JPH03276380A - Character recognizing device - Google Patents

Abstract

PURPOSE:To eliminate the generation of missing of a character line and the disappearance of blank owing to normalization by detecting the line width of a character pattern and executing processing that the character line less than the first fixed width is made thick and the character line more than the second fixed width is made thin. CONSTITUTION:A character segmenting part 10 segments a character pattern at one character unit from the quantized picture data of an inputted character medium. A character frame detecting part 12 detects a character circumscribing frame at every the respective character pattern and finds the character size of the character pattern from the position of the character circumscribing frame. A normalized constant determining part 14 sets a normalized constant corresponding to the character size, a normalizing part 16 normalizes the character pattern based on the normalized constant and a recognizing part 18 executes the recognization of the normalized character pattern. A line width detecting part 20 detects the line width of the character pattern. In a line width converting part 22, both processing that a thickening processing part 28 makes a character line thick for the character pattern of which the line width becomes less than a first fixed width and that a thinning processing part 30 makes the character line thin for the character pattern of which the line width becomes more than a second fixed width is executed.

Description

[Detailed description of the invention] (Industrial application field) The present invention relates to a character recognition device.

(Prior Art) In general printed documents such as newspapers, books, and magazines, the font size of the main text and the font size of headings or titles are significantly different. In order to recognize these characters of different sizes using the same recognition method and circuit, it is necessary to normalize the size of the character pattern.

One method of normalization is to simply thin out character patterns when the character size is larger than a standard value. This method converts the original character pattern into a pattern that is reduced by a normalization constant such as 1/2.1/3.1/4, etc., increasing the processing speed of the character recognition device and reducing the hardware cost. It is possible to reduce the scale.

(Problem to be Solved by the Invention) However, in general printed documents, the main font is Mincho, and the font size of the main text is approximately 3 mm, and the font size of headings and titles is often approximately 12 mm or more.

A normalized pattern memory that stores character patterns with a character size of 3 mm usually has a capacity of 64 x 64 pixels or 128 x 128 pixels, or a large A character pattern with a character size of 12 mm must be compressed (normalized) to 1/4 and stored in a normalized pattern memory.

On the other hand, the line width of the horizontal stroke of a Mincho typeface character with a character size of 12 mm is approximately 0.3 mm, and when the image pattern of this character is obtained using a commonly used scanner with a resolution of 300 clpi, the line width of the horizontal stroke is 3 pixels. Therefore, if the pattern of a Mincho font with a character size of 12 mm is simply thinned out to 1/4, horizontal strokes may be missing. In addition, headings and titles often use characters with extremely thick lines, such as Gothic fonts or specially designed fonts, or in general, as the character lines become thicker, the spaces between the character lines tend to become narrower. Therefore, if the line width becomes extremely thick, the space between the character lines may become too narrow, causing the space to collapse.

These points will be explained in more detail with reference to Figure 9. Figure 5 is a diagram used to explain missing horizontal strokes. Figure 5 (A) shows an example of Mincho typeface characters, and Figure 5 (CB) shows the circle surrounded by the - dotted chain line in Figure 5 (A). An example of a horizontal stroke with a line width of 3 pixels is shown enlarged.

In Fig. 5(B), the pixel position when read by the scanner is represented by 8% dotted line grid points, and the pixel position when thinned out to 1/4 is represented by straight line grid points, and the character pattern is further quantized into black and white binary. The black pixels (character line part) are represented by grid points with black circles, and the white pixels (character background part) are represented by unmarked grid points. On the pattern memory that stores character patterns before normalization, an X-axis representing the scanner's scanning position W and a Y-axis representing the scanner's sub-scanning position are set, and dotted lines parallel to the X- and Y-axes are set. T
Dotted grid points are set at positions where x and Tv intersect and correspond to pixel positions read by the scanner.

When simply thinning out to 1/4, every fourth dotted line Tx and TV\! counting from the origin of the X-Y coordinate system set on the pattern. , Jx, and JV are selected, and the white pixels or black pixels existing at the grid points of these English spellings are directly adopted as pixels of the character pattern normalized by thinning.

Therefore, as shown in FIG. 5(B), when a horizontal stroke with a line width of 3 pixels is located between straight lines Jx that are in contact with 11g5,
There are no black pixels left in the character pattern after thinning that correspond to a horizontal stroke with a line width of 3 pixels, and as a result, the horizontal stroke is missing.

Figure 6 is a diagram used to explain the collapse of spaces between characters. Figure 6 (A) shows an example of Gothic characters, and Figure 6 (B) shows the arrows in Figure 6 (A). An example of a blank space between character lines in a portion with a width of 3 pixels is shown enlarged.

