is 1 is a right pattern of 7,
The linear density is calculated again by the following method.

More specifically, as shown in FIG. 17C, the vertical line density and the horizontal line density 1 shown in FIGS. 17A and 17B are shifted from the center of the width of the circumscribed rectangle in the vertical direction. The density is checked, and when the line density is counted, the line density is checked in the horizontal direction, and it is determined whether or not the right angle linear density is 2 (step 137). Those having a right angle linear density of 2 are integrated as 7 (step 136). Note that the linear density in the multiple directions is 1 for the vertical linear density 1 and the horizontal linear density 1 shown in FIG.

By observing the line density in such a perpendicular direction, it is possible to discriminate a pattern which could not be discriminated by the conventional search in only one direction. Further, for the character “C” shown in FIG. 17F and “L” shown in FIG.
The right angle linear density is 2. In the case of "4" shown in FIG. 17H, the plurality of directions need not be perpendicular directions.

In step 136, if the right-angle linearity is a linearity other than 2, or if the character size is
For the result of T, there is a possibility of a small separation stroke of 5, so the routine returns to the routine of 5, and the character size at the time of integration as 5 is examined. If the condition is satisfied, the characters are integrated, and if the condition is not satisfied, REJE
CT. FIG. 18 shows an example of a character pattern of numbers cut out in this embodiment.

As described above, according to the embodiment, the variable pitch,
Since the average size and pitch of characters are strictly calculated for character strings with variable character sizes, the integration conditions are adaptively changed according to their average and variance values during integration.
Characters can be cut out with high precision. In particular, for the handwritten numeric character strings, accurate and high-speed processing can be performed because the simple recognition processing unit 26 that focuses on small separation strokes is used without using a pattern matching method. That is,
By performing processing focusing on small separation strokes instead of performing uniform processing on all patterns in a character string, it is possible to speed up the entire cutout processing.

Also, the histogram of the widths of all the circumscribed rectangles in the character string is calculated, the average character size is calculated temporarily, and the character size is accurately calculated based on the value. The average character size can be calculated more accurately even when the size fluctuates greatly or when the character string has an overhang. As a result, the characters can be integrated accurately.

Further, the conditions for integrating small separation strokes are adaptively changed in accordance with the character size and the average value and the variance of the pitch in the character string, so that the conditions are independent of the variation in the character size and pitch. More accurate integration can be performed.

When a pitch between characters is calculated including kana dots and separation strokes of numbers in kana characters, a pitch smaller than the actual pitch interval is calculated. By excluding these small separation strokes in advance, it is possible to calculate the pitch more accurately.

Since the threshold for integrating the separated strokes is adaptively changed according to the regularity of the arrangement of the characters in the character string, more accurate character integration can be performed. Furthermore, although the arrangement of the characters in the character string is not regular, the distance ratio between the separation stroke and the pattern located on the left and right of the separation stroke is quantified as a certainty factor, and integration is performed according to the value. I can do it. Embodiment 2 Embodiment 2 will be described below with reference to the drawings. FIG. 19 is a configuration block diagram of the second embodiment. The second embodiment is a character string extracting device. The character string extraction device includes an input pattern unit 12, a connection pattern extraction unit 14, a circumscribed rectangle integration unit 62, a weighting projection unit 64, and a character string axis determination unit 6.
6. A character string extracting unit 68 is provided.

The input pattern section 12 receives a plurality of character string patterns. The connection pattern extraction unit 14 extracts only the connection pattern by performing labeling on the plurality of character string input patterns input by the input pattern unit 12, and obtains each pattern circumscribed rectangle. A character shape determination unit 61 is connected to the connection pattern extraction unit 14.

The character shape determining section 61 determines whether the character pattern has a complicated shape. A circumscribed rectangle integrating unit 62 and a weighting projection unit 64 are connected to the character shape determination unit 61. The circumscribed rectangle integrating unit 62 integrates overlapping circumscribed rectangles. The output of the circumscribed rectangle integrating unit 62 is connected to the weighting projection unit 64.

