JP2018132953A - Image processing method, and computer program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image processing method and a program that can create a binary image appropriate for extracting meaningful information such as characters or signs from images including various colors, particularly from photographed images.SOLUTION: The method includes: a step S2 for identifying a character area excluding a background area from a color image subject to the processing; a step S3 for classifying each pixel included in the background area into L (L>=2) groups; a step S4 for classifying each pixel included in the character area into N+L groups which are the combination of the L of the background area and the character area specific N (N>=2); and a step S5 for considering the groups of the background area as same groups, dividing the groups of N + 1 into two, and creating a binary image showing the pixels included in one portion in a same one solid color and the pixels included in the other portion in another one solid color.SELECTED DRAWING: Figure 2

Description

The present invention relates to an image processing method capable of creating a binary image suitable for extracting meaningful information such as characters and signs from an image including various colors, particularly a photographed image.

In a moving image such as a video on a television, characters are often overlaid on the image, and a function for extracting only characters may be required. Most of the recent images are color images, but conventional color image character extraction processing uses a monochrome binary image processing technique. That is, after a color image is binarized by some method to form a monochrome binary image, character data is extracted from the binary image. However, since a color image includes various colors, the following problems are associated with character extraction. Even if the color of the character in the color image is different from the background color, as a result of the binarization process, both the character and the background color are converted to black (or white). It is a problem of being lost. Patent Document 1 proposes an “image recognition method” that can accurately recognize an image of each color from a color document including various colors.

Japanese Patent Laid-Open No. 2004-21765

According to the invention described in Patent Document 1, the recognition process is performed for each of a plurality of image data obtained by separating the color image data for each color without making the color image a binary image. Therefore, for example, it is possible to make use of the color of the original by representing different characters for each color in the color original. Also, as long as the color of the character and the background color of the color document are different, they can both be converted to black, preventing the loss of the character, preventing the layout from becoming unrecognizable, and smoothly recognizing the character. You can move on to processing.

However, although the target image here is a color, it is just a document, and does not take into account actual images such as landscapes and people. Since the pixel value changes continuously in a live-action video, it is not meaningful to determine the number of colors before clustering as in paragraphs [0031] to [0032] of Patent Document 1.
An object of the present invention is to create a binary image for reliably extracting characters from a natural image including various colors such as a moving image broadcast on a television.

The present invention is an image processing method for binarizing a color image, the step of specifying a character area excluding the background area from the target color image, and each pixel included in the background area as L (L> = 2) Classifying into groups, and each pixel included in the character area is classified into N + L groups that are a combination of L in the background area and N (N> = 2) unique to the character area. Binary that displays the step and the background area group as the same group, divides the N + 1 group into two, and displays the pixels included in one with the same color and the pixels included in the other with the other color And a step of creating an image.

According to the image processing method of the present invention, a plurality of binary images are created after grouping the background area and the character area. Even if complete character data cannot be extracted with only individual binary images, character data can be extracted with high accuracy by combining information obtained from the plurality of binary images.

In the present invention, the input color image is converted into the coordinates of the color space of the L ^* a ^* b ^* color system in pixel units, and the character area is specified for the converted image. It is good to perform the process.
In this way, since the color representation is converted to the L ^* a ^* b ^* value, which is a color expression that better reflects the characteristics of human vision compared to the RGB values, the similarity of colors can be evaluated without a sense of incongruity for humans.

In the present invention, the background region is grouped by a K-means method for L colors, and the character region is grouped by a K-means method for initially M + L colors. It is preferable to repeat the process of deleting the group having the smallest number of pixels belonging to M (M> N) groups unique to, until M reaches the final group number N. The characters often use colors that do not appear in the background. Therefore, character parts can be appropriately grouped by applying the K-means method separately to the background area and the character area. In addition, when grouping character areas, if a character area has a color close to the pixel value obtained by grouping the background area, the pixel belongs to the background. As a result, even in the area specified as the character area, the pixels originally belonging to the background are appropriately classified as being in the background area, so that the accuracy of the binarization process is increased.

