JPH0991378A - Character recognition system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the character recognition system which can stably recognize a correct character by making good use of the effect of the shaping of a character even for a character whose character line is deformed. SOLUTION: This system has a 1st image converting means 2 which scans an image of a read character with a mask of predetermined size and replaces pixels positioned in the center of the mask with a background when the pixels in the mask are all a character line, a 2nd image converting means 3 which measures the length of an array of background pixels from the output image of the 1st image converting means 2 as to two orthogonal directions, accumulates values which are in inverse proportion to the reciprocal of the length of the array at right angles to the measurement directions, and levels the accumulated values respectively, a dictionary part 4 which stores features by characters, and a collation part 5 which extracts features of a character from the output image of the 2nd image converting means 3, compares them with the features of characters stored in the dictionary part 4 to find a similar character category, and outputs the recognition result of the read character or candidates.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition system for optically reading characters, and more particularly to a character recognition system for reading handwritten characters written by various writing instruments.

[0002]

2. Description of the Related Art In a handwritten character recognition method, the effectiveness of a processing method for shaping a character in accordance with the distribution of the line portion or background portion of the character has been confirmed. This process is generally called nonlinear normalization. Such processing has the ability to correct handwritten character distortion, and is said to be particularly effective when reading Chinese characters having a complicated structure. As a representative document of this nonlinear normalization technique,
Handwritten Chinese Character Recognition Based on Hierarchical Misregistration Correction "
(The Institute of Electronics, Information and Communication Engineers Study Group material PRU87-104) is known.

In the above-described non-linear normalization processing, in the input character image, a portion of the character background having a narrow background is enlarged and a portion of the character background having a wide background is reduced. By shaping in this way, the portion that overlaps with the characteristic data for each character category that is held in advance for recognizing the character increases, and the character recognition rate can be improved.

[0004]

However, in the conventional character recognition method as described above, since the processing is performed on the assumption that all the character lines forming the character are obtained, it is possible to use thick magic or a brush. When the character lines of written characters overlap or are in contact with each other, the recognition performance may be deteriorated.

For example, in the case of a character image "bundle" written with a normal pen as shown in FIG. 10A, all the character lines are given as shown in FIG. As shown, the correct character "bundle" is recognized even if the nonlinear normalization is performed. However, when non-linear normalization is applied to the character image "bundle" in which the central portion is crushed as shown in FIG. 10C, the central portion is reduced by shaping the character because there is no background portion in the central portion. The image is converted into an image as shown in 10 (d). For this reason, the correct character "bundle"
The shape difference between and became large, and as a result of collating with the characteristic data for each character category, it was sometimes recognized as an incorrect character.

Further, in Japanese Patent Application No. 60-111444, as a method of recognizing a crushed type character, a portion within a certain range from the edge of the character line portion is set as a collation area.
A method of recognizing the inside as outside the collation area is disclosed. However, since this method relates to a calculation method at the time of matching, character shaping such as non-linear normalization is not performed, and a sufficient effect cannot be obtained when reading a handwritten character. Further, in the publication,
It is described that the width of the character line is extracted, but the specific method is not mentioned.

The present invention has been made in order to solve the problems of the above-mentioned conventional techniques, and it is possible to write characters having various thicknesses and overlapping character lines.
An object of the present invention is to provide a character recognition method that can recognize a correct character stably by making the most of the effect of character shaping.

[0008]

In order to achieve the above object, the character recognition system of the present invention reads characters as an image,
A character recognition method for outputting a recognition result or a candidate of the character, wherein an image of the read character is scanned with a mask having a predetermined size, and when all the pixels in the mask are character lines, From the first image conversion unit that replaces the pixel located in the center of the mask with the background, and from the output image of the first image conversion unit, the length of the array of pixels of the background is measured in each of two perpendicularly intersecting directions, Second image conversion means for accumulating values proportional to the reciprocal of the length in a direction perpendicular to the measured direction so that the accumulated values are uniform, and characteristics for each character category are accumulated. The character portion is extracted from the dictionary unit and the output image of the second image conversion means, the similar character category is searched by comparing with the characteristic accumulated in the dictionary unit, and the read character is read. And having a matching portion for outputting a recognition result or candidate.

