JPH02165288A - Picture processor - Google Patents

Abstract

PURPOSE:To increase a character rotational speed by transforming the rotatory coordinate of an original picture and the rotatory coordinate of a new picture using a prescribed expression when a character obliquely taken in using a camera is to be rotated, returned to its normal direction and recognized, and providing up and down counters for a control circuit. CONSTITUTION:A cylindrical object 14, in which the character is written, is illuminated by an illuminator 13 and photographed by a camera 11, the picture is processed in a picture processor 10, and the written character is read. The analog output signal of the camera 11 is converted into a digital signal in an A/D converter 101 in the device 10, the output of the converter 101 is selected in a selector 102, binarized in a lookup table 103, passed through a filter circuit 104, and stored in a picture memory 105. For this picture, a character part is selected and segmented in X and Y axis projection circuits 110 and 111, rotation-processed, and stored in a picture memory 107, and at this time, the coordinate of the original picture and the coordinate of the new picture, namely, X, Y and X2, Y2, are transformed using expressions X2=AX+BY and Y2=CX+ DY and expanded for (1-2<0.5>) times.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an image processing device that recognizes images of characters, etc. captured by a camera by pattern matching, and in particular, the present invention relates to an image processing device that recognizes images of characters etc. captured by a camera by pattern matching. , relates to a device for correcting and recognizing the normal orientation.

[Conventional technology]

Various technologies have been devised to recognize characters, symbols, shapes of parts, etc. using pattern matching technology.

In pattern matching, first, it is decided that the pattern will be a binary signal normalized to a standard number of rows and two columns of pixels (for example, 16 rows and 16 columns).

Manually create a standard pattern. Next, an image of the object is captured by a camera and a pattern is cut out. From this, a pattern of the object normalized to a standard black and white binary pattern of 16 rows and 16 columns is obtained, and this is compared with the standard pattern. Then, a standard pattern in which the number of pixels matching the pattern of the target object is greater than a required value is determined to be a pattern corresponding to the target object.

When using this method to recognize characters printed on the bottom of a cylinder such as a can, the captured image is slanted in one direction, and by rotating the image, it is returned to the horizontal direction. After that, it is necessary to perform processing such as extraction and normalization.

Various image rotation techniques have been developed so far. For example, Japanese Patent Laid-Open No. 139085/1985 describes a method of rotating an image by calculating addresses after rotation and replacing pixel positions. There is.

[Problem to be solved by the invention]

However, until now, image rotation technology has not been considered as a method that can be combined with pattern matching technology.

In the method of performing pattern matching by rotating and correcting the direction of a single character, etc. to horizontal, the pattern matching includes a process of normalizing the number of pixels, so it has no meaning with normal image rotation. different.

In other words, in normal rotation, the size of the image after one rotation must of course be the same as before rotation, but in the case of pattern matching that performs normalization, it is not a problem that the size changes from before rotation. It won't happen. In addition, the only parts that need to be rotated in pattern matching are target characters and the like.

Considering these conditions, the image rotation method becomes simple and rotation processing can be performed at high speed.

SUMMARY OF THE INVENTION An object of the present invention is to provide an image processing device that can rotate an image captured obliquely, correct it in the normal direction, and then recognize the image.

[Means to solve the problem]

As shown in Figure 2, the point (■) rotated by α around the origin O is Pi, and each coordinate is (x, y).
, (XI, Yl).

Now, the length of vector OI) is r, and the angle between it and the X axis is 0.
Then, the following formula holds.

As can be seen from equation (3), in order to obtain XI and Ys, X and Y must be multiplied by cos α and sin α, respectively.

However, if we limit α to -45'' to +45'', then 1
/cosα is 1~. /T. If it is used for pattern matching, there is no problem in enlarging one image by one time.

1 point P there! If the coordinates of position P2, which is obtained by multiplying 1/cosα by (X Z, Y 2), are as follows.

Next, the following equation holds true for one point P1.

By transforming equation (2), the following equation is obtained.