In the present invention, the initial M colors used for grouping specific to the character area are eight colors of R, G, B, C (cyan), M (magenta), Y (yellow), white, and black. Good. Since there are many pure colors such as black and blue, the K-means processing for the character area is preferably started from these colors.

  Since a plurality of binary images are created from the input color image, character data can be extracted with high accuracy by combining them.

1 is a functional block diagram illustrating a configuration of an image processing apparatus according to an embodiment of the present invention. It is a flowchart explaining the outline of the process which concerns on embodiment of this invention. It is explanatory drawing which illustrates the image after specifying the original image and character area which concern on embodiment of this invention. It is a flowchart explaining the background area grouping process which concerns on embodiment of this invention. It is a flowchart explaining the character area grouping process which concerns on embodiment of this invention. It is a figure for demonstrating the number in the case of dividing a group into 2 in order to produce the binary image which concerns on embodiment of this invention. It is a figure which illustrates the binary image which is an output result concerning the embodiment of the present invention.

An image processing apparatus according to an embodiment of the present invention will be described with reference to the drawings.
Hereinafter, the following items will be described.
<< 1. Functional block configuration of image processing apparatus >>
<< 2. Preprocessing by image processing device (character area identification) >>
<< 3. Main Processing 1 by Image Processing Device (Background Area Grouping) >>
<< 4. Main processing 2 by image processing apparatus (grouping of character areas) >>
<< 5. Main processing 3 by image processing apparatus (binary image creation for processing target figure) >>

<< 1. Functional block configuration of image processing apparatus >>
With reference to FIG. 1, a configuration focusing on the function of the image processing apparatus 1 will be described.
The image processing apparatus 1 is realized by a computer such as a personal computer or a smartphone and a computer program (corresponding to a computer program according to claims 5 to 8) installed in the computer.
The image processing apparatus 1 includes a processing unit 2, a storage unit 3, and a communication interface unit 4. In addition to these, an input operation unit such as a mouse and a keyboard used by the operator during operation, an output unit such as a display and a printer, a camera, and the like are appropriately provided, but illustration is omitted.

The storage unit 3 stores an input image (hereinafter, “processing target image”), a learning sample for specifying a character region, various parameters, various intermediate processing results by the processing unit 2, and the like. It is realized by a storage device such as
The parameters include parameters of a convolutional neural network (hereinafter referred to as “CNN”) used to specify a character area, the number of groups when grouping the background area and the character area, and a representative of each group. Contains the initial pixel value.
The intermediate processing result includes the progress of the specified character area, the progress of the application of the K-means method, such as the group to which each pixel belongs.
The storage unit 3 also includes programs for causing a computer to function as the image processing apparatus 1. These programs are read into the memory, and the read program code is executed by a CPU (not shown) so that the processing unit 2 can execute the program. Each part operates.
Next, the processing unit 2 will be described.

  The processing unit 2 includes an image acquisition unit 21, a character region identification processing unit 22, a background region grouping processing unit 23, a character region grouping processing unit 24, and a binary image creation unit 25. In the following, an outline of processing by the image processing apparatus 1 will be described with reference to FIG.

The image acquisition unit 21 acquires a processing target image from an external communication network or information processing apparatus via the communication interface unit 4, and uses color information of each pixel of this image as coordinates in the L ^* a ^* b ^* color space. (Step S1 in FIG. 2). After the character area specifying process (steps S2 to S5 in FIG. 2), the process is performed based on each pixel after conversion. Here, the conversion is performed because the L ^* a ^* b ^* color space can display coordinates closer to human color recognition than the RGB color space, and therefore follows the human color recognition almost accurately. This is because the colors can be separated.