At this time, the first image conversion means has a first mask having a predetermined size and a second mask formed around the first mask in a similar shape to the first mask. The image of the character read by the mask is scanned, and when all the pixels in the first mask are character lines and the pixels in the second mask have a background, the image is displayed in the center of the first mask. The located pixel may be replaced with the background.

In addition, a character recognition method for reading a character as an image and outputting the recognition result or candidate of the character, and a distance calculation unit for calculating the distance to the nearest background pixel for each pixel of each character line. , The ratio of the number of pixels with the minimum distance to the number of pixels of the character line is calculated from the calculation result of the distance calculation unit, and the threshold value for replacing the pixel of the character line with the background is calculated from the ratio. A first image conversion unit including a threshold value setting unit and an image conversion unit that replaces pixels having a distance longer than the threshold value with a background;
From the output image of the first image conversion means, the length of the arrangement of the background pixels is measured in each of two directions perpendicular to each other, and a value proportional to the reciprocal of the length is measured in the direction perpendicular to the measured direction. Second image conversion means for accumulating and making the accumulated values uniform, a dictionary unit accumulating characteristics for each character category, and character characteristics from an output image of the second image conversion means Is extracted, a similar character category is searched for by comparing with the characteristics stored in the dictionary section, and a collation section that outputs the recognition result or the candidate of the read character is included.

At this time, the threshold setting unit obtains a ratio of the number of pixels having the smallest distance to the number of pixels of the character line from the calculation result of the distance calculating unit, and the pixel of the character line is set as a background from the ratio. And a lower limit threshold value for replacing the lower limit threshold value and an upper limit threshold value obtained by adding a predetermined value to the lower limit threshold value. Pixels at a distance from the threshold may be replaced with the background.

Further, in any one of the above character recognition methods, a normalizing means for converting the read character into a uniform size may be provided.

According to the character recognition method of the present invention configured as described above, the image of the character read by the first image conversion means is scanned with a mask of a predetermined size, and all the pixels in the mask are character. When the line becomes a line, the pixel located in the center of the mask is replaced with the background, and when two character lines overlap or touch, the character lines are separated and an image close to the correct character is obtained. You can

Further, the image of the character read by the first image converting means is scanned by the first mask having a predetermined size and the second mask formed around the first mask, All pixels in the first mask are character lines,
In addition, when the pixel in the second mask has a background, the pixel located in the center of the first mask is replaced with the background, so that not only when two character lines overlap or contact each other, Even when the character lines of a book overlap or are in contact with each other, the character lines are separated and an image close to a correct character can be obtained.

Furthermore, the distance to the nearest background pixel is calculated for each pixel of each character line by the distance calculation unit of the first image conversion means, and the threshold value setting unit calculates the distance to the number of pixels of the character line. By calculating the ratio of the number of pixels with the minimum distance, calculating the threshold value for replacing the character line pixels with the background, and replacing the pixels with a distance longer than the threshold value with the background in the image conversion unit, Even when reading a character having a thickness, when the two character lines overlap or are in contact with each other, the character lines are separated and an image close to a correct character can be obtained. In particular, the threshold value setting unit adds a predetermined value to this threshold value to set an upper limit threshold value, and the image conversion unit converts pixels between the two threshold values into a background. Thus, it is possible to deal with the case where three character lines overlap or are in contact with each other.

[0016]

Next, the present invention will be described with reference to the drawings.