In this way, to find Xz, set tanα to the Y coordinate and Y
To obtain x, it is possible to perform a single process of multiplying the X coordinate by tanα. Even better, when performing image coordinate transformation, the data of the original image is scanned line by line in the is unchanged, only the X coordinate is 1
Since it will continue to increase, the new coordinate will similarly increase the X coordinate by one.

It is sufficient to change the Y coordinate by tanα.

To find the first coordinate of the next row, take the first coordinate of the previous row as a reference, increase the Y coordinate by 1, and set the X coordinate by -tan.
It is sufficient to change only α.

In this way, to obtain new coordinates, tanα is simply added each time, and no multiplication is necessary.

Also, in reality, coordinates can only take integer values, so it is necessary to multiply by 4 times 5 people.

In this way, the image can be rotated at high speed simply by enlarging the image size up to 5 times, but with the above method, it is not possible to rotate the image by an angle greater than -45' to +45''. However, as a method for rotating an image in 90° increments, there is a known method of exchanging the X coordinate and Yi mark, so this technique can also be applied to rotate the image in 90° increments of 4.
I decided to divide it into two ranges and calculate it using the following method.

When the rotation angle α is 45° to 135°, 1/sin α
is in the range of 1 to 5, so it was decided to multiply the image by 1/sin α. The coordinates (xx, Yz) at that time are as follows.

Every time the Y coordinate is fixed and the X coordinate is increased by 1, the Y coordinate is increased by 1 at the new coordinate, and the X coordinate is tan (90'-
α) increases.

For the coordinates of the beginning of the next row (the Y coordinate of the original coordinates is decreased), the X coordinate is increased by one compared to the coordinates of the beginning of the previous row, and the Y coordinate is decreased by tan (90°-α).

When the rotation angle α is from 135° to 225°, 1/cos
Since α ranges from 1 to 5, the image is 1/cos
Multiply by α. The coordinates (X2°Y2) at that time are as follows.

Furthermore, the rotation angle α is from 225° to 315° (45'
) 1jHIili, 1/5incz ranges from 1 to H, so multiply the image by -17 sin α.

The coordinates (X2. Yz) at that time are as follows.

By employing such a method, it is possible to rotate to any angle of 360 degrees.

FIG. 3 shows how many times the image is multiplied with respect to the rotation angle α.

When actually programming, use ta as a table.
Since we are creating a range from n O° to tan45°, we set the conversion coefficients A, B, C, and D in the following equation as shown in Figure 4, and set the range of α to 4.
It is divided into eight ranges of 5° each.

In short, the above purpose is to rotate the image by converting X2=A-X+B-Y, Y2=C-X+D-Y, where the original pixel coordinates are x, y and the new coordinates are X2°Y2. Set a formula, divide one rotation angle into at least 4 types, set a type of conversion coefficient equal to the number of divisions, and enlarge the image after rotation by 1 to 5 times in length than the image before rotation. By allowing , two of the four conversion coefficients A, B, C, and D are set to ±1, thereby achieving this by providing means for performing rotation processing at high speed.

[Effect]

Looking at FIG. 4, the absolute value of either A, D or B, C is always 1. This means that when scanning a single line in which the Y coordinate is fixed and the X coordinate is increased by 1 in the original image, one of the desired coordinates only needs to be increased by 1, and the other coordinate is increased by a certain amount. It shows us that we need to change.

Also, when you finish scanning one line and find the first coordinate of the next scan, set one coordinate to 1 for the first point of the previous line.
This shows that it is only necessary to change the other coordinate by the same fixed amount as before.

This method has greatly simplified calculations compared to conventional methods.

Note that the above rotations are based on the origin, but
In the case of rotation around a point other than the origin, the above calculation can be performed by first performing coordinate transformation so that the center of rotation is (0,0).

For pattern matching, all you need to do is cut out the pattern and rotate only the cut out part.

Moreover, since the relative position of the pattern does not matter, the starting point of scanning of the original image coincides with the center of rotation.

In addition, since the calculation concerns only the amount of change, the fact that the center of rotation is not the origin does not pose a problem in the actual calculation.

