JP3103387B2 - Adaptive binarization method for gray image - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention reads multi-valued gray-scale images of characters and the like handwritten on various forms, including entry frames, and performs multi-valued preprocessing for recognizing the hand-written characters and the like. the grayscale image regarding <br/> to how to convert a binary image.

[0002]

2. Description of the Related Art For example, Japanese Patent Laid-Open Publication No. 1-116782 discloses a method for recognizing a money amount handwritten on a check in the United States. Before recognition processing, the check is read by an image sensor or the like. A multi-valued gray image must be converted into binary data as preprocessing. In the binarization, a fixed threshold value is determined from the distribution state of the gradation of the read grayscale data, and the binarization is performed according to the magnitude of the threshold value.

[0003]

However, in some cases, characters are blurred or interrupted due to handwriting, and these portions are erased by binarization.
In the case where a break is likely to occur, it can be remedied by providing a separate pattern in the recognition process. However, blurring of characters is not thorough and incorrect recognition or recognition is impossible.

[0004] Further, there are halftone dots in the character entry frame in addition to the fine lines, and there are many checks in which the background (ground) portion has a halftone dot pattern, and the background pattern is not uniform. Absent. For this reason, these halftone dots are
In many cases, the noise remains without being erased at the time of value conversion, so that a burden is placed on recognition processing in a later stage, and in many cases, erroneous recognition or recognition becomes impossible.

The present invention has been made in view of the above problems, and an object of the present invention is to provide an input frame and a halftone dot (dot) pattern.
Along with the erase only the noise of the like, etc., interruption of character, blur is a adaptive not put a burden on the subsequent stage of the recognition process to repair 2
It is to provide a Nekakata method.

[0006]

SUMMARY OF THE INVENTION The present invention provides thin lines, broken lines,
A plurality of characters freely designated by hand on a ground portion having a halftone dot pattern are read as a multi-valued grayscale image including the entire entry frame and binarized. One
Relates to law, the object of the present invention, for each pixel of the gray image, binarized to determine a threshold value for each said pixel as compared to the concentration in the vicinity of the pixel, the fill frame
And removal of noise such as the dot-like pattern is the incomplete
The noise part but while creating a first feature image repair blurring portion of the character, from the pixels of the grayscale image
Create a binary image by extracting only
To eliminate the noise and blurred parts of the characters
A binary image is created, and each pixel of these two binary images is
Comparing and determining the blurred portion of the character and the noise
A second feature image from which both have been removed is created, and then each pixel of the first feature image and the second feature image is compared.
This is achieved by comparing and judging to obtain an optimal binary image in which only the noise is removed and the blurred portion of the character is restored.

[0007]

According to the present invention, binarization of a grayscale image is not performed by using a fixed or variable threshold value, but by an adaptive method taking into account the situation around pixels. That is, two types of neighborhood density comparison processes having different bias values are performed on the original image (shade image), and the dot extraction process is performed in parallel. Thereafter, the noiseless image creation process and the neighborhood area search process are performed. ing. Furthermore, a hollowing-out prevention image is created, and hollowing-out occurring in a thick portion of the image is removed to complete the image.

[0008]

DETAILED DESCRIPTION FIG. 1 is a flowchart outlining an 2 Nekakata method of the present invention, blurring of characters, interrupted (dregs of the "Character
Re "hereinafter) and the entry frame and dot-like pattern or the like of noise (hereinafter
An original image (grayscale image) 1 having a lower “noise” is, for example, an original grayscale image that is stored for each pixel in 64 levels of shades, with the darkest part being 0 and the brightest part being 63. The original image (shade image) 1 is subjected to a neighborhood density comparison process (step S10), a dot extraction process (step S20), a neighborhood density large bias comparison process (step S30), and a center hole prevention image creation process (step S40). Processed in parallel. When processed in the neighborhood density comparison process (step S10), the binary image 2 is obtained, and the data is sent to the neighborhood area search process (step S60).