In this embodiment, the character area specifying processing unit 22 specifies a character area excluding the background from the processing target image by the machine learning function implemented in the character area existence determining unit 22b (step S2 in FIG. 2). The present invention is characterized in that it is divided into a background region and a character region, and grouping is performed by applying the K-means method to each. For this purpose, the character area specification processing unit 22 determines whether each pixel included in the processing target image is included in the background area or the character area.
By the way, the entity of the character area existence determination unit 22b is CNN, and the parameters of the CNN are adjusted in advance by the machine learning unit 22a. The machine learning unit 22a and the character area existence determination unit 22b will be described later in << 2. This will be described in “Preprocessing by Image Processing Device (Character Area Identification)”

The background area grouping processing unit 23 classifies each pixel of the area determined as the background of the character area by the character area specifying processing unit 22 into L groups (L> = 2) by applying the K-means method. To do. (Step S3 in FIG. 2).

The character area grouping processing unit 24 groups each pixel included in the character area for each character area specified by the character area specifying processing unit 22 by applying the K-means method. The initial number of groups is M + L, which is a total of M characters unique to the character region and L same as the background region. However, each time the grouping converges, the classification of pixels by the K-means method is stable. Since the group is deleted step by step until the target number of N is reached, the group is finally grouped into a total N + L of N characters unique to the character region and L background regions (step S4 in FIG. 2). Here, N is an integer of N> = 2, and M is an integer of M> N.

The binary image creation unit 25 creates a plurality of binary images based on the result of grouping processing in two stages of the background region and the character region. The purpose of the present invention is to convert the original image into a binary image for character extraction. Therefore, for the pixel in the background area, only the information that the pixel exists in the background area is necessary. Therefore, the L groups classified into the background area are treated as one group without being distinguished, and the image to be processed is selected as one of (N + 1) groups combined with the N groups classified into the character area. Classify each pixel of. The number when (N + 1) groups are color-coded with two colors (usually white and black) is 2 ^{(N + 1)} . However, since the case of all white or black is excluded, a binary image is created for {2 ^{(N + 1)} −2} cases (step S5 in FIG. 2). With the above << 1. The description of the functional block configuration of the image processing apparatus> ends. Subsequently, the operation of the image processing apparatus 1 will be described.

<< 2. Preprocessing by image processing device (character area identification) >>
When receiving the color image to be processed, the image acquisition unit 21 converts the color information of each pixel from the RGB color space or the like to the L ^* a ^* b ^* color space. That is, each pixel is represented by lightness L ^* , hue a ^* , and saturation b ^* in pixel units.
Subsequently, a character area specifying process is performed on the graphic to be processed after being converted into the coordinates of the L ^* a ^* b ^* color space. Hereinafter, this process will be described in detail.

  In the present embodiment, a character area in which character data included in one image exists is specified, and is distinguished from an image area (“background area”) in the background of the character. In order to specify a character area, CNN which is a kind of machine learning is used. Therefore, before describing the processing by the character region identification processing unit 22, the function of machine learning by the image processing apparatus 1 will be described.

The image processing apparatus 1 includes a machine learning unit 22a, collects a large amount of learning images in advance, takes out learning samples, performs machine learning, verifies the results, and adjusts the parameters for machine learning. Specifically, learning images are collected, and positive samples are extracted from the character regions of these images, and negative samples are extracted from other regions. A positive sample that is entirely contained within the character area has a likelihood that its center is contained in the character area, and a negative sample that is not contained within the character area has its center located in the character area. Is assumed to be 0.0. This likelihood is teacher data, and the learning sample is associated with the teacher data.
Each time a new learning sample is input, the likelihood is calculated, and if this likelihood deviates from the teacher data, the parameters are adjusted. For example, the learning sample extracted from within the character region is necessarily within the character region, so the likelihood should be 1.0. However, if the output result from machine learning and the original likelihood of 1.0 are different from each other, parameter adjustment, that is, learning of CNN is performed until a desired accuracy is realized so as to reduce this difference as much as possible. To do.
The parameter adjusted in this way is exported to the character area existence determination unit 22b.
The machine learning has been briefly described above. Returning to the description of the character area specifying process.

