JPS58123170A - Character reading method - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character reading method that can read with high precision low-quality characters in which part of the character line is blurred or crushed.

One of the major factors that influences the reading accuracy in character reading methods is the print quality of input characters. In printed characters, the causes include smearing of the typeface, incorrect printing pressure, and uneven immersion of the ink ribbon, and in handwritten characters, parts of the character lines may be damaged due to problems with the writing utensil, fluctuations in writing speed, and the habits of the person writing. The image becomes blurred or crushed. In particular, in handwritten letters written with a single ballpoint pen, the ink does not flow smoothly due to the uneven rotation of the ball, resulting in frequent blurring of the letter lines. of numbers
There is no need to give examples of CI with ON and English letters.

When reading characters, the most important thing is the separation and combination of the character lines, so as mentioned above, the character lines may become blurred or
It is normal for the reading accuracy of characters that are crushed or crushed to be greatly reduced.

Conventionally, for the purpose of reading this type of low-quality characters, the average density level, average line width, and character scanning direction K of the character pattern are
A method has been proposed in which characters are read by converting input character patterns into binary values of black pixels and white pixels using a threshold value calculated based on one-dimensional local contrast. This is a basic and excellent method, and is effective for printed characters that contain noise caused by ink droplets, and for lowercase printed characters that are prone to smearing. However, because this method was devised primarily to improve the reading accuracy of printed characters, it is not very effective for handwritten characters that have large variations in density level and line width within the same character pattern. It has its drawbacks.

Another method that has been proposed to overcome the above drawbacks is to utilize two-dimensional local contrast. In other words,
If the difference between the density level of one pixel in a two-dimensional geographic plane and the density level of its neighboring pixels is greater than or equal to a predetermined threshold,
The center pixel of interest is defined as a black pixel, and the other pixels are defined as white pixels. 9, which uses contrast, is a natural idea, and is also effective for text line areas with low density levels and crushed areas with relatively high density levels. but,
This method has the drawback that if there is unevenness in the character line part KII, a part of the inside of the character line may be determined to be a white pixel, or the pentagram stain may be emphasized and determined to be a clearly black pixel.

In addition, in other conventionally proposed methods, two types of thresholds are provided, a high threshold and a low threshold at the 111 degree level, and the high threshold is used to binarize and connect to pixels determined to be black pixels, and the low threshold is used to connect to pixels determined to be black pixels. A pixel determined to be a black pixel is defined as a black pixel. This method has the advantage that it is easy to remove isolated noise that is far from characters, and is also effective against blurred character lines, but it is difficult to remove noise that is combined with characters, and it is vulnerable to characters being cluttered together. There are drawbacks. 4IIC, it is difficult to set a low threshold value; if it is set high, it becomes weak against blurring, and when it is set low, it becomes weak against blurring. Particularly in the case of low-quality characters, the density level of the blurred portions of the character lines is often higher than that of the faded portions, so it is impossible to correctly binarize both into black and white using this method.

As mentioned above, the various character reading methods that have been proposed so far have one or more drawbacks. As mentioned above, the essence of character information lies in the connection and separation of character lines, and the width of character lines is a secondary matter, so blurred parts are [The character reading method must be able to guarantee the correct connection of character lines. To achieve this, character patterns, unlike ordinary figures, must be recognized as black pixels and as white pixels. The method should utilize the information that it is a special pattern composed of character lines.

This invention was made in view of these points, and its purpose is to thin a binarized pattern obtained using a low threshold under certain conditions, and then perform local processing to make black pixels line up with character lines. It is an object of the present invention to provide a character reading method that reliably detects only character line portions and reads low-quality characters with high accuracy by checking whether or not they are part of the text.

An embodiment of the present invention will be described in detail below with reference to the drawings.

Figure 1 is an example of the character pattern (2) of the number 8 written on a recording medium (1) such as paper, and (2m) and (2b) are dirt and stains on the recording medium (1). This is noise generated by dust, etc. Noise (2a) is located on the character line, and noise (2b) is located away from the character line.

In addition, this character pattern/(2) has extremely poor character shape and printing quality, and the upper right corner is faded and the lower loop part is crushed. It is very difficult to prevent the faded parts from being cut and the loop parts from being crushed together with the strips for removing (2b), but this is an example of the optimal character pattern to demonstrate the effects of this invention. be.