Further, a dot extraction process (step S20)
Both the binary image 3 processed in step (1) and the binary image 4 processed in the neighborhood density large bias comparison processing (step S30) are sent to a noiseless image creation processing (step S50).
The binary image 6 created in that process is sent to the nearby area search process (step S60). The binary image 7 processed in the nearby area search processing (step S60) is sent to the center loss prevention processing (step S70). At the same time, the binary image 5 created in the centering prevention image creation process (step S40) is also sent to the centering prevention process (step S70), and the processed image becomes a binary processed image 8.

FIG. 2 shows a neighborhood density comparison process (step S1).
0) is a flowchart showing an operation example of FIG. 3A. First, a pixel number is set to i (step S11), and (D) and (D) of FIG.
(E) and FIG. 4 (C), the pixel _{a i} as shown in (D)
The maximum value of the density of the pixels in the 3 × 3 area (the value of the brightest pixel in the area) is set to max1 (step S12). Note that the hatched portions in (D) and (E) of FIG. 3 and (C) and (D) of FIG. 4 indicate images. Next, let the maximum value of the shading of the pixel in the 9 × 9 area centered on the pixel a _{i be} max4. In the case of a halftone dot pattern as shown in FIG. 3D, the maximum values max1 and max
4 is as shown in FIG. 3A, and the maximum values max1 and max4 in the case of a character portion as shown in FIG.
(B). The dashed line in FIG. 7B shows the case where characters are blurred. In addition, FIG.
Indicates the maximum values max1 and max4 when the background (ground) is read.

As described above, the maximum values max1 and ma
After finding x4,

[0012]

## EQU1 ## It is determined whether or not max1> (max4-bias) (step S14). If so, it is determined that the pixel is not a character, and pixel a _i = 0 (step S15). If not, it is regarded as a character and the pixel a _i = 1 (step S16). Therefore, in the case of (A), (D), and (C) of FIG. 3, a _i = 0, and in the case of (B), (E) of FIG. 3, a _i = 1. That is, the blurred portion of the character is a _i = 1, which means that the character has been restored. By repeating the above operation for all pixels (steps S17 and S18), the entire image can be binarized, and a small halftone dot pattern does not remain as shown in image 2 in FIG.

However, as shown in FIG.
Since the reading density in the case of the character portion is as shown in FIG. 4A, the expression 1 is established and the pixel a _i = 0, and a hollow portion (white portion in the image) as shown in image 2 occurs. . In addition, since the reading density in the case of a large halftone dot pattern as shown in FIG. 4D is as shown in FIG. 4B, the equation 1 does not hold and the pixel a _i = 1. That is, a large halftone dot pattern remains as noise. As described above, when the neighborhood density comparison processing (step S10) is performed on the original image 1, the binary image 2 is obtained. This binary image 2 corresponds to a first feature image. Although noise removal is incomplete in this neighborhood density comparison process,
The faint parts of the letters have been restored.

FIG. 5 shows a dot extraction process (step S20).
6 is a flowchart showing an example of the operation. First, a pixel number is set to i (step S21), and a pixel a as shown in FIG.
_For eight pixels a _ij (j = 1 to 8) around _i

[0015]

(Equation 2) Is calculated, and

[0016]

(Equation 3) Is obtained (step S22).

And

Equation 4] determines whether the delta ≦ -K ₁ (step S23), leaving the pixel as a pixel _a i = 1 if so (step S2
4). In other words, subtracting the gray level of the surrounding pixels from the pixels a _i, that the sum thereof that is a constant value -K ₁ below there are some pixels are several brighter than a _i around a _i In this case, a _i is regarded as a dot and a _i = 1. The same applies to the pixel _ai shown in FIG.
If Equation 4 does not hold in step S23,

[0018]

[Equation 5] N ≧ K₃  Is determined (step S26).
If so, the process proceeds to step S24; otherwise, the pixel a_i
= 0 and the pixel is erased (step S27). One
Mari, the sum is -K₁Does not becomeToutEven pixel a
_iPixels that are significantly brighter than

Equation 3] also satisfy) is around _{a i,} is set to _a i = 1 considers _{a i} a dot if a certain value _{K 3} above. After the pixel a _i = 1 is left in step S24, a _ij (j = 1 to 8) = 1 is set so that surrounding pixels are also left.
(Step S25). This is to more reliably represent the dots. The above operation is repeated until all pixels are completed (steps S28 and S29).

