JPH0534712B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention relates to an image processing device, and particularly to an image processing device that performs smoothing, gradient operation, Laplacian operation, pole extraction, threshold processing, noise removal processing, etc. on grayscale images. The present invention relates to an image processing apparatus that executes a series of processes for extracting the outline of an object as a binary image in real time and at high speed in synchronization with the horizontal scanning signal of a TV camera.

[Technology background]

Recently, industrial robots have been used in various fields. For example, in the case of machine assembly robots, an object is moved for the next process in a short period of time until the robot finishes one operation, such as loosening a screw. If you don't recognize what it is, you won't be able to continue the action. To achieve this, the robot's eyes must process multilevel image signals input from a TV camera at high speed to quickly recognize what the object is. To do this, we extract parts with rapid changes in brightness,
It is necessary to capture the outline of an object.

[Conventional technology and problems]

In general, image data obtained using a TV camera etc. has a screen size of, for example, 256 x 256 (or 128 x
128) Since it is a two-dimensional, 8-bit gray level of pixels, the amount of information is extremely large. Conventionally, when attempting to handle such image data, the image signal from the TV camera is stored in an image memory for one frame, and the data is transferred to a large computer for processing, or Techniques have been used such as determining the slice level of a video signal from a camera using some method, converting it into a binary value, compressing the data, storing it in a memory, and performing logical operations on the data.

Therefore, while the former can perform advanced image processing, it has the disadvantage of requiring large memory and high-speed large-scale computers to achieve high-speed processing, resulting in high costs. It is difficult to determine the appropriate slice level for digitization, and since multilevel information is omitted to binary level information, the amount of information is missing, so even if this is processed, it is difficult to It has the disadvantage that it cannot process, for example, images with complex gradation changes.

Recently, some general-purpose image processing devices have been designed to speed up image processing by arranging one basic calculation module for each pixel or multiple pixels and performing parallel calculations. In order to do this, it is necessary to arrange a large number of basic arithmetic modules, which has the drawbacks that the processing device itself becomes large and costly, and address control becomes complicated.

Therefore, in order to improve such drawbacks, as shown in FIG. 1, an apparatus for extracting a contour line drawing of an object from a grayscale image is used, as shown in FIG. The applicant has developed a system in which the AD converter 2 converts a photographed image into a multilevel digital grayscale image, and then the binary image extraction circuit 3 converts this digital grayscale image into a binary image representing the outline of an object. Gansho 57-
It has been filed as No. 104742.

In this case, as the binary image extraction circuit 3, a spatial filter calculation unit consisting of multiplication, addition, and division circuits is used, and by appropriately selecting the weight coefficient of this spatial filter calculation unit, noise reduction and image enhancement are performed. The object's contour line is extracted as a binary image by extracting horizontal and vertical poles from the data.

This method is very effective when the target is a single object, the background has a uniform gray level, and there is a high contrast between the target and the background.

However, with normal image data acquisition systems, the surface condition, color, shape, lighting conditions, lens brightness, aperture, etc. of the target object can cause uneven shading levels, and it is difficult to obtain high contrast between the target and the background. Nakatari. In such a case, a wedge-shaped pattern appears in the background and the S/N ratio decreases. In order to eliminate this background noise, threshold processing is first required. Furthermore, if the objects overlap and three or more regions intersect, it is not enough to simply extract the horizontal and vertical poles from the data after smoothing and Laplacian processing, as shown in Figure 2. at the intersection
A discontinuous portion occurs as shown as N ₀ (the reason will be explained later). Further, as shown in FIG. 3, whisker-like noise _N1 branching from the contour line, isolated noise _N2 of several pixels or less due to external factors, etc. cannot be removed and remain.

[Purpose of the invention]

In order to improve the above-mentioned problems, it is an object of the present invention to extract not only polar points in the horizontal and vertical directions but also polar points in diagonal directions of 45° and 135° for grayscale image data after applying a spatial filter or FFT. By performing processing, we extract sudden changes in brightness on the original image and eliminate the occurrence of discontinuous parts, eliminate background noise by threshold processing, and further remove whisker-like noise and isolated noise by noise removal processing for binary images. It is an object of the present invention to provide an image processing device capable of extracting only the outline of an object as a series of connections and obtaining an image with a good S/N ratio.