(First Embodiment) FIG. 1 is a block diagram showing the arrangement of a character recognition system according to the first embodiment of the present invention. In FIG. 1, the character recognition device continuously scans an input unit 1 for reading characters by optical means and an image of the input characters with a mask of a predetermined size, and all the lines in the mask are character lines. In the case of, the first image conversion unit 2 that replaces the pixel at the center of the mask with the background, and the two vertical and horizontal directions that vertically intersect based on the background distribution of the first shaped image output from the first image conversion unit 2. For each of them, the second image conversion unit 3 that shapes the character image so that the background distribution is uniform, the dictionary unit 4 that stores the characteristics of each character category, and the second image conversion unit 3 outputs The second done
An output for outputting the character recognition result of the collation unit 5 and the collation unit 5 for extracting the character features from the shaped image and performing the comparison operation with the character features accumulated in the dictionary unit 4 to recognize the input character. And part 6.

When performing character shaping as described in the conventional example, it is important to obtain a character line which is a precondition as accurately as possible. In the present embodiment, attention is paid to the thickness of the character lines, and for thick lines assumed to overlap character lines, the first image conversion means 2 forcibly provides a background in the central portion. Then, after being processed by the first image conversion unit 2, the second image conversion unit 3 shapes the character according to the non-linear normalization, and the collation unit 5 recognizes the character.

In such a configuration, the character captured by the input unit 1 is output to the first image conversion means 2 as an input character image. Here, the input character image is a binary image for one character in which a value of white (background) or black (character line) is set for each of a plurality of pixels arranged in a grid. The first image conversion means 2 which has received the input character image scans the entire surface of the input character image with a mask having a predetermined size. This mask is a processing unit for determining whether each pixel of the input character image contained in the mask is a character line or a background.

Here, when all the pixels in the mask are character lines, the first image converting means 2 converts the pixel corresponding to the center of the mask into the background pixel, and the other pixels are the input character image. Output without converting. This first
The character image processed by the image conversion means 2 is used as a first shaped image.

The second image conversion means 3 creates a second shaped image based on the distribution of background pixels of the first shaped image. The second image conversion means 3 detects the lengths of the vertical and horizontal arrangements of the background pixels of the first shaped image, and the horizontal background length has a value proportional to its reciprocal. The background length in the vertical direction accumulates a value proportional to the reciprocal thereof in the horizontal direction. Then, the image conversion standard is set so that the cumulative distribution becomes uniform, and based on the image conversion standard, a portion having a high cumulative value is compressed and a portion having a low cumulative value is expanded. This processing is the non-linear normalization processing described in the conventional example, and the character can be accurately recognized by this processing.

FIGS. 2 and 3 show the states of these processes. FIG. 2 shows the conversion process from the first shaped image to the second shaped image by taking the character “T” as an example.
As shown in FIG. 2, the character "Ta" distorted in the first shaped image is accurately shaped in the shape of the character "Ta" with a uniform background in the second shaped image. Also, FIG.
10 shows a case where the character “bundle” whose central portion is crushed as shown in FIG. 10C is taken as an example. Here, FIG. 3A shows the first shaped image, and FIG.
Represents the second shaped image. Here again, it can be seen that the second shaped image is shaped into an image closer to the character "bundle" as compared with the conventional case.

The collation unit 5 extracts the characteristics of the character from the second shaped image and compares the characteristic data of each character category registered in the dictionary unit 4 in advance to find a similar character. Then, the characters determined to be similar are output as a character recognition result, and the recognition process for one character is completed. Here, the signal of the character recognition result output from the collating unit 5 is composed of the comparison calculation result and one or a plurality of character codes derived from the comparison calculation result.

As the method of extracting the feature of the character by the collating unit 5, a method of extracting the feature of the character edge part (focusing only on the character edge part) and a method of extracting the feature of the direction (character line part or edge) are used. Direction of lines), a method of extracting features by thinning, a method of extracting features by image compression processing (compressing an image to absorb fluctuations), a method of extracting features by convolution calculation (mask An appropriate coefficient is given to each position, and the product-sum operation with the input image is locally performed), and a method of extracting features by dimension conversion by a matrix product (of the product-sum operation of the entire image and the corresponding mask) Combination), etc., any of these methods may be adopted based on the processing speed, the reading target, and the recognition accuracy requirement, and the features of the characters may be extracted in any combination.