In FIG. 6(CB), the same components as those shown in FIG. 5(B) are given the same reference numerals, and detailed explanation thereof will be omitted.

Figure 6 (8) If a blank space of 1 pixel width and 3 pixels width is located between adjacent M lines Jx as shown in Fig. As a result, the spaces between character lines are collapsed.

If you normalize the character pattern by simply thinning it out in this way,
If the line width of a character pattern with a large character size is thin or thick, character lines may be missing or the spaces between character lines may be collapsed in the character pattern after normalization. impossibility) increases and recognition accuracy deteriorates.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character recognition device H that eliminates or reduces missing character lines and collapsed spaces due to normalization, in order to solve the above-mentioned conventional problems.

(Means for Solving the Problem) In order to achieve this object, the character recognition system of the present invention includes a character cutting section that cuts out a character pattern for each character from quantized image data of a character medium; A character frame detection unit that detects a character circumscribing frame for each character pattern and determines the character size of the character pattern from the position of the character circumscribing frame, a normalization constant determination unit that sets a normalization constant according to the character size, A line width detection unit that detects the line width of a character pattern in a character recognition device comprising a normalization unit that normalizes a character pattern based on a constant and a recognition unit that recognizes the normalized character pattern. , thicken the character lines of character patterns whose line width is less than or equal to the first predetermined width! - a line width conversion unit that performs both or one of the processing of - and the second processing of thinning character lines of a character pattern whose line width is equal to or greater than a second predetermined width. Features.

(Function) According to such a configuration, the line width of the character pattern is detected,
A first process of thickening the character lines of a character pattern whose line width is less than or equal to a first predetermined width, and/or a second process of thinning the line width of a character pattern whose line width is greater than or equal to a second predetermined width. Do the following.

If a character pattern is detected whose line width is less than or equal to the first predetermined width, this character pattern has thin character lines that may be missing after normalization. Prevent lines from missing.

In addition, if a character pattern whose line width is equal to or larger than the second predetermined width is detected, this character pattern has spaces between character lines that may be collapsed after normalization, so the width of the spaces between character lines can be reduced by narrowing the character lines. Make it wider to prevent spaces between character lines from being collapsed after normalization.

Therefore, by performing the first process and/or the second process, it is possible to eliminate or reduce the occurrence of missing character lines and/or collapsed spaces due to normalization.

(Examples) Examples of the present invention will be described below with reference to the drawings.

FIG. 1 is a functional block diagram for explaining the invention in detail. As shown in the figure, the character recognition device of this embodiment includes a character cutting section 10 that cuts out character patterns character by character from quantized image data of a character medium, and a character circumscribing frame bubble for each character pattern. A character frame detection unit 12 detects the character size of a character pattern from the position of a character circumscribing frame, a normalization constant determination unit 14 sets a normalization constant according to the character size, and a It includes a normalization section 16 that normalizes a pattern, a recognition section 18 that recognizes a normalized character pattern, a line width detection section 20 that roughly detects the line width of the character pattern, and a line width detection section 20 that roughly detects the line width of the character pattern. A first process of thickening the character lines of a character pattern whose line width is equal to or larger than a second predetermined width is performed, and a second process of thinning the character lines of a character pattern whose line width is equal to or greater than a second predetermined width. The line width conversion section 22 is also provided.

In FIG. 1, 24 and 26 indicate a photoelectric conversion section and a pattern register.

This embodiment will be explained in more detail below.

The photoelectric conversion unit 24 optically scans the form to be processed, receives a large optical signal LV from the form, photoelectrically converts the optical signal to obtain electrical signal image data, and converts the image data into, for example, black and white binary. Quantize.

The character cutting section 10 is connected to a photoelectric conversion section 2 in an image memory (not shown).
The image data from No. 4 is saved, the image data is scanned, character patterns are cut out character by character from the image data, and the cut out character patterns are stored in a pattern register 26.

The character frame detection unit 12 scans the character pattern in the pattern register 26 and detects the character circumscribing frame of the pattern. An X-Y coordinate system is set on the pattern register 26, and the upper, lower, left, and right end positions of the character circumscribing frame represented by this coordinate system are detected. Upper and lower end positions Y, Y, are the starting ends of the character circumscribing frame in the Y-axis direction,
The end position N, and the left and right end positions XL and Xll represent the start and end positions M of the character circumscribing frame in the x-axis direction, and the character frame detection unit 12 detects the character pattern from the top and bottom positions Y as YB. Calculate the height and further left end and right end positions XL
, XR to calculate the width of the character pattern.