[Structure of the invention]

To achieve this purpose, the image processing device of the present invention includes a data shift circuit that holds and stores a certain length of data, and a data shift circuit that holds and stores a certain length of data, in order to process a grayscale image pixel by pixel and extract its outline. A space equipped with a multiplication circuit that performs multiplication by a preset load coefficient, an addition circuit that adds the results, and a division circuit that keeps the increased level due to multiplication and addition within an appropriate range. In an image processing device that has a filter operation unit and a pole extraction unit that extracts pole points in the horizontal and vertical directions, a pole point extraction unit that extracts pole points in the horizontal and vertical directions and also extracts pole points in the diagonal direction by comparing the gray levels of adjacent pixels. It includes an extraction section, a threshold processing section for removing background noise, and a noise removal section for removing whisker-like noise and isolated noise of several pixels or less, and a first spatial filter operation as the spatial filter operation section. means and a second spatial filter calculating means, the first spatial filter calculating means and the second spatial filter calculating means are provided.
The spatial filter calculation means of the invention are configured so that they can be selectively connected in series or in parallel, and the outline of a grayscale image is extracted as binary pixels in real time in synchronization with the horizontal scanning signal of a TV camera. Features.

[Key points of the invention]

The key point of the present invention is to extract pixels with locally high gray levels from gray scale image data after applying a spatial filter or FFT as candidate points for the contour line of a target object. corners and
Processing of each pole in the 135° diagonal direction was performed to extract an uninterrupted contour even when three or more areas intersect, and threshold processing was performed to remove low-level background noise. Furthermore, even if the number of contour lines extracted as a binary image is small, it is assumed that a closed loop is formed by, for example, 8 connections, and whisker-like noise and isolated noise can be removed by logical operations within the local window. Through this process, only the outline of the object is extracted from the original image.

[Embodiments of the invention]

An embodiment of the present invention will be described with reference to the accompanying drawings starting from FIG.

FIG. 4 shows a simple configuration diagram of the image processing apparatus of the present invention.

Its basic configuration is a first spatial filter calculation section 1
2. Second spatial filter calculation unit 13, pole extraction unit 1
4, a threshold processing unit 15 and a noise removal unit 16, the first spatial filter calculation unit 12 and the second spatial filter calculation unit 13 are connected in series or in parallel by a changeover switch unit 17, and in the case of a parallel connection state, the absolute A value calculation section 18 is inserted.

When the object 4 in FIG. 5 a-1 is viewed from above, the image shown in FIG. 5 a-2 is obtained, and if this is scanned along the line X-Y, the output signal shown in FIG. 5 a-3 is obtained.
If this is subjected to differential processing, the output shown in FIG. 5b-3 can be obtained. From this differential output, slice levels A, B,
If it is set to C, the outputs shown in FIG. 6a, b, and c can be obtained, but it is difficult to select the slice level B in order to obtain the ideal contour shown in FIG. 6b. Therefore, the fifth
Figure 5b obtained by differential processing of the signal in Figure a-3
By extracting the extreme points from -3, it is possible to obtain an output as shown in FIG. 6b without particularly setting the slice level. Note that FIGS. 5 and 6 will be described in detail later.

The first spatial filter calculating section 12 and the second spatial filter calculating section 13 have the same configuration, and the details thereof will be explained using the first spatial filter calculating section 12 as a representative example.

In the spatial filter calculation, the calculation given by the following equation (1) is performed in hardware.

Y _n,o = ₁ 〓 ^i=-1 ₁ 〓 ^j=-1 X _n+i,o+j・f _i,j ／H ...(1) Here, X _n,o and Y _n,o are As shown in FIG. 7, the gray level values of pixels in m rows and n columns of the screen are shown before and after filtering, where f _i,j is a weighting coefficient and H is a normalization coefficient. This filter converts the value of Y _n,o to the pixel of interest X _n,o and its surrounding eight pixels X _n+i,o+j (i=1, 0,
-1, j=1, 0, -1 except i=j=0)
As shown in FIG. 13, the characteristics of the filter can be changed in various ways by selecting the value of f _{i,j .}

Therefore, the load coefficient f _i,j is determined as shown in Figure 7 (b),
When the normalization coefficient H is set as H=16, Y _n,o is given by the following formula
Required by (2).

Y _n,o =1/16 [X _n-1,o-1 ×1+X _n,o-1 ×2+X _n+1,o-1 ×
1+X _n-1,o ×2+X _n,o ×4+X _n+1,o ×2+X _n-1,o+1 ×1
+X _n,o+1 ×2 +X _n+1,o+1 ×1〕...(2) In the state shown in Fig. 4, the first spatial filter calculation unit 12
and the second spatial filter calculation section 13 are connected in series by a changeover switch section 17. In this case,
The first spatial filter calculation unit 12 in the first stage is used as a smoothing filter, and the second spatial filter calculation unit 1 in the second stage is used as a smoothing filter.
3 for an image enhancement filter (high pass, Laplacian) to reduce image noise and extract the outline of the object in real time.