Also, for the comparison calculation performed by the matching unit 5, any of various distance calculations (city block distance, Euclidean distance, Mahalanobis distance, Batachariya distance, etc.), similarity calculation of inner products, DP matching, etc. is adopted. You may.

Since the above-mentioned feature extraction calculation and comparison calculation are well known techniques, detailed description thereof will be omitted. Further, as for the processing contents of the second image conversion means 3, there is a detailed description in the cited document "Handwritten Chinese character recognition based on hierarchical misregistration correction processing" and its reference document, which are shown in the conventional example, and thus detailed description thereof. Is omitted.

Next, the mask used in the first image conversion means 2 of this embodiment will be described with reference to the drawings.

FIG. 4 shows an example of a mask used in this embodiment. In FIG. 4A, a pixel marked with a circle in the center of the mask is a pixel on which the process of whether or not to replace with the background is performed, and hereinafter, this pixel is referred to as an evaluation pixel. The mask 7 is larger than the thickness of the character line with a margin depending on the writing instrument used to enter the character, the size of the entered character, and the resolution of the input unit 1 (the character font and the resolution in the case of the typeface). Experimentally, the vertical and horizontal widths are set to 9 pixels in this embodiment.
FIG. 4B shows the relationship between the size of the mask 7 and the size of the input character image at this time.

The first image conversion means 2 scans by using the mask 7 so that all the pixels in the input character image become the evaluation pixels, and when the mask 7 is all the character lines, the evaluation pixels are set to the background. Set to a value, otherwise, make the evaluation pixel the same as the input character image. The scanning order at this time is arbitrary.

In the first image conversion means 2, it is assumed that the periphery of the input character image is the background, and the mask fits inside the input character image (the shaded portion in FIG. 4B is the evaluation at that time). (Corresponding to pixels) only,
Pixels that have not become evaluation pixels (portions other than the shaded areas in FIG. 4B) are output with the same values as the pixels of the input character image.

By performing the above-mentioned processing by the first image conversion means 2 before shaping the characters, it is very simple and correct for a character in which two character lines are overlapped or in contact with each other. It is possible to obtain an image that is closer to the characters, and it is possible to reduce recognition of incorrect characters.

As shown in FIG. 4, the mask 7 of the present embodiment has a square shape, but a rectangular mask 7 is used for a pattern such as a type having a clearly different vertical and horizontal widths. May be used. Also, a diamond shape (based on city block distance) so that the distance from the center to the mask periphery is constant.
Alternatively, a pseudo circular (Euclidean distance based) mask shape may be adopted.

(Second Embodiment) Next, a second embodiment of the present invention will be described. In the present embodiment, only the processing of the first image conversion means is different from that of the first embodiment, and other configurations and processing contents are the same as those of the first embodiment, and therefore description thereof will be omitted.

First, the mask used in the first image conversion means of this embodiment will be described.

FIG. 5 is a diagram showing the configuration of a second embodiment of the mask used in the first image conversion means of the character recognition system of the present invention. The mask used in the first image conversion means of the present embodiment is further processed around the mask shown in the first embodiment in order to perform processing assuming that three character lines overlap each other. It has an added configuration. The shaded portion (inner mask) shown in FIG. 5 corresponds to the same mask as in the first embodiment, and the portion without hatching (peripheral mask) is the portion added in this embodiment. The additional portion of this mask is for confirming that all the corresponding portions of the input character image are not character lines, and its size is set in the same way as the mask of the first embodiment. Here, as an example, the first
A mask having a width of 5 pixels each in the vertical and horizontal directions is added around the mask of the embodiment.