The character area specifying processing by the character area specifying processing unit 22 is performed using CNN after machine learning as follows.
The character area identification processing unit 22 scans the processing target image and extracts a small area (hereinafter referred to as “unit area”) having the same size as the learning sample. For example, scanning is performed from the upper left of the image toward the right end with a predetermined movement amount, and when reaching the right end, the image is moved downward by a predetermined movement amount and scanned toward the left end. This is repeated over the entire processing target image.
The extracted unit area is input to the character area existence determination unit 22b each time, and the character area specification processing unit 22 obtains the likelihood that the center of the unit area exists in the character area as an output result.

  When the character region specifying processing unit 22 scans the entire processing target image and acquires the likelihood of the center of the unit region, it determines whether the likelihood is within the character region depending on whether the likelihood is equal to or greater than a preset threshold value. judge. Based on the determination result, it is determined whether each pixel of the processing target image belongs to the character area or the background area. For example, the processing target image shown in FIG. 3A is separated into a character area and a background area like the binary image in FIG. In FIG. 3B, the background area is blacked out and the character area is outlined. In this example, there are three character areas, chA, chB, and chC. In the subsequent character area grouping process, these three character areas are processed separately. This is because the characters that are gathered at a grouped position often have the same color, and the characters that are separated from each other often have a different color.

  The machine learning unit 22a described above operates asynchronously with the color image binarization processing of the present embodiment. Therefore, the machine learning unit 22a may be realized by a computer different from the image processing apparatus 1. The character area presence / absence determining unit 22b, which is the result of the machine learning, may also be realized by another computer so as to transmit / receive data to / from the character area specifying processing unit 22 via a communication line.

As described above, << 2. Although the preprocessing by the image processing apparatus (character area specification) >> has been described, the significance of the character area specification processing in the present invention will be described before the background and character area grouping processing.
That is, it is assumed that character data is extracted without omission in a situation where characters are overlaid on a part of a color live-action video image. Therefore, it is desirable to specify the overlaid area and the background area as a mask in advance. This is because it is possible to first group only the background portion and then extract characters as colors that stand out from the background. If the entire processing target image in which the background and characters are mixed without grouping the character area is simply grouped by the K-means method, the background mixed in the obtained binary image cannot be sufficiently removed. It cannot withstand the use of subsequent processing such as recognition.

<< 3. Main Processing 1 by Image Processing Device (Background Area Grouping) >>
The background area grouping processing unit 23 classifies each pixel included in the background area of the processing target image into a predetermined number L of groups according to the K-means method. Hereinafter, a description will be given with reference to the processing flow of FIG.

First, the number of groups L is set, and the processing counter Np of the K-means method is initialized to 1 (step S31). Initial values bgCLR (l) (l = 1, 2,..., L) of the representative pixel values are set in the L groups (step S32). Note that the representative pixel value will eventually converge when the process by the K-means method is repeated, so the initial value may be arbitrarily set.

The pixel value of the pixel of interest (i, j) is Pij (three-dimensional real vector), and norm S (l) = | Pij−bgCLR (l) | (l = 1, 2,..., L) is calculated. (Step S33). S (l) is the distance between the pixel of interest and the provisional representative pixel value of each group in the coordinates of the L ^* a ^* b ^* color space. A group corresponding to the smallest value among S (l) (l = 1, 2,..., L) is a group to which the pixel of interest (i, j) belongs. This calculation is performed for pixels in all background regions, and all pixels are classified into one of L groups.

Next, it is determined whether the grouping in step S33 is stable. If the process counter Np is greater than 1, that is, if the process of the K-means method is not the first time (“Np> 1” in step S34), the group to which all the pixels belong is the same as the application result of the previous K-means process. If the number of non-identical pixels exceeds a predetermined threshold (including 0) (No in step S35), the process proceeds to step S36. Even when the process of the K-means method is the first time (“Np = 1” in step S34), the process proceeds to step S36. In step S36, the pixel value average of the pixels included in the same group is calculated, the representative pixel value of each group is updated with the obtained average value, and 1 is added to the processing counter Np. When the process of step S36 is completed, the process returns to the process of step S33. That is, the processes in steps S33 to S36 are repeated until the grouping of all pixels is stabilized.