The above-described dot extraction processing (step S
20), a dot as shown by a hatched portion in FIG. 6 and a thin line as shown by a hatched portion in FIG. 7 are extracted.
Therefore, the original image 1 in FIG. 1 becomes like a binary image 3 in the dot extraction process (step S20).

Near density large bias comparison processing (step S
30) is the neighborhood density comparison process described above (step S10).
Is the same as that of b in the expression 1 in step S14.
The value of ias becomes relatively large, and bias2 (> bi
as)). FIG. 8 explains the operation, and even a slightly larger dot as shown in FIG.

[0021]

## EQU6 ## Since max1> (max4-bias2), the pixel is erased as _ai = 0. This is,
In the neighborhood size comparison process (step S10), the size of the bias is

[0022]

## EQU7 ## Since max1 <(max4-bias), the pixel a _i = 1 remains as it is. Also, in the case of faint characters shown in FIG.
Holds, the pixel is erased as _ai = 0. FIG.
(B) shows the same faint character as the broken line in (B) of FIG.

As described above, in the neighborhood density large bias comparison processing (step S30), a slightly large dot or faint character disappears because of the large value bias2. For this reason, the original image 1 in FIG. 1 becomes a binary image 4 by performing the neighborhood density large bias comparison process (step S30).

The binary image 3 created in the dot extraction process (step S20) and the binary image 4 created in the neighborhood density large bias comparison process (step S30) are processed in a noiseless image creation process (step S50). Thus, a binary image 6 without noise is created. This binary image 6 corresponds to a second feature image. Here, in addition to noise, blurred portions of characters are also removed. This will be described with reference to the flowchart of FIG.

First, the "1" pixel of the image 4 is compared with the "1" pixel of the image 3 for each pixel (step S51).
The pixel of image 3 corresponding to the pixel of “1” of image 4 is “1”
It is determined whether or not this is the case (step S52). If the corresponding pixel is “1”, the “1” pixel of the image 4 is erased to “0” (step S53). If the corresponding pixel is not “1”, the “1” pixel of the image 4 is deleted. It is left as it is (step S54). The operations of steps S52 to S54 are repeated until the correspondence of all “1” pixels of the image 4 is completed (step S55).

Noiseless image creation processing (step S5)
The binary image 6 created in step (0) is processed in the neighborhood area search processing (step S60) together with the binary image 2 created in the neighborhood density comparison processing (step S10). That is, as shown in the flowchart of FIG. 10, first, the image 2 is compared with the image 6 (step S61), and "1" of the image 2 is determined.
A 7 × 7 area is searched for the pixel of the image 6 corresponding to the pixel (step S62). FIG. 11 shows this state. Next, it is determined whether or not there is one or more “1” pixels in the area (step S63). If so, the “1” pixel of the image 2 is left as it is (step S6).
4) If not, the pixel of "1" of the image 2 is erased and set to "0" (step S65). For example, the pixel A indicated by the mark X in FIG. 11 has one or more pixels of “1” in the corresponding 7 × 7 area, so that the pixel A is left as “1”, while the pixel B indicated by the mark X is Since there is no “1” pixel in the corresponding 7 × 7 area, the pixel B erases “1” to “0”. The above steps S62 to S65 are repeated until the correspondence of all “1” pixels of the image 2 is completed (step S66). As a result, noise is removed,
Breaks due to blurring or the like are interpolated, and an image as shown in FIG. 1 is obtained.