By the way, since the image processing device scans from the top left to the right and down one line at a time, the horizontal axis is the X axis and the vertical axis is the Y axis, but as you go down, the Y coordinate increases. It has the shape shown in Figure 5. Therefore, to view it in the same direction as Figure 2, turn it upside down.

Moreover, you have to look at it from behind.

In that case, the direction of O is clockwise, which is opposite to that in FIG.

With the above method, the area to be expanded is 36 degrees or less by 5 times.
Rotation of ° is possible.

After rotating in this way, cut out the letters, etc., and
By normalizing to a binary pattern of 6 rows and 16 columns, the effect of doubling can be removed.

In addition, in the example of black text, if you enlarge it five times, white dots will appear between the black parts and the text will become a little thinner, so you may want to perform filtering to enlarge or reduce the black dots to improve the image quality of the text. There is also.

〔Example〕

FIG. 1 shows an embodiment of a hardware configuration diagram of an image processing apparatus employing the present invention. This is an example of recognizing characters written on the bottom of a cylindrical object.

A cylindrical object 14 with letters written on the bottom is a light 13
It is illuminated by the camera 11 and photographed by the camera 11.

This image is processed by an image processing device 10 to read the characters written on the bottom of the object 14.

A/D converter 101 converts the analog output signal of camera 11 into a digital signal and sends it to selector 102 . When inputting an image, the selector 102 selects the output of the A/D converter 101, so this signal is sent to the lookup table LUT.
The image is binarized in step 103, passed through a filter circuit 104, and stored in an image memory 105.

This image is output from the output circuit 112 and monitored on the monitor television 12.

If this image contains a noise signal and needs to be filtered by enlarging or reducing black points, the image in the image memory 105 is selected by the selector 109 .

102 and sends it to the filter circuit 104 through the look amplifier table LUT 103. Filter circuit 104
It has 3 lines of image memory, and ANDs 8 surrounding points and 9 points of your own image information.

Expand the sunspot by taking OR and FOR.

Perform processing such as shrinkage.

In this way, the image undergoes various processes during one rotation through the image bus 117, so this loop is referred to as an image loop.

Commands for these processes are inputted to the keyboard via /B 115, and are executed by the control unit CPU 13 according to a program stored in the memory 114.

The image in the image memory 105 is divided into parts of one character based on the projection distribution (the number of white dots in each row in the X-axis direction, the number of white dots in each column in the Y-axis direction) obtained by the X-axis projection circuit 110 and the Y-axis projection circuit 111. is selected, cut out, subjected to one-rotation processing, and stored in the image memory 107. Furthermore, each character is cut out and compared with a standard pattern stored in memory 114 for recognition. These results are shown on monitor TV 12
will be displayed. In addition, if necessary, input/output circuit■101
16, the data is output to other computers, etc.

Address control circuits AC106 and AC108 control the addresses of image memories 105 and 107, respectively.

A case in which this hardware is used to read characters written on the bottom of the object 14 will be described below as an example.

A binarized image 20 of the bottom of the object is shown in FIG. 2
1 is an image of the bottom of the object, and 26 is a written character. In addition, square marks 24 and 25 for detecting the rotation angle are written. The first mark 24 is larger than the last mark 25, and the length of one side thereof is longer than the line width of the character.

22 is a window, which is set so that even if the position of the object varies, the line around the object does not fall within the window 22, and the character string does not protrude outside the window 22. Ru.

The main flow of the subsequent processing is shown in FIG.

First, the rotation angle of the character is determined using 200. To do this, the filter circuit 104 is used to perform black point contraction on the image 20, and the resultant image is stored in the image memory 107, so that the thin characters 26 disappear from the image.

Only marks 24.25 remain.

Next, the image in the window 22 is searched and a rectangle circumscribing the image is found. As a method, window 22
Find the black point in , find the outline of the image, and find the maximum and minimum Y coordinates of the outline, and the leftmost coordinate.

This can be achieved using a method such as finding the rightmost X* marker.

Let the center coordinates of the larger area be (Xa, Ya).

When the smaller center coordinate is (X4. Y4),
The rotation angle θ of the character when viewed clockwise from the coordinates is determined by the following formula.