The weighting projection unit 64 calculates a projection histogram by performing weighting projection on the circumscribed rectangle at the center of the height. This weighted projection unit 64
Is connected to a character string axis determination unit 66.

The character string axis determining section 66 searches for the peak value based on the projection histogram and determines the axis connecting the respective peak values as the character string axis. A character string extracting unit 68 is connected to the character string axis determining unit 66.

The character string extracting unit 68 includes a character string axis determining unit 66
The character string is extracted based on the character string axis determined in. Further, a character size calculation unit 74 for calculating an average character size based on the result of the integration, and a contact character removal unit 7 for removing a circumscribed rectangle having a predetermined size or more of the average character size as a contact character block.
5 is provided. For example, these characters are used when a character touches between upper, lower, left, and right character strings and it is known in advance that the character size in the input data is substantially uniform. <Process of Second Embodiment> The second embodiment will be described below with reference to the drawings. FIG. 20 is a processing flowchart of the second embodiment. First, the connection pattern extracting unit 14 inputs a character string pattern from the input pattern unit 12 (Step 201).

Then, in order to distinguish individual patterns, the connected pattern extracting unit 14 extracts patterns connected by eight connections by labeling, and at the same time, obtains a circumscribed rectangle as shown in FIG. 21 (step 202).

Next, the character shape determining section 61 determines whether or not the character shape is complicated (step 203). The case where the character shape is complicated refers to the case where one character is composed of a plurality of partial patterns like a kanji. At this time,
The complexity is determined based on whether the number of overlaps between the partial patterns is equal to or greater than the threshold value.

On the other hand, if this is not the case, that is, if the character shape is simple, then the weighted projection unit 64
Calculates a projection histogram by performing weighted projection on the center of height as shown in FIG. 22 on the circumscribed rectangle (step 204). The weight value is
It has a peak at the center of the height and decays toward the periphery. At this time, if the entire height line segment has a value, there may be no valley in the projected histogram. For this reason, the value has n% (n;
Real number) only. In FIG. 22A, the shape is triangular in consideration of speeding up the processing.

The projection histogram resulting from the projection will have relatively distinct peak values. The character string axis determination unit 66 searches for the peak value based on the projection histogram, and determines the axis connecting the respective peak values as the character string axis (step 205). This determined character string axis X1,
X2 is shown in FIG.

Then, the character string extracting section 68 calculates the distance between each circumscribed rectangle and the character string axis, determines that the circumscribed rectangle belongs to the character string axis having the shortest distance, and extracts the character string. (Step 206).

On the other hand, in step 203, when the character pattern has a complicated shape, a clear peak may not be generated in the weighted projection. Therefore, as preprocessing of the projection, the circumscribed rectangle integrating unit 62
Merge overlapping circumscribed rectangles (Step 2
07) After performing this integration process, it is determined whether or not there is any contact between characters between the upper and lower (left and right) character strings (step 208). If there is no contact between characters, the process proceeds to step 205.

On the other hand, if there is a contact between characters between the upper and lower (left and right) character strings and it is known in advance that the character size in the input data is substantially uniform, the character size calculation unit 74 Calculates the average character size using the result of the integration (step 209). Then, the contact character removing unit 75 removes the contact character block that is a large circumscribed rectangle based on the size information (step 210), and thereafter performs the weighted projection process. Thereafter, the character string axis determination unit 66 searches for the peak value based on the projection histogram obtained by the projection processing, and determines the axis connecting the respective peak values as the character string axis (step 205).
Then, the character string extracting unit 68 calculates the distance between each circumscribed rectangle and the character string axis, determines that the circumscribed rectangle belongs to the character string axis with the shortest distance, and extracts the character string (step 206). ).

By the integration process, a clearer projection histogram can be obtained. In addition, a character string can be extracted at high speed and with high accuracy by weighted projection. Furthermore, even when the character shape is complicated and there is a character contact, the contact character block is removed, so that the character string can be accurately extracted.

