JPH0517598B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a method for cutting out characters in a character reading device such as an optical character reading device.

(Prior Art) A block diagram of a general optical character reader (OCR) is shown in FIG. 5a. OCR includes a photoelectric conversion section 11, a multi-value data buffer 12, and a filter circuit 1.
3 and a recognition section 14, and the recognition section 14 is composed of a binary data buffer 14a and a recognition control circuit 14b. The character portion on the form is photoelectrically converted into a multi-gradation digital signal by the photoelectric converter 11, and then stored in the multi-value data buffer 12 as multi-value data. The stored multi-value data is divided into "0" and "1" by the filter circuit 13.
After being digitized and smoothed, it is stored in the binary data buffer 14a of the recognition unit 14 as binary data. At this time, in the binary data buffer 14a,
As explained later, in order to normalize the font size,
Binary data thinned out according to a predetermined compression rate is stored. Based on the binary data stored in the binary data buffer 14a, the recognition control unit 14b
Character recognition is performed. Note that, hereinafter, "0" in the binary data is referred to as a white point, and "1" is referred to as a black point.

Prior to character recognition, binary data buffer 14
For the binary data in a, there is a task of determining the recognition target range, which is the range to be recognized, and this is called character extraction.

The internal configuration of the binary data buffer 14a is shown in FIG. 5b. The binary data buffer 14a is a data buffer 141 that stores binary data, and a projection that stores the result (projection image) of projecting a character pattern portion (i.e., a collection of black dots) in the x direction of the binary data stored in the data buffer 141. Batsuhua 14
2, and a projection buffer 143 that similarly stores the results of projection in the y direction.

Further, examples of character patterns stored in the binary data buffer 14a are shown in FIGS. 6a and 6b.

Next, a character extraction procedure under the control of the recognition control circuit 14b of the recognition section 14 will be described with reference to FIG. In FIG. 6a, character patterns 20, 21,
22 is stored in an area 23 corresponding to two characters in the data buffer 141, and projected images 24, 25, and 26 obtained by projecting these in the y direction are stored in the projection buffer 143. These projected images 24,2
The range corresponding to the projected image 25 from the breaks 5 and 26 can be regarded as the width of one character. Furthermore,
As shown in FIG. 6b, a projected image 27 of the character pattern 21 projected in the x direction is stored in the projection buffer 142. The range corresponding to this projected image 27 is regarded as the height of the character. Therefore, the projected image 25,
A rectangular area 28 surrounded by 27 is called a character frame.
This character frame serves as a circumscribing frame for a single character pattern within the recognition target range, and indicates the size of the character (hereinafter, character size is defined as the sum of the height and width of the character frame. ).

Next, the normalization of the character size during the above-mentioned character extraction will be explained.

First, an example of thinning out binary data is shown in FIG.
In the figure, 1 in the multilevel data buffer 12
1, 12, . . . 17, . . . , 67 indicate address numbers, and multi-value data is stored in each address. 11 in the binary data buffer 14a,
13, 15, 17, . . . , 57 indicate storage positions in the multi-value data buffer 12 of multi-value data corresponding to the binary data stored in the binary data buffer 14a. This figure shows a state in which the binary data that has been binarized by the filter circuit 13 is thinned out every other data and stored in the binary data buffer 14a.

In this way, the rate of thinning, such as thinning out every other item or every two items, is determined in advance by the form format, and the method for determining it is as follows.

The position where you write the characters on the form is usually 1 character.
It is determined by the entry frame for each entry. Multiple types of entry frame sizes are available, and it is necessary to normalize the size of written characters assuming that they will change depending on the size of the entry frame. Specifically, normalization is a compression/enlargement process that converts a writing frame into a frame of a specific size (hereinafter referred to as a standard pattern), which is determined so that the font size should be written at a constant size. Yes, in order to match the height and width of the standard pattern with the height and width of the entry frame, determine the compression rate of how much to compress, and enter this compression rate as the thinning rate in the format.

FIGS. 8a and 8b are diagrams for explaining normalization using a writing frame, and 31 and 32 are writing frames, and 33 is a standard pattern having a height 33a and a width 33b. A compression ratio is given to each of the entry frames 31 and 32 in the x direction and the y direction.

(Problems to be Solved by the Invention) However, the normalization using the form entry frame in the conventional character extraction method described above has the following problems.