The normalization constant determining unit 14 determines a normalization constant for each character from the height and width of the character pattern. If the character pattern stored in the pattern register 26 is referred to as a character pattern of interest, first, the larger value of the height and width of the character pattern of interest is detected as the character size S of the pattern. Assuming that it is known in advance that the minimum character size of the form to be processed is, for example, 3 mm, for example, when 3≦S<6, N=1.6≦S<9
When N=1/2.9≦S〈12, N=1/3,・
...When 3n≦S<3(n+1), N=i/n(n
is a natural number), and a normalization constant In corresponding to the classification is given to the character pattern of interest.

When image data is obtained using a scanner with a resolution of 300 dpi, if the character size of the character pattern of interest is 3 mm, the size of the character pattern is about 35 x 35 pixels, and if the character size is 6 mm (the size of the character pattern is 70 x
Therefore, if the normalization constant N is assigned as described above in this case, the size of the character pattern can be normalized to 35x35 to 70x70 pixels regardless of the character size. Therefore, the capacity of pattern memory to store character patterns after normalization! 128x12
It can be within 8 pixels.

The line width detection unit 2o scans the character pattern of interest in the pattern register 26 and calculates the line width of the pattern.

Although the calculation of the line width may be performed by any conventionally known suitable method, in this embodiment, for example, the line width detector 2o has a shift register configuration similar to a conventionally known filter circuit, and the following equation (1) is used. The line width W is calculated according to the approximate formula shown below.

In equation (1), Q is the total number of times that all points in the 2×2 window became black pixels, A is the total number of black pixels in the character pattern, and the line width detection unit 2o detects all the points in the 2×2 window. Count the number of times a point becomes a black pixel and the number of black pixels in the character pattern to calculate the total number of times and the total number Q, and calculate the line width WIF according to formula (1).
Calculate r.

The line width conversion section 22 of this embodiment includes a thickening processing section 28, a thinning processing section 30, a data switching section 32, and a processing selection section 34.

Based on the line width W and the normalization constant N, the processing selection unit 34 selects a character pattern of interest, a character pattern obtained by thickening the line width of the character pattern of interest by the first process (hereinafter referred to as thick pattern], and a character pattern obtained by narrowing the line width of the character pattern of interest by the second process (
Hereinafter, it is determined which of the narrow patterns) is to be used for character recognition.

For this reason, the process selection unit 34 first calculates the number of times MCNT of process executions when executing the first process and the number of process executions NCNT when executing the second process.

When the first process is performed once, the line width increases by MuP, and as a result of repeating the first process MCN 7 times to increase the thickness, a line width of the first predetermined width C, (for example, C, = 3) is obtained. Then, the predetermined width C1 can be expressed as shown in the following equation (1).

C+ = N * (MCN T * Mup+W)・
...(1) The setting value of the line width increment M u p is, for example, when all the white points touching the edge of the character line Va are to be replaced with black points, Mu = 2, and the setting value of the line width increment Mu When replacing the white points adjacent to the upper and right edges with black points, or when replacing the white points adjacent to the upper and left edges of the character line w4 with black points, Mup"1 may be used.

The following equation (2) is obtained from equation (1).

M CN T f,t is an integer greater than or equal to 0, (2)
If MCNT calculated according to the formula is a negative value, MCNT
=O. In addition, if the calculated MCNT is not an integer, the calculated MCNT is converted into an integer by, for example, rounding up, rounding down, or rounding off the decimal point.
Preferably, the right side of equation (1) is the smallest integer that is larger than the predetermined width C1! It is better to use MCNT.

Also, if you perform the second process once, the line width will be M. , and as a result of thinning the second process by repeating the NCN seven times, if a line width of the second predetermined width C2 (for example, c2=3) is obtained, the predetermined width C2 is calculated as shown in the following equation (3) It can be expressed as

C2=N*(W-NCNT*Mo5)--(3) Decrement M of line width. The setting value of N is, for example, llI on the edge of the character line.
When replacing all touching white points with black points, M, , = 2
If you want to replace the white dots adjacent to the lower and right edges of the character line with black dots, or add w4 to the upper and left edges of the character line.
M to change the touching white point to a black point. ,=1.

The following equation (4) is obtained from equation (3).

NCNT is an integer greater than or equal to 0, and when NCNT calculated according to equation (4) is a negative value, NCNT=◯.

In addition, if the calculated NCNT is not an integer, the calculated NCNT is converted into an integer by rounding up, rounding down, or rounding off to the nearest whole number, or preferably,
It is preferable to set the maximum integer VNC:NT that is smaller than the right side of equation (3) or the predetermined width C2.