For example, if a character touches between upper, lower, left, and right character strings and it is known in advance that the character size in the input data is substantially uniform, the height of the character relative to the circumscribed rectangle is determined. Is provided with a character height calculation unit 71 (not shown) for calculating the average of. Further, a contact character removing unit 72 (not shown) that removes a circumscribed rectangle having a size equal to or more than a predetermined size of the average character height as a contact character block may be provided. Then, step 2 is applied to the circumscribed rectangle excluding the contact character block.
The processing after 04 may be performed.

As a result, a character string can be accurately extracted even when there is contact between characters at the top and bottom. Third Embodiment Next, a third embodiment will be described. FIG.
FIG. 9 is a configuration block diagram of a third embodiment. Weighting projection unit 6
In 4a, the upper and lower ends of the vertical or horizontal line segment of the circumscribed rectangle of the pattern are peaked, and weighted projection is performed according to the distance from the upper end and the distance from the lower end. Character string axis determination unit 66
“a” determines the center axis of the character string from the peak value at the upper end and the peak value at the lower end of the projection histogram. String extractor 6
8a extracts a character string to which the pattern belongs based on the distance between the center axis and the center of each circumscribed rectangle for the circumscribed rectangle excluding the contact character string block.

Further, there is provided a character contact removal section 73 for removing a circumscribed rectangle crossed by a plurality of character string axes as a contact character string block. <Processing of Third Embodiment> FIG. 25 is a processing flowchart of the third embodiment. Next, processing of the third embodiment will be described. As shown in FIG. 22 (b), the weighting projection unit 64a sets the vertical and horizontal line segments of the circumscribed rectangle of the pattern at the upper and lower ends as peaks, and performs weighted projection according to the distance from the upper end and the distance from the lower end. Perform (Step 301). In this method, weighted projection is performed so that the value decreases from the upper (lower) end toward the center of the height of the circumscribed rectangle. As a result, FIG.
The projection of the upper end and the projection of the lower end shown in FIG. 2B are obtained.

Then, the character string axis determining unit 66a obtains the peak value at the upper end and the peak value at the lower end of the projection histogram, and determines the center of the peak value as the central axis of the character string (step 302).

Next, the contact character removing unit 73 excludes a circumscribed rectangle that is crossed by a plurality of character string axes from among the circumscribed rectangular rectangles as characters that have been vertically touched (step 30).
3). Here, they are the circumscribed rectangles K1, K2 in FIG. Then, the character string extraction unit 68a calculates the distance between each of the circumscribed rectangles and the character string axis with respect to each of the remaining circumscribed rectangles, and assumes that the circumscribed rectangle belongs to the character string axis with the minimum distance. Then, a character string is extracted (step 304).

FIG. 26 shows the method of the second embodiment and the result of projection of black pixels according to the conventional method. FIG. 27 shows the weighted projection results of the center of the circumscribed rectangle in the method of the second embodiment and the conventional method. 26 and 27, the histogram at the right end of the character string is the processing result of the second embodiment, and the histogram at the left end is the processing result of the conventional method.

As shown in FIG. 27, the histogram of the prior art (left end of the character string) is obtained by simply performing weighted projection of the circumscribed rectangle despite the fact that the characters are in contact with each other up and down. It is. On the other hand, in the second embodiment, in contrast to steps 207 to 210 in FIG. 20, in the second embodiment, weighted projection is performed after removing the contact characters as shown in steps 207 to 210 in FIG. Therefore, highly accurate character string extraction is possible.

FIG. 28 shows the result of character string extraction when there is a character contact between the upper and lower character strings. The leftmost histogram shows the result of the top projection, and the rightmost histogram shows the result of the bottom projection. Then, the peak of the histogram of each projection is shown. The string axis is the center of this peak.

As described above, according to the second and third embodiments, even when the character strings are close to each other and a part of a certain character belongs to another character string, the vertical (width) of the circumscribed rectangle of the pattern can be reduced. By performing weighted projection according to the height position on the direction line segment, a character string can be extracted with high accuracy and high speed.