Since the compression rate is determined by the size of the entry frame, it does not contribute to normalizing the size of the actually written characters. For example, as shown in FIG. 8b, an actually written character 34 is written in the writing frame 31 smaller than the standard pattern 33 (that is, very small compared to the writing frame 31), and the writing frame 31 is written in the standard pattern. When compressed to fit 33, the character 34
Since the size of the cutout itself becomes smaller and the cut out characters become smaller, it becomes difficult to accurately recognize the characters.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character extraction method that eliminates the above-mentioned problems and allows accurate character recognition by normalization according to the size of the actually written character.

(Means for Solving the Problems) In order to solve the above-mentioned problems, the present invention normalizes binary data obtained by reading characters on a form, and adds character patterns to the normalized binary data. In the character extraction method of a character reading device, which extracts characters by finding a character frame, which is a circumscribed rectangular area, the binary data obtained by reading characters on a form is extracted using the entry frame of the form format and a standard pattern. A first step in which a compression rate of 1 is determined, and a character frame of a character pattern is determined for the binary data normalized by the first compression rate, and the size of the character frame determined in the first step is a second step of calculating a second compression ratio based on the size of the character pattern and the size of the standard pattern; A third step of determining whether or not to correct the second compression ratio based on the ratio to the inner area; and determining whether the compression ratio in the height direction or the lateral direction is corrected based on the determination in the third step. a fourth step of determining a third compression ratio by correcting the first compression ratio according to the second compression ratio corrected to be equal to the smaller one or the second compression ratio that is not corrected; and a fifth step of normalizing the binary data obtained by reading the characters by thinning out according to a third compression ratio and cutting out the characters again.

(Operation) The character extraction method of the character reading device of the present invention operates as follows. In the first step, as in the past, binary data is normalized according to the first compression rate determined by the form format (for example, depending on the size of the entry frame on the form and the size of the standard pattern). Obtain a character frame for value data (binary data by thinning). In the second step, a third compression ratio is determined based on the size of the character frame and the size of the standard pattern obtained in this way. In the second step, the second compression ratio is corrected based on the ratio between the number of black dots within the character frame and the area of the character frame, and the ratio between the number of black dots within the inclusive line (convex hull) and the area within the inclusive line. It is determined whether or not. For example, a character pattern with almost no white dots inside the character frame or within the comprehensive line,
A second compression ratio is determined to compensate. Fourth
In this step, the second compression rate corrected based on the determination in the third step (for example, vertically long and horizontally long character patterns are aligned to the smaller compression rate for both height and width), or The third compression rate is determined by correcting the first compression rate according to the compression rate of (for example, by multiplying the corrected second compression rate or the second compression rate and the first compression rate) It is determined).
In the fifth step, normalization is performed to thin out the binary data obtained by reading the characters on the form according to the third compression ratio, and character extraction is performed again. In this way, in addition to the second compression rate based on the size of the character frame, the second compression rate is also corrected based on the shape of the characters (by enclosing lines) to achieve the first compression rate based on the conventional form format. By correcting the ratio, a third compression ratio that is faithful to the actually written characters is determined, and normalization is performed using this third compression ratio, so optimal character extraction can be performed. In addition, it becomes possible to accurately recognize characters. Therefore, the problems of the prior art described above can be solved.

(Embodiment) FIG. 1 is a flowchart showing the processing procedure of a character segmentation method of an optical character reading device in an embodiment of the present invention, and shows the recognition unit 14 described in FIG.
This is performed under the control of the recognition control circuit 14b. FIGS. 2a, 2b, and 2c are diagrams showing examples of character patterns for specifically explaining the normalization of character size in character extraction in this embodiment.

First, the character extraction procedure will be explained with reference to FIG.

As in the past, binary data is thinned out at the compression rate determined in advance by the form format, that is, the compression rate determined by the entry frame and standard pattern (y-direction compression rate PY ₁ , x-direction compression rate PX ₁ ). Character cutting is performed on the character, and a character frame is determined. The height Y and width X of this character frame are the standard pattern height Y ₀
and the width X ₀ , and the compression ratio in the y direction PY
=Y ₀ /Y, the compression ratio in the x direction PX = X ₀ /X is determined (step 1). Next, the number of black dots in the character frame is compared with the area of the character frame, and the magnitude of the comparison result is determined with respect to a certain constant _C1 (step 2).