If the process of the K-means method is executed twice or more (“Np> 1” in step S34) and the K-means process has converged (Yes in step S35), the final representative pixel value bgCLR of each group. (L) is confirmed (step S37), and these values are referred to in the subsequent character area grouping process.

<< 4. Main processing 2 by image processing apparatus (grouping of character areas) >>
The character area grouping processing unit 24 classifies each pixel included in the character area of the processing target image into a predetermined number of groups according to the K-means method. Hereinafter, a description will be given with reference to FIG.

The character areas are finally divided into groups unique to N (2 = <N <8) character areas. However, since the background may appear around the character even within the character area, we want to separate the pixels belonging to this background from the pixels of the character itself. Therefore, L groups obtained by grouping the background regions described above are also used in the processing by the K-means method.
The number of groups unique to the character area is finally N, but initially eight is set. That is, the process by the K-means method is started with a total of (8 + L) groups of 8 character specifics and L backgrounds (step S41 in FIG. 5). Here, the processing counter Np of the K-means method is initialized to 1.
By the way, the characters are often drawn in a clear color or bordered. Therefore, the initial value of the character specific group starts from 8 pure colors. The eight colors are R, G, and B of the three primary colors of light, C (cyan), M (magenta), Y (yellow), and white and black.

The representative pixel value of the background group is bgCLR (l) (l = 1, 2,..., L), and the representative pixel value of the character specific group is chCLR (m) (m = 1, 2,..., 8). ). Let Qij be the pixel value of the pixel of interest (i, j). The pixel value is a three-dimensional real vector, and norm bgS (l) = | Qij−bgCLR (l) | (l = 1, 2,..., L) and norm chS (m) = | Qij−chCHR (m ) | (M = 1, 2,..., 8) is calculated (step S42). bgS (l) is the distance between the pixel of interest in the coordinates of the L ^* a ^* b ^* color space and the fixed pixel value of each group in the background region, and chS (m) is a provisional value for each group specific to the character region. This is the distance from the representative pixel value.

The minimum value among {bgS (1) * β,..., BgS (L) * β, chS (1), chS (2),..., ChS (8)} calculated for all groups. Let the corresponding group be the group to which the pixel of interest (i, j) belongs. Here, β (for example, β = 1.5) multiplied by the distance to the background group is a bias. It is preferable to apply the bias because pixels in the character region are determined as characters as much as possible even if they are close to the color of the group in the background region. However, β = 1, that is, no bias may be applied. This calculation is performed for pixels in all character regions, and all the pixels are classified into any of (8 + L) groups (step S42).

Next, it is determined whether or not the grouping in step S42 is stable. If the processing counter Np is greater than 1, that is, if the process of the K-means method is not the first time (“Np> 1” in step S43), the group to which all the pixels belong is the same as the application result of the previous K-means process. If the number of non-identical pixels exceeds a predetermined threshold (including 0) (No in step S44), the process proceeds to step S45. Even when the process of the K-means method is the first time (“Np = 1” in step S43), the process proceeds to step S45. In step S45, the pixel value average of the pixels included in the same group is calculated, and the representative pixel value of each group is updated with the obtained average value. However, the group belonging to the background area is not updated. After updating the representative pixel value, 1 is added to the processing counter Np (step S45), and S42 to S45 are repeated with the same number of groups until the grouping is stabilized (No in step S44). If the grouping is stable (Yes in step S44) and the number of character unique groups exceeds N (No in step S46), the character unique group having the smallest number of pixels in the group is deleted (step S47). The number of groups is a total (7 + L) of 7 character areas and L backgrounds. The number of groups belonging to the background area remains L and is not to be deleted. After deleting one group, the process counter Np is reinitialized to 1, and the process returns to step S42 again. Here, the pixels classified into the group deleted in step S47 are absorbed by the group having the closest representative pixel value among the remaining groups in step S42 to be executed again. The group that has absorbed such pixels recalculates the pixel value average including the pixels that have been absorbed in the subsequent step S45.