Note that the configuration of the second embodiment and the configuration of the third embodiment may be used in combination. Further, for the online handwritten character string, the weighting projection unit performs a weighted projection that attenuates from the right end to the left end and a weighted projection that attenuates from the left end to the right end of the circumscribed rectangle,
Find each peak of the projection histogram. The center position of one character and the region where one character exists may be estimated from the respective peak values.

For example, when writing "ma" in online handwritten characters, first, the right end projection and the left end projection of the upper part of "ma" are performed at the first stroke shown in FIG. Next, at the second stroke shown in FIG.
Projection of the right end and the left end of the lower part of. Then, as shown in FIG. 31 (c), when the projections are combined for each end, the existence area of one character is obtained.

In the case of a character string having an overhang, as shown in FIG. 32 (b), the right-edge projection and the left-edge projection are performed for each of the 63 characters, thereby obtaining the one-character existence area.

According to this, since a clear valley is generated in the projection histogram, the right and left ends and the center position of each character can be easily and accurately determined.

[0129]

According to the present invention, the average size and pitch of a character are strictly calculated for a character string having an indefinite pitch and a variation in character size. Since the integration conditions are adaptively changed, characters can be cut out with high precision.

Further, even when character strings are close to each other and a part of a certain character belongs to another character string, the vertical position of the circumscribed rectangle of the pattern in the vertical (width) direction is also set at the height position. By performing appropriate weighted projection, a character string can be extracted with high accuracy and high speed.

[Brief description of the drawings]

FIG. 1 is a principle diagram of a character cutout device according to a first invention.

FIG. 2 is a principle diagram of a character string extracting device according to a second invention.

FIG. 3 is a configuration block diagram of a first embodiment.

FIG. 4 is a processing flow of the first embodiment.

FIG. 5 is a flowchart of an average character size calculation process.

FIG. 6 is a flowchart showing small separation stroke integration processing by simple recognition.

FIG. 7 is a diagram showing labeling.

FIG. 8 is a diagram showing calculation of a circumscribed rectangular coordinate value of a partial pattern.

FIG. 9 is a diagram showing an average character size calculation method.

FIG. 10 is a diagram illustrating a method of calculating a pitch.

FIG. 11 is a diagram showing integration.

FIG. 12 is a diagram showing a linear density calculation method.

FIG. 13 is a diagram illustrating an example of failure when calculating a horizontal linear density for a horizontally long stroke.

FIG. 14 is a diagram illustrating a slope calculation method.

FIG. 15 is a diagram showing angles of separation strokes of 5 and 7;

FIG. 16 is a diagram illustrating a linear density calculation method.

FIG. 17 is a diagram illustrating a method of calculating a linear density in a plurality of directions.

FIG. 18 is a view showing an example of a character cutout pattern;

FIG. 19 is a configuration block diagram of a second embodiment.

FIG. 20 is a processing flowchart of the second embodiment.

FIG. 21 is a diagram showing a circumscribed rectangle of each character.

FIG. 22 is a diagram showing weighted projection.

FIG. 23 is a diagram illustrating a character string axis.

FIG. 24 is a configuration block diagram of a third embodiment.

FIG. 25 is a processing flowchart of the third embodiment.

FIG. 26 is a diagram showing a result of projecting black pixels by the method of the second embodiment and the conventional method.

FIG. 27 is a diagram showing weighted projection results of the center of a circumscribed rectangle by the method of the second embodiment and the conventional method.

FIG. 28 is a diagram showing a character string extraction result when there is a character contact between upper and lower character strings.

FIG. 29 is a diagram illustrating an example of overhang between character strings.

FIG. 30 is an average character pitch calculation processing flow.

FIG. 31 is a diagram illustrating an example of a one-character existence area in the case of writing “ma” with online handwritten characters.

FIG. 32 is a diagram illustrating an example of a one-character existence area in the case of a character string having an overhang.