FIG. 2 shows a multivalued character pattern (3) in which the density levels of the character pattern (2) in FIG. 1 are displayed in 16 levels.

Concentration levels range from 0 to 99. It is expressed in hexadecimal numbers from A to F, where F is the highest density level and 0 is the lowest density level. However, to make the figure easier to read, kO is replaced with a blank space.

FIG. 3 shows the multivalued character pattern (3) with a low darkness value TH.
This is a 200 million character pattern (4) binarized with t. In this example, THL=2, pixels having a density level equal to or higher than THL are defined as black pixels, and other pixels are defined as white pixels. As shown in FIG. 3, unless otherwise specified, black pixels are represented by X marks, and white pixels are represented by blank spaces. 200 million character patterns (4
), because the threshold was set to a low value, the blurred part at the top right is connected, but the loop at the bottom is collapsed and noise (
4a)j (4b) also appears clearly.

FIG. 4 shows the multivalued character pattern (3) at a high threshold TH.
This is a 200 million character pattern (5) binarized with H. In this example, 'rHH=5, pixels having a density level equal to or higher than THH are defined as black pixels, and the others are defined as white pixels. 2
Billion character pattern (5) has a high threshold value, so
Although the noise has been removed and the loop at the bottom is clearly visible, the faint part at the top has been cut off.

FIG. 5 shows a thin line character pattern (6) obtained by performing ll1A line processing on the binary character pattern (4) according to the method proposed by the present invention. The thinning process is performed using this kind of E, S
, Deutsch's method (Thinning algo
rithms on rec-1 angular, he
xagonal, and triangular
arrays', Com-munica
tions of the ACM, V(11,
15, fJn 9. pp. 827-837 (1972
), ) can be used. However, the purpose here is not to obtain a embroidery pattern with a line width of lii elements, but to remove noise while maintaining the connected state of character lines, so the above line thinning method is as follows. changes are being made.

(1) The black pixels to be removed are only pixels whose density level is not THH grooved.

(2) Remove the endpoint preservation condition.

(3) Also remove the isolated point preservation condition.

(4) Lines are sequentially thinned starting from the pixels with the lowest degree level.

Note that the above item 4 is to leave the ridge part of the density level of the character line as a black pixel, so that the 0 of the thin line character pattern (6), which is a change point, is □'' to prevent deformation of the character shape.
The mark is a black pixel with a density level below THH, the pixel in the upper right faded part (6a) is not scraped off to preserve the connectivity of the character line, and the pixel in the lower loop (6b)
Since it is a pixel within the character line, it is saved as a black pixel. Regarding the two types of noise, both of them were deleted because the preservation conditions for end points and isolated points were removed. follow [,
If the black pixels (6b) in the lower loop of this thin line character pattern (6) can be converted to white pixels, the object of the invention will be achieved. Black pixel (6a) e
Since the pixels (6b) are located at the ridges and valleys of the density gradient, the best way to determine whether these pixels are white or black pixels is to examine the contrast between the pixels and neighboring pixels. As mentioned in the explanation of the prior art, if you try to judge the entire character pattern as black and white simply by using contrast, various defects will occur due to density unevenness. If it is used only for a limited time, the above-mentioned drawbacks will not occur and great effects will be brought about.

FIG. 6 is an example of a local processing mask for examining contrast. Let the center pixel be XO.

16 pixels 2 chessboard distance apart from x1% x
It is expressed as ts. contrast)! Many methods have been proposed for the determination, but here, in order to avoid the influence of density unevenness, the following method is used as an example. Of course, the present invention is not limited to the following method.

Contrast) C is algebraically ** using the following formula.

C=*! s((!Z(1+β5r)−Σf(Zat4)
C1 However, χ0 is the density level of pixel Xo, χ4 (A=l~1
6) is the density level of pixels X1 to x16, and α, β, and r are non-negative constants. mis (αχ.+βsr) is the seventh
It is the function value shown in the figure, and Σf(χ0,4) is the pixel X1~
White pixel X of X1.