Now, the hardware configuration that realizes the calculation of the above equation (1) will be explained. As shown in FIG. 8, this spatial filter operation section includes a data shift circuit 30, a multiplication circuit 3,
1, an adder circuit 32, and a divider circuit 33.

The data shift circuit 30 records and stores data of 2 rows + 3 pixels using a shift register, as shown in FIG. 9, in order to perform 3×3 spatial filtering processing on image data for one screen. , X _n+i,o+j (i=-1, 0, 1, J=
It has the function of transmitting nine pixels (-1, 0, 1).

In FIG. 9, the case of screen data with 1 row of p pixels and 2 ^q gray levels will be described. The data displayed by the A/D converter 11 in FIG. 8 with q bits per pixel is held using shift registers for 3 pixels in the first stage and p pixels in the second and third stages. And regarding the last three pixels,
The latch or flip-flop circuit R1~
By cascading R9 or using a serial-in-parallel-out shift register, nine pixels can be transmitted. The second and third stages are shift registers 3 for (p-3) pixels.
4 and 35 are used to hold data of pixels other than those sent out. And this data shift circuit is
It is shifted pixel by pixel according to the scanning signal of the camera 10.

As shown in FIG. 10, the multiplication circuit 31 applies a weighting coefficient f _i,j (i=-1,
It has a function to obtain the following equation (3) by multiplying by 0, 1, j=-1, 0, 1).

W _n+i,o+j =X _n+i,o+j ×f _i,j ...(3) (i=-1, 0, 1, j=-1, 0, 1) Therefore, the As shown in FIG. 10, nine multipliers from the first multiplier 36 to the ninth multiplier 44 are used, and the first multiplier 36 is set with load coefficients f _-1 and _-1 , and the data shift circuit Created pixel X _n-1,o-1
Multiplication is performed. The second multiplier 37 is set with a weighting coefficient f _{-1,0 ,} and is multiplied by the pixel X _n-1,o sent out from the data shift circuit. In this way, the multiplication in equation (3) can be performed using the multiplication circuit shown in FIG.

The adder circuit 32 calculates the nine values W _n+i,o+j (i=-1, 0, 1, j=-
1, 0, 1), it has an addition function that calculates the following equation (4), which is the sum of these.

Z _n,o = ₁ 〓 ^i=-1 ₁ 〓 ^j=-1 W _n+i,n+j ...(4) (i=-1, 0, 1, j=-1, 0, 1) Therefore As shown in FIG. 11, the first adder 4
The fifth to eighth adders 52 are used.

Furthermore, the division circuit 33 adjusts the gray level Z _n,o increased by the multiplication and addition operations in the multiplication circuit 31 and the addition circuit 32 to fit within the gray level range of the image memory 19 and the monitor TV shown in FIG. By dividing by the normalization coefficient H, Y _n,o =Z _n,o /H (=Equation (1)) ...(5) is obtained. Here, the normalization coefficient H is ₁ 〓 ^i=-1 ₁ 〓 ^j=-1 f close to _i,j and larger than 2 ^k
(k=0, 1, 2...). As shown in FIG. 12, this division circuit 33 is composed of a first data selector 53 to an eighth data selector 60, and by performing the hydrochloric acid calculation of the above equation (5), the final calculation result of the above equation (1) is obtained. Y _n,o is obtained. This figure 12 shows 8-bit data obtained by multiplying and adding 8-bit gray pixels to 16-bit gray level data _Zn,o.
In this example _, 2 to 9 bits of the 16-bit data Z _n,o are input to the first data selector 53, and the 16-bit data Z n, _{o is input to the second data selector 54.} 3 to 10 bits of _o are input, 4 to 11 bits are similarly input to the third data selector 55, and 5 to 12 bits are input to the fourth data selector 56.
6 to 13 bits to the fifth data selector 57, 7 to 14 bits to the sixth data selector 58, 8 to 15 bits to the seventh data selector 59, and 9 to 13 bits to the eighth data selector 60. Input 16 bits each. Each data selector 53-60 selects and outputs any one of the first to eighth bits among the 8-bit inputs in response to a 3-bit selection signal transmitted from the selector line SL. For example, when quoting Z _n,o by 2 ⁷ , each data selector 53 is set so that the first bit is output from each data selector 53 to 60.
~60 is controlled, so the first data selector 5
3 outputs the first bit of 2 to 9 bits, that is, the 2nd bit of 16 bits, and the second data selector 54 outputs the first bit of 3 to 10 bits, that is, 16 bits. The third data selector 55 outputs the first bit of the 4th to 11th bits, that is, the 4th bit of the 16 bits. 60 outputs the first bit of 9 to 16 bits, that is, the 9th bit of 16 bits, and as a result, each data selector 53 to 60 outputs 16 bits Z _n,o by 2. It will output 2 to 9 bits, which is the value divided by ⁷ . Similarly, Z _n,o is 2 ⁶
When quotient by 25, it is sufficient to output the second bit from each data selector 53 to 60, that is, 3 to 10 bits of 16 bit data; when quotient by ²⁵ , each data selector 53 to 60 It is only necessary to output the third bit and output the 4th to 11th bits from the 16-bit data. In this way, the ^215-N to ^{28-N bits of the 16-bit data are input to the Nth data} selector, and the value of k is set in the selector line corresponding to the dividing number ^2k . Division can be realized at high speed. In this way, by using the data selector, it is possible to speed up division.