[Explanation of symbols]

 12. Input pattern part 14. Concatenated pattern extraction part 16. Character string extraction part 18. Average character size calculation means 20 Small separation stroke extraction part 21 Character pitch calculation part 22 Average character size Pitch integration unit 24, certainty integration unit 26, simple recognition processing unit 32, histogram calculation unit 34, provisional average character size calculation unit 36, character size calculation region determination unit 38, average character size calculation unit 42・ Small separation stroke linear density calculator 44 ・ ・ Slope calculator 46,50 ・ ・ Character size discriminator 48 ・ ・ Right adjacent rectangular line density calculator 52 ・ ・ Right angle linear density calculator 61 ・ ・ Character shape determiner 62 ・A circumscribed rectangle integrating unit 64 a weighting projection unit 66 a character string axis determining unit 68 a character string extracting unit 71 a character height calculating unit 72, 73, 75 a contact character removing unit 74 ..Character size calculator

Claims (9)

[Claims] 1. A pattern extracting means (14) for extracting a partial pattern of a character based on character string connection information, and weighting a vertical or horizontal line segment of a circumscribed rectangle circumscribing the extracted partial pattern. Weighting projection means (64) for obtaining a projection histogram by performing projection; character string axis determination means (66) for determining a character string axis based on the peak value of the projection histogram; and a center of the character string axis and a circumscribed rectangle. Character string extracting means (68) for extracting a character string to which the partial pattern belongs based on the distance to the character string. 2. The character weighting means according to claim 1, wherein said weighted projection means (64) performs weighted projection according to a distance from the center of a vertical or horizontal line segment of a circumscribed rectangle of the pattern as a peak. The column axis determining means (66) determines the central axis of the character string from the peak value of the projection histogram, and the character string extracting means (68) determines the central axis based on the distance between the central axis and the center of each circumscribed rectangle. A character string cutout device for extracting a character string to which a pattern belongs. 3. A method according to claim 1, further comprising: calculating means for calculating an average of the heights of the characters with respect to the circumscribed rectangle; A character string cutout device comprising: 4. A removing means (7) for removing a circumscribed rectangle crossed by a plurality of character string axes as a contact character string block.
The weighted projection means (64) performs weighted projection in accordance with the distance from the upper end and the lower end, with the upper and lower ends of the vertical or horizontal line segment of the circumscribed rectangle of the pattern as peaks, The character string axis determining means (66) determines the central axis of the character string from the peak value at the upper end and the peak value at the lower end of the projection histogram, and the character string extracting means (68) determines the contact character string chunk. A character string cutout device for extracting a character string to which a pattern belongs based on the distance between the center axis and the center of each of the circumscribed rectangles to be excluded. 5. The method according to claim 1, wherein said weighted projection means (64) performs weighted projection on an upper end and a lower end of a circumscribed rectangle of the pattern, and said character string axis determining means (66) operates on an upper end of the projection histogram. And the center position of the character string is determined from the candidate position of the upper end and the candidate position of the lower end. The character string extracting means (68) determines the central axis and the respective circumscribed rectangles. Character string extracting device for extracting a character string to which a pattern belongs based on a distance from the center of the character string. 6. The method according to claim 2, further comprising:
A character string clipping device including a circumscribed rectangle integrating unit (61) that integrates overlapping circumscribed rectangles when the respective circumscribed rectangles overlap. 7. A method according to claim 6, further comprising: calculating means for calculating an average character size based on the result of the integration; and removing means for removing a circumscribed rectangle having a predetermined size or more of the average character size as a contact character block. 75). 8. The weighted projection means (64) according to claim 1, wherein said weighted projection means (64) attenuates from one end to the other end of the circumscribed rectangle and a weighted projection attenuates from the other end to one end. A character string cutout apparatus that obtains respective peaks of the projection histogram, and estimates a center position of one character and an existing area of one character from the respective peak values. 9. An extraction step of extracting a character partial pattern based on character string connection information, and performing weighted projection on a vertical or horizontal line segment of a circumscribed rectangle circumscribing the extracted partial pattern. A projecting step of obtaining a projection histogram by: a decision step of deciding a character string axis based on a peak value of the projection histogram, and a character string to which the partial pattern belongs based on a distance between the character string axis and the center of a circumscribed rectangle. A character string cutting method including an extracting step of extracting.