On the other hand, in FIGS. 4A and 4C, it has been described that a hollow portion occurs in a thick portion in the neighborhood density comparing process (step S10). A dropout prevention process (step S70) is performed.

As a prerequisite for such processing, in the present invention, the centering prevention image creation processing (step S4) is performed on the original image 1.
0). That is, as shown in the flowchart of FIG. 12, first, the pixel number is set to i (step S41).
It calculates an average value a _m of the contrast of pixels in an area of 5 × 5 centered pixels a _i (step S42). FIG. 13 shows this state. Then, the average value _{a m} threshold K
₄ Small whether determined than (step S43), and if so it is left as the pixel _a i = 1 (step S44), the mean value _{a m} is the long threshold _{K 4} or more pixels a _i is erased to "0" (step S4
5). By repeating the above steps S42 to S45 for all the pixels (steps S46 and S47), it is possible to create the center hole prevention image 5 as shown in FIG.

The anti-dropout image 5 created as described above is processed together with the binary image 7 (step S).
70), and the processed image 8 is obtained by adding them as shown in FIG.

[0030]

As described above, in the present invention, near-density comparison processing, dot extraction processing, and near-density large bias comparison processing are performed in parallel on a grayscale image, and further, noiseless image creation processing and neighborhood area search processing. Is performed to obtain a binary processed image, so that only characters can be surely binarized even if noise, blurring, or the like exists. In addition, although a hollow portion occurs in a thick character portion, the hollow portion can be prevented by creating and adding a hollow portion preventing image.

[Brief description of the drawings]

1 is a flow diagram showing schematically the 2 Nekakata method of the present invention.

FIG. 2 is a flowchart illustrating an operation example of a neighborhood density comparison process.

FIG. 3 is a diagram for explaining the operation of a neighborhood density comparison process.

FIG. 4 is a diagram for explaining a problem in the neighborhood density comparison process.

FIG. 5 is a flowchart illustrating an operation example of a dot extraction process.

FIG. 6 is a diagram for explaining a dot extraction process.

FIG. 7 is a diagram illustrating a dot extraction process.

FIG. 8 is a diagram for explaining a neighborhood density large bias comparison process;

FIG. 9 is a flowchart illustrating an operation example of a noiseless image creation process.

FIG. 10 is a flowchart illustrating an operation example of a nearby area search process.

FIG. 11 is a diagram for explaining a nearby area search process.

FIG. 12 is a flowchart illustrating an example of an operation of a center defect prevention image creation process.

FIG. 13 is a diagram for explaining a center hole prevention image creation process.

FIG. 14 is a flowchart showing a hollowing out prevention process.

[Explanation of symbols]

 Reference Signs List 1 Original image (shade image) 2 Binary image 3 Binary image 4 Binary image 5 Dropout prevention image 8 Processed image (binary image)

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-2-245882 (JP, A) JP-A-2-2522084 (JP, A) JP-B-49-32019 (JP, B1) (58) Field (Int.Cl. ⁷ , DB name) G06K 9/38 G06K 9/40

Claims (1)

(57) [Claims] An input frame is designated by a thin line, a dashed line, a halftone dot, or the like, and a plurality of characters handwritten on a ground portion having a halftone dot pattern are multi-valued grayscale images including the entire input frame. in how binarized read as, for each pixel of the gray image, binarizing the to threshold comparison to determine for each said pixel and the concentration in the vicinity of the pixel, the entry frame and removal of noise such as the dot-like pattern is incomplete
There is one for creating a first feature image repair blurring portion of the character, the noise from each pixel of the grayscale image
Create a binary image by extracting only the part and
The noise and the blurred part of the character
Create an erased binary image and make each of these two binary images
Pixels are compared and determined, and the blurred part of the character and the noise
Create a second feature image to remove both's, accordingly
Then, each pixel of the first feature image and the second feature image
Comparing and judging to obtain an optimal binary image in which only the noise is removed and a blurred portion of the character is restored.
Law .