In step 2, (number of black dots) / (character frame area) < C ₁
In this case, find the comprehensive line (convex hull) of the character pattern, compare the number of black dots within the comprehensive line and the area within the comprehensive line, and judge the magnitude of the comparison result for a certain constant C ₂ . (Step 3).

In step 3, (number of black dots) / (area within the comprehensive line) <
When C ₂ , the compression ratio PY, PX obtained in step 1
The compression ratio PY 1 obtained from the respective entry boxes is PY ₁ ,
Multiply by PX ₁ to find the final compression ratio PY ₀ and PX ₀ (step 9). Character extraction is performed again on the thinned out binary data based on the obtained compression ratios PY ₀ and PX ₀ (step 10).

On the other hand, in step 2, if (number of black points)/(character frame area)≧C ₁ , the height Y and width X of the character frame are compared, and the comparison result is calculated for certain constants C ₃ and C ₄ . Judging the size (steps 4 and 5). Y/X>C ₃ in step 4, or X/Y>C ₄ in step 5
, or if (number of sunspots)/(area within the comprehensive line)≧C ₂ in step 3, PX and PY are compared (step 7). If PX<PY in step 7
Put the value of PX into PY (step 8a), and if PX≧PY, put the value of PY into PX (step 8b). That is, in steps 7, 8a, and 8b, PY and PX are adjusted to the value of the smaller compression ratio. Multiply this y-direction compression ratio PY by the compression ratio PY ₁ obtained from the entry frame to obtain the final y-direction compression ratio PY ₀ , and similarly multiply the x-direction compression ratio PX by the compression ratio PX ₁ to obtain x The final compression ratio PX ₀ in the direction is determined (step 9). Character extraction is performed again on the thinned out binary data based on the compression ratios PY ₀ and PX ₀ determined in step 9 (step 10).

Also, if Y/X≦C ₃ in step 4 and X/Y≦C ₄ in step 5, the compression ratio is not changed, so 1 is entered in both PX and PY (step 6), and steps 9 and 10 are set. I do.

In this way, after normalization is performed according to the size of the actually written characters using the character frame and comprehensive line, character segmentation is performed again.

Next, a character cutting procedure will be specifically explained with reference to character pattern examples shown in FIGS. 2a, b, and c.

First, if the character patterns are three types of character patterns 101, 102, 103 of "1" shown in FIG. From X, find each compression ratio PY, PX. In step 2, if the constant C ₁ is C ₁ ≒ 1, character pattern 101 has almost no white dots within the character frame 101a, so the processing from step 4 onwards is performed, and character patterns 102 and 103 are Because there are many white spots,
Processing in step 3 is performed. Therefore, the character patterns 102 and 103 are as shown in FIG. 2b.
If the respective comprehensive lines 102b and 103b are obtained and the constant C ₂ is C ₂ ≈1, then the character pattern 103 has almost no white dots within the comprehensive line, so the processing from step 7 onwards is performed.

FIG. 2c is an example of a character pattern for which the processing in step 4 is performed based on the judgment in step 2. Step 4
If the constant C ₃ is C ₃ ≒ 1, then the character pattern 101
Since it is vertically long, the processing from step 7 onwards is performed. Further, the character patterns 104 and 105 are processed in step 5. The constant C ₄ in step 5 is
If C ₄ ≒ 1, almost square character pattern 10
4 undergoes the processing from step 6 onwards, and the horizontally long character pattern 105 undergoes the processing from step 7 onwards.

Next, an example of a procedure for determining a comprehensive line will be explained.
The character pattern used here is shown in FIG.

First, with the x direction as the main scan and the y direction as the sub scan, take the number of white points from the left end of the character frame 107 of the character pattern 106 to the black point, and take the y direction as the horizontal axis and the vertical axis as (character width) - ( A histogram of the obtained white point numbers is shown in FIG. 4a. At this time, the maximum value A of the histogram indicates the character width, and the point at which this maximum value is reached is called the maximum peak. A comprehensive line is found for each tip point of this histogram. First, look at the positive direction of the y direction, which is the right side from the maximum peak, and calculate the rate of change to each tip as shown in Figure 4b.
Shown below. The distance from the maximum peak point to each tip in the y direction is expressed as the denominator, and the distance in the x direction is expressed as the numerator. In other words, when expressed as (ordinate, abscissa), when the maximum peak is m, n, m-M, n
When looking at the point -N, the rate of change is expressed as M/N. With respect to the rate of change thus determined, the point where the rate of change is the smallest is defined as the first global contact point (FIG. 4, b). Next, the rate of change from this first comprehensive contact point is taken for each tip on the right side of the first comprehensive contact point on the histogram, and the point where the calculated rate of change is the smallest is set as the second comprehensive contact point ( (c in Figure 4). below,
Continue until the right end d in the y direction becomes a comprehensive contact point.
In this way, all comprehensive contacts on the right side of the maximum peak are found, and then comprehensive contacts in the negative direction of the y direction on the left side of the maximum peak are found.