If the group classification by the K-means method converges (Yes in step S44) and the number of groups specific to the character area reaches N (Yes in step S46), the pixels in the character area are represented by the character area N. It is grouped into a total of N and L backgrounds (N + L). The process shown in FIG. 5 is executed by the number of character areas specified by the character area specifying process. In the example of FIG. 3B, since there are three character areas chA, chB, and chC, the character area grouping process is executed three times.
In the present embodiment, the processing by the K-means method is started from eight groups unique to characters in addition to the L groups of backgrounds, but these eight are more than the target N. The reason why the initial value of the number of character unique groups is set to 8 is that the K-means method is affected by the initial value, and therefore it is desirable to reduce the value stepwise from a value larger than the target value. Furthermore, if the process by the K-means method is performed with a very small number of groups, there is a possibility that a group of character colors to be left disappears. Considering these, it is appropriate to start processing from eight groups. However, if the number of groups remains eight, the number of binary images that are output results of the present embodiment is 510, which takes too much time for mounting. Therefore, we decided to reduce the number of groups step by step.

<< 5. Main processing 3 by image processing apparatus (binary image creation for processing target figure) >>
The binary image creation unit 25 receives a result of the processing target image being classified into N + L groups, and performs a binarization process for converting pixels belonging to each group into white or black.

Here, once the grouping process for the background and the character area is completed, the specific representative pixel value of each group is no longer necessary as information. For background pixels, only the result of being classified as background is meaningful. Therefore, for each pixel, whether it is classified into one group of backgrounds or one of N groups of character areas remains as meaningful information.
In the case of N = 2, the number of groups including the background is 3, so as shown in FIG. 6, there are eight numbers (1) to (8) when each group is painted with white or black. However, (1) and (8) are meaningless because all groups have the same color, and six binary images (2) to (7) may be created.

FIG. 7 shows two of the binary images obtained from the processing target graphic illustrated in FIG.
In both FIGS. 7A and 7B, character data corresponding to the original image in FIG. 3A cannot be extracted without omission. However, even if a single binary image is not sufficient, omission of character data can be reduced by combining a plurality of binary images. Character data can be extracted with high accuracy even in an image in which gradation is applied to the color of characters or characters are drawn in a striped pattern.
When N = 3, the number of binary images is 14, and when N = 4, the number of binary images is 30, which increases the accuracy of character extraction. An appropriate value of N may be set in consideration of the number of colors included in the processing target image and the required accuracy.

The plurality of output binary image data is sent to, for example, an external device that performs character recognition or displayed on the screen. How to use the obtained binary image is a subject of an invention different from the present invention.

Although one embodiment of the present invention has been described above, the present invention can be variously modified in accordance with the gist disclosed in the claims.

For example, in the above embodiment, the original data is converted into the coordinates of the L ^* a ^* b ^* color space, but the original color information of the original data may be used as it is. L ^* a ^* b ^* is only desirable because it matches the characteristics of human vision. Further, although eight pure colors are used as initial values when grouping character areas, three colors of RGB or four colors of CMYK may be used.

  Further, in the above-described embodiment, after the color expression is converted, it is used as input data for the character area specifying process immediately. However, noise removal by smoothing may be performed prior to the character region specifying process. That is, if a large amount of noise is included in the binary image obtained in the final stage of the present invention, the accuracy of subsequent processing (for example, character recognition processing) based on these binary images is reduced. It is desirable to smooth the image with a filter or the like and output a binary image with less noise.

In the above embodiment, the purpose of binarization is extraction of character data. However, the present invention may be intended to extract not only characters but also pictograms (pictograms) and traffic signs. This is because, like letters, they appeal to the eye and convey information and call attention.
Further, the present invention is suitable for use in extracting character data overlaid on a television image, but can also be used for an image obtained by scanning a color printed matter with a scanner.