Since the number of pixels is higher than the density level -j, the contrast C has a small value at the valley of the density gradient and a large value at the ridge. Therefore, if the contrast C is a predetermined value of 1 or more, the pixel Xo remains as a black pixel; otherwise, the pixel
, is converted into a white pixel. The constants α and βer*a have the property of selecting the optimum values empirically by applying them to the actual character pattern.Also, the size of the local processing mask for determining the contrast shown in Figure 6 is , and is determined by referring to the average character line width. For example, the size of one side of the mask is determined to be the average character line width + 2.

FIG. 8 shows the constant α=1. β-Tomi6. γ=9°δ=4
and black pixel (6a) of thin line character pattern (6)
*This is the final character pattern obtained as a result of re-judging (6b). Since the black pixel (6a) is located in the ridge, the contrast C becomes 6 or more and remains a black pixel, but the black pixel (6b) is located in the valley, so the contrast C becomes 2 or less and it remains a white pixel. is converted to

As a result of the above, the final character pattern obtained is improved to be of high quality, with no noise, faded portions not being cut, and crushed portions clearly separated.

Figure 9 makes it easier to understand the above processing steps from a programming perspective! , is expressed as a flowchart.

In other words, the multilevel character pattern obtained by scanning the character with a stepper and photoelectrically converting it is
Value it, step C! In υ, a high threshold value THH is determined so that the average line width of the binarized pattern becomes the set value, in step (2), black pixels below the high threshold value 781 are thinned, and in step (c), the density level It is determined whether or not there is a black pixel that satisfies THH, and if no, the process ends; if yes, the validity of the black pixel that satisfies the above condition is re-determined by local processing, and the process ends.

High-quality character patterns make up the majority of the characters read.
There are few black pixels whose density level is less than THH after line thinning (in that case, local processing using contrast or the like becomes unnecessary and the processing time is improved.

Next, we will discuss how to select the two threshold values THL and THH. Due to the characteristics of this invention, the degree of freedom in determining the threshold value is large, and the threshold value can be determined relatively freely regardless of character quality. However, the low threshold value THL is lower than the density level of the faded part of the character line and is free from dirt and dust. A value higher than the density level of the recording medium itself.

This value can be easily determined by examining an appropriate number of character patterns.

The high threshold value THH must be set higher than the density level of a collapsed portion of a character line or a contact portion of two parallel lines (and a value that does not cause the character line to disappear).

If the value is set too high, the character pattern may degenerate and the character line may disappear because the end point preservation condition is deleted in the line thinning process. An excellent method for determining THH is to measure the average line width Z of the 200 million character pattern (5) shown in Figure 4, and then determine whether this value is the closest to the predetermined average line width W, i.e. the set value W. This is the method of determining 'rHH. In other words, by changing THH, the difference between W and d =
Find 'rHii that minimizes lW - wl. 10th
The figure represents this as a flowchart from a programming standpoint. That is, in step (to), d is set to a sufficiently large value, 'rHt is set to THH in step 6υ, and this THH is used to perform binarization in step (2).

At step (to), measure the average line width W, step 104)
Then, determine IW-wlkud.

If l W −w l < d is yes, step (
In step (to), Iw−vl is set to d, and in step (to), the value obtained by adding 1 to THH is set as THH, and the process returns to step (to).

Hereafter, repeat the procedure of proceeding to Step 9 and Step (B). If IW-ml-d is NO in step (B), then in step (B) the result obtained by subtracting 1 from THii is set as THH and the process ends. The method for determining the average line width is a well-known technique for engineers in this type of technical field.
Artz's method ("The I BM 1975?T
ptical Page Reader, Part
II :Video thre@holding S
system', IBM Res, Develop
.

This can be achieved by using the following: pp. 354-363 < 1968).

Furthermore, another example of how to determine the high threshold value THH may be as follows. That is, it is determined by K as follows from the average density ■ of the multivalued character pattern shown in FIG.

THH=-1-V-b Here, a and b are constants, and ■ is given by the following formula.

-Σ■L V=- (2) is the density level of a point whose density level is not 0, and N is the number thereof. In FIG. 2, it is V-5.4.

In this example, B = 1.5, b-3, but this should also be determined experimentally like the above-mentioned α, β, γ, and δ.

FIG. 11 shows an example of the configuration of an apparatus for carrying out the present invention. Here, (7) is a scanning device, (8) is a first storage device that stores a multivalued character pattern, (9) is a binarization device,
αq is a high threshold value determining device, αq is a second storage device for storing binary character patterns, and (b) is a thinning device.