The mask shape of this embodiment may be rectangular as described in the first embodiment, and the width of the mask added around the mask does not have to be uniform vertically and horizontally.

Next, the processing procedure of the first image conversion means of this embodiment will be described with reference to FIG. FIG. 6 is a flowchart showing the processing procedure of the second embodiment of the first image conversion means of the character recognition system of the present invention.

In FIG. 6, the first image conversion means first confirms whether or not the current mask position is within the collation range (processing range) of the input character image (step S).
1) If it is within the collation range, the process proceeds to step S2 and the same process as in the first embodiment is performed using the inner mask.
In the process by the inner mask, it is confirmed whether or not all the pixels in the inner mask are the character lines (step S3). If all the pixels are the character lines, the process moves to step S4 and the positions of the peripheral masks are in the collation range. Check if it is inside. If the peripheral mask is within the collation range, the process proceeds to step S5 to start the process using the peripheral mask. In the process by the peripheral mask, it is confirmed whether or not all the pixels in the peripheral mask are character lines (step S6), and when all the pixels in the mask are not character lines, the evaluation pixel is converted into the background (step S).
7).

Here, if the mask position is not within the collation range in step S1, the background is present in the inner mask in step S3, and if all the surrounding masks are character lines in step S6, the evaluation pixel is selected. It is set to the same value as the input character image (step S8). If the peripheral mask position is not within the collation range in step S4, step S4
Moving to 7, the evaluation pixel is converted into the background. When the above processing is completed, it is confirmed whether or not the processing is completed for all the collation ranges (step S9), and if all the processing is not completed, a mask is set at the next collation position (step S10). , And returns to step S1 to repeat the above process.

FIG. 7 shows the appearance of the shaped image at this time. FIG. 7A shows a state in which the center of the character “East” is collapsed, and FIGS. 7B and 7C show the process for the collapsed character based on the first embodiment. The first shaped image and the second shaped image in the case of being shown are shown. Further, with respect to FIG. 7A, a first shaped image and a second shaped image created based on the second embodiment are shown in FIGS. 7D and 7E, respectively.

As shown in FIGS. 7 (b) and 7 (c), in the processing of the first embodiment, there is a possibility that it is recognized as "bundle" instead of "east". By performing the process of this embodiment as shown in (e), a correct character image closer to "East" can be obtained. Therefore, even for a character in which three character lines overlap or are in contact with each other, an image closer to a correct character can be obtained by a simple process, and recognition of an incorrect character is reduced.

In the processing of the first image converting means of this embodiment, the processing is performed assuming that the periphery of the input character image is the background as in the first embodiment, and therefore the mask added to the periphery is When the area is outside the input character image, the latter half of the processing is unnecessary, so that the first shaped image can be created only by the first half of the processing.

(Third Embodiment) Next, a third embodiment of the present invention will be described. Also in this embodiment, only the first image conversion means is different from the first embodiment, and the other configurations and processing contents are the same as those in the first embodiment, and therefore the description thereof is omitted.

The first image conversion means of this embodiment sets the threshold value by determining the thickness of the character line in order to deal with characters written by various writing instruments having different thicknesses. Pixels corresponding to character lines thicker than the threshold value are converted into the background.

FIG. 8 is a block diagram showing the configuration of the third embodiment of the first image conversion means of the character recognition system of the present invention.
In FIG. 8, the first image conversion means 20 of this embodiment is
The distance calculation unit 21 that calculates the distance to the closest background pixel for each character line pixel of the input character image, and the result calculated by the distance calculation unit 21 and the character line pixel are converted to the background from a predetermined value. A threshold setting unit 22 that sets a threshold for the calculation, and an image conversion unit 23 that outputs a first shaped image from the processing results of the distance calculation unit 21 and the threshold setting unit 22.
And is constituted by.