  Widespread demand is expected as an image processing technique for color images, particularly as a technique used for character extraction from television telops.

1: Image processing device 2: Processing unit 21: Image acquisition unit 22: Character region specifying processing unit 23: Background region grouping processing unit 24: Character region grouping processing unit 25: Binary image creation unit 3: Storage unit 4: Communication interface part

The present invention is an image processing method for binarizing a color image , the step of specifying an area (hereinafter referred to as a “character area”) in which characters are displayed superimposed on a background from a target color image ; Classifying each pixel included in an area excluding the character area from the color image (hereinafter, “background area”) into L (L> = 2) groups, and each pixel included in the character area Are classified into N + L groups, which are a combination of L in the background area and N (N> = 2) unique to the character area, and the group in the background area is regarded as the same group, and N + 1 groups are represented by 2 And creating a binary image in which pixels included in one are displayed with the same one color and pixels included in the other are displayed with the other one color.

In the present invention, it is desirable to specify the character region by a machine learning function.
Note that, as in the following embodiment, the input color image is converted into coordinates of the color space of the L ^* a ^* b ^* color system in units of pixels, and the character is applied to the converted image. Processing after the step of specifying the region may be performed.
In this way, since the color representation is converted to the L ^* a ^* b ^* value, which is a color expression that better reflects the characteristics of human vision compared to the RGB values, the similarity of colors can be evaluated without a sense of incongruity for humans.

In the present invention, when there are a plurality of specified character areas, it is desirable to group the character areas for each specified character area. This is because the color used for the character is often different for each character area.
As in the following embodiment, the initial M colors used for grouping specific to the character area are R, G, B, C (cyan), M (magenta), Y (yellow), white, It should be 8 black colors. Since there are many pure colors such as black and blue, the K-means processing for the character area is preferably started from these colors.

Claims (8)

An image processing method for binarizing a color image, the step of specifying a character area excluding a background area from a target color image, and L (L> = 2) pixels included in the background area A step of classifying each pixel included in the character region into a group of N + L that is a combination of L in the background region and N (N> = 2) unique to the character region, Consider a group of regions as the same group, divide N + 1 groups into two, and create a binary image that displays the pixels contained in one with the same color and the pixels contained in the other with the other color An image processing method comprising: steps. The image processing method according to claim 1, wherein the input color image is converted into coordinates in a color space of the L ^* a ^* b ^* color system in pixel units. The background region is grouped by the K-means method for the number of colors L, and the character region is grouped by the K-means method for the number of colors initially M + L, and the M region unique to the character region. 3. The process of deleting a group having the smallest number of pixels belonging to (M> N) groups is repeated until M reaches a final number N. 4. Image processing method. The initial M colors used for grouping specific to the character area are eight colors of R, G, B, C (cyan), M (magenta), Y (yellow), white, and black. The image processing method according to claim 3. In order to binarize the color image, a step of identifying a character area excluding the background area from the target color image, and L (L> = 2) pixels included in the background area A step of classifying the pixels into the group, a step of classifying each pixel included in the character region into N + L groups that are a combination of L (N> = 2) unique to the character region, and a background region. A binary image in which N + 1 groups are divided into two, and pixels included in one are displayed with the same color and pixels included in the other are displayed with the other color. And a computer program for executing the above. 6. The computer program according to claim 5, wherein the input color image is converted into coordinates in a color space of the L ^* a ^* b ^* color system in units of pixels. The background region is grouped by the K-means method for the number of colors L, and the character region is grouped by the K-means method for the number of colors initially M + L, and the M region unique to the character region. 7. The process of deleting a group having the smallest number of pixels belonging to (M> N) groups is repeated until M reaches a final number N. 8. Computer program. The initial M colors used for grouping specific to the character area are eight colors of R, G, B, C (cyan), M (magenta), Y (yellow), white, and black. The computer program according to claim 7.

Images