Through the above steps, all comprehensive contact points on the histogram are found, and then a comprehensive line connecting the comprehensive contact points is found.
FIG. 4c shows a method for determining the maximum peak and the comprehensive contact point or the comprehensive line between the comprehensive contact points.
P _p and P _x are comprehensive contacts, and the rate of change of P _x based on P _p is M/N. Assuming that there is no comprehensive contact between P _p and P _x , the rate of change from P _p is M /N or less and which is equally distant from P _p in the y direction, the point with the largest rate of change (in this case, the white point) is defined as the point on the comprehensive line. Fourth
In Figure c, if the rate of change from P _p to P ₂ is M-1/N-1, if M-1/N-1>M/N, P ₂ is not on the comprehensive line and N-1 When /N-1≦M/N, P ₂ is on the comprehensive line. If P ₂ is on the comprehensive line, then y from P _p
P 1 whose distance in the direction is equal to P ₂ and whose rate of change from P _p is M-2/N- ₁ is not on the comprehensive line.

From the above, the comprehensive line 108 of the left half of the character pattern 106 shown in FIG. 3 can be found by finding the collection of points on the obtained comprehensive line, the maximum peak, and the comprehensive contact point.

On the other hand, from the right end of the character frame 107 in FIG. 3, the negative direction in the x direction is used as the main scan, and the direction in the y direction is used as the sub scan, and the number of white points is taken until it hits a black point. A comprehensive line 109 for the right half of the character pattern 106 shown in FIG. 3 is determined. As a result of the above, the inclusive line 11 of the character pattern 106 in FIG.
Find 0.

Although the present embodiment describes a character cutting method for an optical character reading device, the method of the present invention is also applicable to character reading devices other than optical type.

(Effects of the Invention) As explained in detail above, according to the present invention, binary data is normalized faithfully to the size and shape of the character actually written using the character frame and comprehensive line. Optimal character extraction can be performed. Therefore, it becomes possible to accurately recognize characters.

[Brief explanation of the drawing]

Fig. 1 is a flowchart showing an embodiment of the present invention, Fig. 2 a, b, and c are explanatory diagrams of normalization in this embodiment, Fig. 3 is a diagram showing an example of a character pattern, and Fig. 4 a,
Figures b and c are diagrams explaining the determination of comprehensive lines, Figures 5 a and b are block diagrams of general OCR, Figures 6 a and b are diagrams showing examples of character patterns, and Figure 7 is thinning of binary data. Figures 8a and 8b, which illustrate examples, are explanatory diagrams of normalization using entry boxes. 11...Photoelectric conversion unit, 12...Multi-value data buffer, 13...Filter circuit, 14...Recognition unit,
14a...Binary data buffer, 14b...Recognition control circuit.

Claims (1)

[Claims] 1. Normalize binary data obtained by reading characters on a form, and extract characters by finding a character frame, which is a circumscribed rectangular area of a character pattern, from the normalized binary data. In a character extraction method for a character reading device, a first compression rate is determined for binary data obtained by reading characters on a form using a form entry frame and a standard pattern, and the first compression rate is A first step of determining a character frame of a character pattern for the normalized binary data, and a second compression ratio based on the size of the character frame determined in the first step and the size of the standard pattern. A second compression ratio is determined based on the second step to be determined, the ratio of the number of black dots in the character frame to the area of the character frame, and the ratio of the number of black dots in the comprehensive line of the character pattern to the area within the comprehensive line. a third step of determining whether or not to correct; and a second compression ratio corrected to be equal to the smaller of the compression ratio in the height direction or the lateral direction based on the determination in the third step; a fourth step of determining a third compression rate by correcting the first compression rate according to the uncorrected second compression rate; 1. A method for cutting out characters for a character reading device, comprising: a fifth step of normalizing the characters by thinning out according to a compression ratio of the characters and cutting out the characters again.