(to) is a third storage device that stores thinned character patterns, α◆ is a local processing device, and (g) is a recognition logic device that determines character names. The character pattern (2) on the recording medium (1) is converted into an electric signal proportional to the density by a scanning device (7) K composed of a lamp, a lens, and a semiconductor optical sensor (not shown). It is converted into a 16-step digital signal representing the level (first storage device (8)
will be forwarded to. The binarization device (9) converts the multivalued character pattern (3) in the first storage device (8) to a low threshold 'l'.
Convert 200 million black and white images using HL and send to the second storage device αQK. The thin line processing device (b) is 1, and the second storage device (g) is
The binary character pattern (4) in the first memory device fill (
Using the multi-value character pattern (3) of 8), black pixels below the high threshold value THi determined by the high threshold value determination device (g) are determined. Next k, the local processing device a◆ is connected to the third storage device α
Contrast information, etc., is obtained by referring to the multi-valued character pattern (3) in the first storage device (8), targeting black pixels of less than the threshold THH of the thin line character pattern (6) in the field, and Re-evaluate validity. The character name of the final character pattern obtained through the above processing is determined by the recognition logic device (2)k.

In addition, although the case where the binarization device, the thinning device 9 local processing device, and other devices that individually perform independent processing are combined is described above, the present invention is not limited to this, and can be applied to some or all of these devices. It may be processed by a general-purpose arithmetic processing device.

Moreover, although the case of handwritten characters has been described above, it may also be used to read printed characters or symbols similar to characters.

As described above, in the character reading method according to the present invention, among the black pixels □ of the binary character pattern 7 obtained using a low threshold value,
In the original multi-valued character pattern, black pixels that are less than a high threshold are thinned, and then the validity of black pixels that are less than a high threshold is re-determined by local processing using contrast information. It has the great advantage of being able to read with high precision even low-quality characters with blurred parts, unnecessary contact, crushing, or noise.

[Brief explanation of the drawing]

Figure 1 shows an example of a character pattern to be read, Figure 2 shows a character pattern expressed in multilevel density levels, and Figure 3 shows a character pattern binarized with a low threshold. 4 is a diagram showing a character pattern binarized with a high threshold value, FIG. 5 is a diagram showing a thinned character pattern, FIG. 6 is a diagram showing an example of a local processing mask, and FIG. 7 is a diagram showing an example of a mask for local processing. The figure shows an example of a function for obtaining contrast information.
The figure shows the finally obtained character butter /, Figure 9 is a flowchart of the processing procedure, Figure 10 is a flowchart of the process for determining the average line width of characters, and Figure 11 is a block diagram showing an example of the configuration of the device. It is a diagram. In the figure, (1) is the recording medium, (2) is the character pattern, (
3) is a multivalued character pattern, (4) is a character pattern binarized using THL, and (5) is a character pattern binarized using THH. (6) is a thin line character pattern, (7) is a scanning device, (8)
is the first storage device, (9) is the binarization device, (g) is the high threshold value determination device, α force is the second storage device, (2) is the thinning device, (2) is the third storage device, α◆ is local processing device,
(to) is a recognition logic device. In the drawings, the same or corresponding parts are designated by the same reference numerals. Agent Makoto Kuzuno - Figure 2 322222 1 3C6224A! 1 4D4 141 D33C5 32 31st! I XXXX X× Figure 6 Figure 7 Figure 8 xx Figure 9 Figure 10 433-

Claims (1)

[Claims] In a character reading method that reads characters recorded on a recording medium such as a form, a scanning means scans the characters and converts them photoelectrically, and a multilevel character pattern obtained by the scanning means is converted into black and white using a low threshold. A binarization means for digitizing, a threshold determining means for determining a high threshold such that the average line ring of 200 million patterns of characters becomes a predetermined value, and regarding the lI & li elements determined to be black by the binarization means. converting only black pixels having density levels below the high threshold into white pixels 1111;
111 conversion means, and targets black pixels having a density level below the high threshold value in the thinned pattern, and determines whether the black pixel should be made white or black based on the density levels of the black pixel and its neighboring pixels. A character reading method characterized by comprising a local processing means for redetermining.