However, θ is in the range from Oo to 180', and if Y3-Y4 is negative, 180° is added to the result.

In this case, we are finding the angle, but the calculation uses t
Since it is an O, it may be used as is for calculation of rotation. However, considering the possibility of finding angles using other methods, a method for finding angles will be described here.

Once the angle of the character has been determined, the next step is step 3 in the flowchart in Figure 7.
00 rotates by that angle to correct the character position.

FIG. 9 shows a more detailed flow for rotating characters.

In step 301, first, a rotating portion is cut out.

The time required for rotation is proportional to the parts being rotated, so remove unnecessary parts before rotating.

Therefore, the X of the window 22 portion of the image memory 105 is
The axial projection distribution and the Y-axis projection distribution are determined, and a window 23 is set as shown in FIG. 10 in the range where the black dot exists.

In this case, in addition to the character 26, marks 24 and 25 will also be rotated, but such a wasteful amount cannot be helped.

Next, at 302, the coordinates of the upper left point of the window 23 are set to
, Ys, and the widths are Xa and Yd to find the coordinates of the window 23.

In step 303, the image in the image memory 107 used to find the position of the marks 24 and 25 is cleared, and preparations are made to record the rotated image.

From the range of angle α to be rotated at 304 (0 determined at 200 is equal to 1 α), coefficients A, B, C, and D are determined using the table shown in FIG. The following description will be made assuming that the angle α is in the range of 00 to 45 degrees.

In other words, assume that it rotates according to the following formula.

In step 305, the window 23 in the image memory 105 is rotated and recorded in the image memory 107.

For this purpose, the image memory 105 is set to read mode and the image memory 107 is set to write mode, and the data is rotated in an image loop. By this method, the address A of the original image in the image memory 105 is obtained. information is moved to address A of the new image in the image memory 107. Each address is controlled by address control circuits AC106 and AC108.

The number of pixels of the image captured from camera 11 is 256x2
In the case of 56, the address of the image memory takes the origin at the upper left as 0 and increases by 1 for each pixel in the scanning direction. One line is 8 bits, and 9 or more bits represent the number of lines, but the number of lines is also 25.
Since there are 6 lines, there are 8 bits. Therefore, two 8-bit registers can be used to provide an X-axis address and a Y-axis address. From now on, the address Ao is (Xo, Yo
), address A0 is expressed as (xn, y,).

FIG. 11 shows the flow of the rotation processing section.

First, at 400, X o = X s , Y o =
Ys, and.

Set the initial values as Xo”xs and Yo=Ys.

In 401, the address of the original image increases Xo by 1, and Y
o is kept constant, and this is repeated Xd times in 402. The address of the new image changes in synchronization with this, but Xo increases by 1 and Y increases by tan O. However, since the actual address is an integer, the value obtained by rounding off the value of Yn is used as the address. This allows you to rotate one line.

Next, rotate the next row. To do this, it is necessary to remember the start address value of the previous line, so the address of the original image is XO'IYO' + the address of the new image is Xn'r.
yn', the start address of the first line is as follows. Xo' = Xs, Y.

=Ys, Xn'=xs', Yn=Ys In 403, the start address of the second line is Y for the original image.
o is incremented by 1, and Xo is incremented by -tan O. This start address will be used to calculate the start address of the next line, so remember it and set this value to the address A of the original image.
In order to substitute o and the address An of the new image, do the following. X o ” X o 'Yo=Yo', Xn=X1
1', Yn=Yn' Also, the address is the value obtained by rounding off Xo.

Rotation of one line is performed based on this first address.

When the operations from 401 to 403 are repeated Yd times in 404, the rotation of the window 23 is completed.

The above explanation applies when α is Oo ~ 45'', but α is 4
In the case of 5° to 90', rotation is performed using the following formula.

The processing at this time is X in the formula shown in 401. .

Yo remains unchanged, but changes are made such that tan (90'-a) is added to X and Yn is added by 1. Also, 40
3, X6'HYo' remains unchanged, Xn' subtracts 1, and Yn' adds tan(90''-a).