Processing counts MCNT and NCNT obtained as described above
are both 0, and when both MCNT and NCNT are not 0, the processing selection unit 34 outputs a first instruction signal indicating that the character pattern of interest in the pattern register 26 is to be used for character recognition.

When the number of processes MCNT≠0 and NCNT=O indicates that the line width of the pattern of interest is less than or equal to the first predetermined width C7, the process selection unit 34 selects a thick pattern whose line width is thickened in the first process as a character. A second instruction signal representing use for recognition is output.

When the rough processing number NCNT≠0 and MCNT=○, it means that the line width of the pattern of interest is less than or equal to the second predetermined width C2, so the narrow pattern whose line width is narrowed in the second process is used for character recognition. A third instruction signal indicating use is output.

When the thickening processing section 28 receives the second instruction signal from the processing selection section 34, it performs a process of thickening the character lines of the character pattern of interest. The thick processing section 28 of this embodiment has a shift register configuration similar to a conventionally well-known filter configuration, and performs the first process of mirroring the character Sa using, for example, a 3.times.3 window.

Figure 2 shows a 3x3 window, Figure 3 (A) and (B)
FIG. 2 is a diagram showing a character line before the first process and a character line after the first process. FIG. 3(A) shows a character line (vertical stroke) having a line width W less than a predetermined width C1 and all white pixels adjacent to the edge of this character line, and FIG. 3(B) shows (
Figure 3 shows the character lines obtained by enlarging the character lines shown in A) in the first process. In these figures, white circles represent white pixels, black circles represent black pixels, and white circles with dots inside the circles represent white. Represents a pixel that has been changed from a pixel to a black pixel.

The thick processing unit 28 changes the pixel of interest a9 to a black pixel when the pixel of interest aS in the 3x3 window (see FIG. 2) is a white pixel and any of the surrounding pixels a, to a6 is a black pixel, All white pixels that are in contact wA with the edge of a character line with a predetermined width C1 or less are made black pixels, and the line width is increased by 2 (Fig. 3 (A) and (B)
? (see).

When the narrowing processing section 30 receives the third instruction signal from the processing selection section 34, it performs a process of narrowing the character lines of the character pattern of interest. The thinning processing section 30 of this embodiment has a shift register structure similar to a conventionally known filter structure, and performs a second process of thinning a character line using, for example, a 3.times.3 window.

FIGS. 4(A) and 4(8) are diagrams showing character lines before the second process and character lines after the second process. Figure 4 (A) shows a character line (vertical stroke) having a line width W less than a predetermined width C2 and all of the white pixels that are in contact with the edge of this character line. The character line obtained by thinning the character line shown in (A) through the second process 1 is shown. In FIG. 4, white circles represent white pixels, black circles represent black pixels, and white circles with an x inside the circle represent pixels that have been changed from black pixels to white pixels.

The narrowing processing unit 30 calculates the number H1 of the number of times a black pixel appears next to a white pixel when the target pixel a9 of 3×3 breath (see FIG. 2) is a black pixel, and the surrounding pixels a to a8 are sequentially scanned, and the black pixel is a black pixel. Count the number of times H2 that a white pixel appears next to a pixel, and when the sum of these counted values H,+H2 becomes H,+H2=2, change the pixel of interest ae% to a white pixel, and restrict character lines with a predetermined width C2 or more. All black pixels located at are made white pixels, and the line width is decreased by 2 (see FIGS. 4(A) and 4(B)).

The data switching unit 32 selects the character pattern of interest from the pattern register 26 when the first instruction signal is input from the processing selection unit 34, and selects the target character pattern from the thick processing unit 2 when the second instruction signal is input.
When the third instruction signal is inputted, the thinning pattern created by the thinning processing section 30 is inputted, and the inputted pattern is outputted to the normalization section 16.

The normalization unit 16 of this embodiment normalizes the pattern by thinning out the character pattern of interest, large pattern, or narrow pattern input from the data switching unit 32 in the same manner as in the past. In this case, the pattern input from the data switching section 32 is output as is to the recognition section 18, and O<N<1.
In this case, a pattern obtained by reducing the pattern inputted from the data switching section 32 by N times is output to the recognition section 18 . Hereinafter, the pattern output by the normalization unit 6 will be referred to as a normalized pattern.

The recognition unit 18 extracts the character features according to the normalized pattern, compares the character features with a dictionary (not shown), performs literary journal m, and outputs, for example, a character code as the recognition result to the next stage device.