In such a configuration, the distance calculator 21 calculates the distance from the pixel of the character line (inside the hatched line) of the input character image of FIG. 9A to the pixel of the closest background (the portion other than the hatched line).
9 and each value is written on each pixel.
An image called a distance image as shown in (b) can be obtained. Here, it is assumed that the distance of the pixel corresponding to the background is always 0, and the distance value is calculated by using the city block distance calculation.

The distance may be calculated not only by the city block distance but also by the Euclidean distance or the chessboard distance.

In the distance image output from the distance calculation unit 21, the number of pixels having the minimum distance of 1 represents the length of the edge of the character, and the number of pixels other than 0 is the pixel of the character line. Will represent the number of. Therefore, when the number of pixels having a distance other than 0 is divided by the number of pixels having a distance of 1, a value close to half the average character line width of the character can be obtained. The threshold setting unit 22 adds a preset value in consideration of the variation in the width of the character line to the average character line width thus obtained, and sets it as the estimated maximum line width.

With this estimated maximum line width as a threshold value, pixels whose distance is equal to or larger than the estimated maximum line width are converted into a background image by the image conversion unit 23, and the other portions (pixels with a distance of 0 are excluded).
The first shaped image is created by using the character line.

By carrying out such a processing, even when reading a character line of various thicknesses that cannot be specified by the writing instrument, it is possible to perform
It is possible to obtain a character image close to a correct character from a character in which the character lines of a book overlap or are in contact with each other, and it is possible to reduce recognition of a wrong character.

By the way, the above-mentioned estimated maximum line width is set as a lower limit threshold value, and a value obtained in advance by experiments or the like is added to it to set as an upper limit threshold value. The first shaped image may be created by converting the pixels in between to the background and using the other pixels (except for the pixel having a distance of 0) as the character line.

By carrying out such processing, it is possible to deal with not only characters having two character lines overlapping or touching, but also characters having three character lines overlapping or touching. .

Before the processing of the first image converting means of each of the above-described embodiments, the size of the circumscribed character frame of the input character image, or the first and second moments from the center of gravity are made uniform. If a normalizing means to convert the character size is provided in, if the size of the written characters varies, for example, when reading both the sender and the recipient of the mail, or the size of the written characters. Even for a character with a large change in, the variation in the thickness of the character line can be made relatively small, and the character written by various writing instruments can be reliably recognized by a simple process.

Further, the first image conversion means, the second image conversion means, the dictionary section, and the collation section can be realized by the processing according to the program by combining the central processing unit (CPU) and the memory. However, in consideration of speeding up, dedicated hardware (DSP or the like) may be used.

[0055]

Since the present invention is constructed as described above, it has the following effects.

According to the first aspect of the present invention, when the two character lines overlap or are in contact with each other, the first image converting means separates the character lines to obtain an image close to a correct character. It is possible to reduce the recognition of the wrong character. Further, since the character shaping by the non-linear normalization is performed by the second image conversion means, the character can be recognized more accurately. In particular, according to the second aspect, the same effect can be obtained even for a character in which three character lines overlap or are in contact with each other.

According to the third aspect of the present invention, even when the character lines of various thicknesses which cannot be specified by the writing instrument are read, the characters which are overlapped or are in contact with each other by the first image conversion means are used. It is possible to obtain a character image close to a correct character from, and it is possible to reduce recognition of a wrong character. Further, since the character shaping by the non-linear normalization is performed by the second image conversion means, the character can be recognized more accurately. In particular, according to the fourth aspect, the same effect can be obtained even for a character in which three character lines overlap or are in contact with each other.

In the fifth aspect of the present invention, when the size of the written characters varies, for example, when both the sender and the recipient of the mail are read, or the size of the written characters changes. Even for a large character, the variation in the thickness of the character line can be made relatively small, and the character written by various writing instruments can be surely recognized by a simple process.

[Brief description of drawings]

FIG. 1 is a block diagram showing a device configuration of a first embodiment of a character recognition system of the present invention.