By appropriately selecting the weighting systems of the first spatial filter calculation unit 12 and the second spatial filter calculation unit 13 in FIG. 4, random noise on the image can be reduced and brightness can be enhanced to emphasize sudden changes. This load coefficient can be retrieved by the CPU 21 from the load coefficient table of the load coefficient holding section 22, for example.

For example, FIG. 5b shows the state after smoothing and Laplacian processing are applied to the original image in FIG. 5a. Fifth
Figure a-1 shows the case where the light is hitting the cylindrical object 4 from diagonally above the right, Figure a-2 shows the view from directly above, and Figure a-3 shows the case where the light hits the cylindrical object 4 from the upper right side. X-Y1
It shows the change in brightness in the dimensional direction. Further, Fig. 5 b-1 is a three-dimensional display of the gray level after smoothing + Laplacian calculation, and Fig. 5 b-2 shows isobright lines when viewed from directly above. Furthermore, FIG. 5 b-3 shows a cross section of FIG. 5 b-2 in the X-Y one-dimensional direction.

Let us now consider extracting the outline of the object from FIG. 5 b-1 by threshold processing. When levels A, B, and C3 are selected as slice levels and binarized, different contour lines are obtained as shown in FIGS. 6a to 6c. In Figure 6a, the level is too high and the outline is interrupted, and in Figure 6c, the level is too low and the outline becomes thick. Therefore, it is necessary to repair the discontinuity or thin the contour line again. These processes are not suitable for real-time processing because their algorithms are complex and time-consuming. It would be convenient if the slice level could be set just right as shown in Figure 6b, but the shading level of the original image depends on the surface condition of the target object.
It is difficult to find an appropriate slicing level because the color, shape, lighting conditions, lens brightness, aperture, etc. vary each time. After all, it is difficult to extract an object as a series of line drawings and obtain an image with a good S/N ratio only by threshold processing.

Therefore, in the present invention, contour candidate points are first extracted by extreme point extraction processing. The principle of this location extraction operation will be explained with reference to FIG. To simplify the explanation, an example will be described in which the extreme point R is determined when the original image P shown in FIG. 14A has the density distribution in the one-dimensional direction as shown. Let P _L be a left-shifted image obtained by shifting the original image P by one pixel Δ to the left, and P _R be a right-shifted image obtained by shifting the original image P by one pixel Δ to the right.
Then, compare the original image P and the left-shifted image P _L and the density of the original image is greater or equal (hereinafter referred to as P
≧P _L ) Extract the part Q _L , and compare the original image P and the right-shifted image P _R to find the part Q where the density of the original image is greater or equal (hereinafter this will be referred to as P≧P _R ). Extract _R. The extreme point (maximum point) found at this time
Since R is considered to be at least at a level greater than or equal to the pixels on both sides, the extreme point R can be found by extracting the part Q _L and Q _R as shown in FIG. 14E. However, P=
Exclude when P _L = P _L.