The first predetermined width C7 represents the average line width of character lines in the normalized thick pattern, and when this average line width is a value or in what numerical range it is normalized. The predetermined width C7 is set to an arbitrary suitable value by statistically examining in advance whether or not missing strokes in the cursive pattern can be substantially eliminated. Similarly, the second predetermined width C2 represents the average line width of the character lines in the normalized pattern of the narrow pattern, and what value is this average line width? ! It is statistically investigated in advance whether it is possible to substantially eliminate the collapse of spaces between character lines in the normalized pattern in the case of a normalization pattern, and the predetermined width C2 is set to an arbitrary suitable value that can substantially eliminate the collapse of spaces.

The present invention is not limited to the embodiments described above, and therefore, the operation, configuration, processing flow, numerical conditions, etc. of each component can be changed as desired.

In the embodiment described above, the number of processing times MCNT and NCNT
It is determined which of the character pattern of interest, a thick pattern, or a thin pattern is to be used for character recognition according to the numerical value of .In addition, the line width W of the character pattern of interest For example, if W<10, a thicker pattern obtained by repeating the first process twice for the character pattern of interest may be used. If 20≦W, use the narrow pattern obtained by repeating the second process twice for the character pattern of interest.1) Use the pattern to be used for character recognition; if 1o≦W<20, use the pattern of the character of interest. may be used as a pattern for character recognition.

In addition, in the embodiment described above, the first and second predetermined widths c1 and C2, which are common to all the normalization constants N, are used, or the 118 normalization constants N of these predetermined widths C, C2 are used.
It may be changed depending on the size of the value of . For example, when the normalization constant N=1, C,=C2=3,N=
When N=1/2, C+=C2=4, and when N=1/3, C,=C2=5, . . . Further, the values of the predetermined widths C1 and C2 may or may not be equal.

The line width W is l! other than that defined by the above equation (1). The J width may be used, and conventionally known methods other than thinning may be used to normalize the pattern.

(Effects of the Invention) As is clear from the above description, according to the character recognition device of the present invention, the first process of thickening the character lines of the character pattern in which the line width is equal to or less than the first predetermined width; Alternatively, a second process of thinning the line width of a character pattern whose line width is equal to or greater than a second predetermined width is performed.

If a character pattern is detected whose line width is less than or equal to the first predetermined width, this character pattern has thin character lines that may be missing after normalization. Prevent missing lines. Also, if a character pattern with a line width greater than or equal to the second predetermined width is detected, this character pattern has spaces between character lines that may be collapsed after normalization, so by narrowing the character lines, the character pattern can be changed. ! Widen the width of the dark space to prevent the space between character lines from being collapsed after normalization.

Therefore, by performing the first process and/or the second process, it is possible to eliminate or reduce the occurrence of missing character lines and/or collapsed spaces due to normalization.

As a result, for example, a character pattern with a large character size and thin line width can be normalized without or with almost no character lines missing, and a character pattern with a large character size and thick line width can be normalized without collapsing the amount of character space. Alternatively, it is possible to normalize with almost no distortion, thereby reducing misreading and unreadability and improving character recognition accuracy.

[Brief explanation of drawings]

Fig. 1 is a functional block diagram schematically showing the configuration of an embodiment of the present invention, Fig. 2 is a diagram showing a 3 x 3 window, and Figs. 3 (A) and (B) are before and after the first processing. Figures 4(A) and 4(B) are diagrams showing character lines before and after the second process; Figures 5(A) and 5(B) are diagrams illustrating missing horizontal strokes; FIGS. 6(A) and 6(B) are diagrams for explaining the collapse of character amount spaces. 1o...Character cutting section, 12...Character frame detection section 1
4--Normalization constant determining section 16... Normalization section 18... Recognition section 2o...
- Line width detection section, 22...Line width conversion section.

Claims (2)

[Claims] (1) A character cutting unit that cuts out a character pattern one character at a time from quantized image data of a character medium, and detects a character circumscribing frame for each character pattern, and detects the character circumscribing frame from the position of the character circumscribing frame. a character frame detection unit that determines the size; a normalization constant determination unit that sets a normalization constant according to the character size; a normalization unit that normalizes the character pattern based on the normalization constant; A character recognition device comprising: a recognition unit that recognizes a character pattern, the line width detection unit that detects the line width of the character pattern; and a character recognition unit that detects the line width of the character pattern; a line width conversion unit that performs both or either of a first process of thickening a line and a second process of thinning a character line of a character pattern in which the line width is equal to or larger than a second predetermined width; A character recognition device comprising: (2) The character recognition device according to claim 1, wherein the line width conversion unit sets the first and second predetermined widths by changing them according to the magnitude of a value of a normalization constant.