FIG. 2 is a diagram showing a conversion process from a first shaped image which is an output of the first image converting means shown in FIG. 1 to a second shaped image which is an output of a second image converting means. .

FIG. 3 is a diagram showing another example of a first shaped image output from the first image conversion unit shown in FIG. 1 and a second shaped image output from the second image conversion unit. is there.

FIG. 4 is a diagram showing a configuration of a first embodiment of a mask used in the first image conversion means shown in FIG.

FIG. 5 is a diagram showing the configuration of a second embodiment of a mask used in the first image conversion means of the character recognition system of the present invention.

FIG. 6 is a flowchart showing a processing procedure of a second embodiment of the first image conversion means of the character recognition system of the present invention.

FIG. 7 is a diagram showing an example of a first shaped image and a second shaped image.

FIG. 8 is a block diagram showing the configuration of a third embodiment of the first image conversion means of the character recognition system of the present invention.

9 is a diagram showing an example of a processing result of the first image converting means shown in FIG.

FIG. 10 is a diagram showing a recognition example when a character is recognized by a conventional image recognition method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 input part 2 and 20 1st image conversion means 3 2nd image conversion means 4 dictionary part 5 collation part 6 output part 7 and 17 mask 21 distance calculation part 22 threshold value setting part 23 image conversion part

Claims (5)

[Claims] 1. A character recognition method for reading a character as an image and outputting a recognition result or a candidate of the character, wherein an image of the read character is scanned with a mask having a predetermined size, First image conversion means for replacing the pixel located at the center of the mask with the background when all the pixels are character lines, and the length of the arrangement of the pixels of the background from the output image of the first image conversion means. In a direction perpendicular to the measured direction, accumulating values proportional to the reciprocal of the length, and making the accumulated values uniform. Means, a dictionary section in which the characteristics of each character category are stored, and character characteristics extracted from the output image of the second image conversion means, and compared with the characteristics stored in the dictionary section to be similar. Out locate the character category, character recognition method, characterized by having a verification unit for outputting a recognition result or candidates of the character read. 2. The character recognition system according to claim 1, wherein the first image conversion means comprises: a first mask having a predetermined size; and the first mask in a similar shape to the first mask. Scan the image of the character read by the second mask formed around the mask,
When all the pixels in the first mask are character lines and the pixels in the second mask have a background, the pixel located in the center of the first mask is replaced with the background. Character recognition method. 3. A character recognition method for reading a character as an image and outputting a recognition result or a candidate of the character, the distance calculating section calculating a distance to a nearest background pixel for each character line pixel. , The ratio of the number of pixels with the minimum distance to the number of pixels of the character line is calculated from the calculation result of the distance calculation unit, and the threshold value for replacing the pixel of the character line with the background is calculated from the ratio. A first image conversion unit including a threshold value setting unit and an image conversion unit that replaces pixels having a distance longer than the threshold value with a background; and a background pixel from an output image of the first image conversion unit. The lengths of the lines are measured in two directions that intersect vertically, and values proportional to the reciprocal of the lengths are accumulated in the direction perpendicular to the measured direction so that the accumulated values are uniform. First Image conversion means, a dictionary unit in which the characteristics of each character category are accumulated, character characteristics are extracted from the output image of the second image conversion means, and compared with the characteristics accumulated in the dictionary section. A character recognition method, which comprises: a character category which is being searched for and which outputs a recognition result or a candidate of the read character. 4. The character recognition method according to claim 3, wherein the threshold value setting unit obtains a ratio of the number of pixels with the minimum distance to the number of pixels of the character line from the calculation result of the distance calculation unit. , A lower limit threshold value for replacing the character line pixels with the background from the ratio, and an upper limit threshold value obtained by adding a predetermined value to the lower limit threshold value, and the image conversion unit A character recognition method, wherein pixels at a distance between the lower limit threshold and the upper limit threshold are replaced with a background. 5. The character recognition system according to claim 1, further comprising a normalizing means for converting a read character into a uniform size.