Other ranges of α can also be accommodated by changing the equations 401 and 403 according to the table in FIG.

In this case, when changing the address, incrementing or decrementing the address by 1 uses an up counter or a down counter as shown in Japanese Patent Application No. 58-9874.
By including it in 08, it can be processed in a very short time. Therefore, 401.

In step 403, adding real numbers such as janα is α
The fact that it is possible to have only one image in each area can be said to reflect the effect of allowing the image to be multiplied up to five times.

On the other hand, in the conventional method using equation (3), it is necessary to add real numbers such as sinα and CO8α to both the X and Y coordinates.

Since the rotation of the characters has been completed in this way, pattern matching is then performed in step 500.

To do this, first cut out one character. As for the method of cutting out characters, it is possible to find out from the contours as in the case of cutting out marks. Next, cut out the image by 16 dots l-X 16
The characters are normalized to dots and matched with a standard pattern stored in the memory 114 to recognize the characters.

The normalization process to 16 dots x 16 dots eliminates the problem that characters are enlarged to different magnifications from 1 to 5 times depending on the rotation angle.

〔Effect of the invention〕

According to the present invention, when characters, etc. captured diagonally by a camera are rotated and then returned to the normal direction for recognition, the address control circuit includes an up counter and a down counter, so that the characters, etc. can be rotated at high speed. can be done.

In this case, there is a drawback that the characters are enlarged by a factor of 1 to 2 depending on the rotation angle of the characters, but there is no problem if the characters are normalized to 16 dots x 16 dots or the like.

Therefore, the present invention is highly effective when adopted for recognizing rotated input characters.

[Brief explanation of the drawing]

FIG. 1 is a hardware configuration diagram showing an embodiment of the image processing device of the present invention, FIGS. 2 and 3 are diagrams for explaining the theory of image rotation, and FIG. 4 is a diagram showing the rotation angle and coefficients. 5 is a diagram showing the method of expressing coordinates in an image processing device, FIG. 6 is a diagram showing an image of an object, and FIG. 7 is a flowchart showing an embodiment of the main flow in the present invention. I.
, FIG. 8 is a diagram showing a rotation angle detection method, FIGS. 9 and 11 are flowcharts showing examples of 300 in FIG. 7 and 305 in FIG. 9, respectively, and FIG. FIG. 10... Image processing device, 11... Camera, 12...
・Monitor TV, 13...Lighting, 14...Object,
20... Binarized image, 21... Image of the bottom of the object,
22.23...Window, 24.25...Mark, 26...Character, 101・A/D converter, 102,1
09-f! Rector, 103... Lookup table LUT, 104... Filter circuit, 105, 107...
...Image memory, 106.108...Address control circuit AC, 110...X-axis projection circuit, 111...Y-axis projection circuit, 112...Output circuit, 113...Control circuit C
PU, 114...Memory, 115...Keyboard/
B, 116... Input/output circuit I10°

Claims (3)

[Claims] 1. It is equipped with a plurality of image memories and a plurality of image memory address control circuits that capture an image of an object with a camera, etc., binarize it, and perform image processing. It is also possible to add 1 or subtract 1, and there is a control circuit that controls the entire device,
In a device having a memory for storing a program for the control circuit, in order to rotate the image, let the coordinates of the original pixel be X, Y and the coordinates of the new coordinates be X_2, Y_2, X_2=AX+BY, Y_2= Set the conversion formula of CX+DY, divide the rotation angle into at least 4 types, set the type of conversion coefficient equal to the number of divisions, and make sure that the image after rotation is longer than the image before rotation (1 to √2). By allowing the expansion by a factor of two, the four conversion coefficients A, B
, C, and D are set to ±1 to perform rotation processing at high speed. 2. Cut out characters etc. from the rotated image and print 16 dots
2. The image processing apparatus according to claim 1, wherein a pattern is obtained by normalizing to a predetermined number of pixels, such as 16 dots, and the pattern is recognized by comparing it with a standard pattern stored in a memory. 3. 2. The image processing device according to claim 1, which includes a filter circuit and improves the image quality of the rotated image by filtering such as black point enlargement or reduction.