By the way, when there is a pole in a pattern as shown in L in Fig. 14, the pole sides l ₁ and I ₃ of the pattern L can be extracted by scanning only in the X direction, but the pole sides I ₂ and I ₄ cannot be extracted. Since this is not possible, the pole sides I ₂ and I ₄ can be extracted by scanning this in the Y direction as well. Therefore, for example, the 14th
If we use the extraction window W' having pixels A to I as shown in the figure, we can use D, E, and F to determine the extreme point in the X direction of the two-dimensional gray image, and B, E, and H to determine the extreme point in the Y direction. C, E, and G can be used to extract a 45° diagonal pole, and A, E, and I can be used to extract a 135° diagonal pole.

Now, one frame shown in FIG. 14H is sequentially scanned by scanning h ₀ , h ₁ , h ₂ . By performing comparison processing, it is possible to sequentially extract pole points in the horizontal, vertical, 45°, and 135° directions.

The pole extractor 14 shown in FIG. 4 is composed of a data shift circuit 62 and a data comparison circuit 63, as shown in FIG.

As detailed in FIG. 16, the data shift circuit 62 sequentially holds data of 2 rows + 3 pixels using a shift register, and the data comparison circuit 63 of the next stage
It has a function of transmitting nine pixels A to I relative to each other, and is configured in the same manner as the data shift circuit shown in FIG. The output of the spatial filter 61 made up of the first spatial filter calculation section 12 and the second spatial filter calculation section 13 is transferred to a data shift circuit 62 comprising registers r ₁ to r ₉ and N-3 bit length shift registers 64 and 65. As a result, pixels A to I are output from each register r ₁ to _{r 9} and input to the data comparison circuit 63 .

As shown in FIG. 17, this data comparison circuit 63 includes comparators C-1 to C-8, OR circuits OR-1 to
OR-9, AND circuit AN-1 to AN-4, NAND circuit NAN-1 to NAN-4, peak register
It is composed of RP, etc. and comparator C-
1 is the register that is the spatially filtered data.
Pixels D and E of r ₄ and r ₅ are compared, and pixels F and E are also compared by comparator C-2. Therefore, if pixel E is the pole in the horizontal direction (X direction), E
>D, E>F, D≠E≠F, so OR-2,
Since "1" is output from both OR-3 and NMN-1, and "1" is output from OR-1,
Pixel E is held in the peak register PR as a horizontal peak value, and this is output as an extreme point.

Also, the comparator C-3 compares pixels B and E,
Since pixels E and H are compared in comparator C-4,
When pixel E is a vertical pole, AN-2 outputs "1" and OR-1 outputs "1", so pixel E is held in the peak register PR. Similarly, pixels C, E, and G in a 45° diagonal direction are compared by comparators C-5 and C-6, and if pixel E is the pole, "1" is output from AN-3, and comparator C- 7. C-8 compares pixels A, E, and I in the 135° diagonal direction, and if pixel E is the pole in the 135° diagonal direction, "1" is output from AN-4. Become. Therefore, if pixel E is at the pole in any one of the horizontal, vertical, 45°, and 135° diagonal directions, OR-1 outputs "1" and pixel E is held in the peak register PR. and will be output as an extreme point. In this way, if the pixel of interest E is an extreme point, it can be extracted.

As shown in FIG. 18, the threshold processing section 15 consists of a data shift circuit and a threshold circuit, and the data shift circuit 62 uses the same circuit as that of the pole extraction processing section. The latter has a comparator 66, an AND circuit 67, and a register 68, and the comparator 66 compares a preset slice level with the pixel of interest (for example, E), and when the pixel of interest has a density greater than the slice level, Comparator 66 outputs "1" and turns on AND circuit 67. By controlling the level of the pixel of interest E using this signal, it is possible to extract only those pixels that are at the extremes in each of the horizontal, vertical, 45°, and 135° directions at a certain level or higher.

In this way, in the present invention, the extreme point extraction section 14 extracts contour candidate points, and the threshold processing section 15 performs threshold processing.

In other words, the contour of the object to be contour extracted is considered to correspond to a portion where the brightness changes suddenly, in principle. Therefore, the Laplacian filter (2
In the data after applying a gradient filter (first-order differentiation, boundary line extraction), the portion corresponding to the contour of the object has a higher level than other portions. Therefore, polar point processing is performed on the filtered data in horizontal, vertical, 45° diagonal, and 135° diagonal directions to extract contour candidate points whose brightness changes suddenly on the original image. If this is all that is done, a lot of background noise due to external factors will be extracted.
However, since the level of background noise is usually lower than the contour points of interest, it can be removed by thresholding. Since the purpose of the threshold processing in the present invention is not to directly extract contour points from the data after filtering, the slice level is such that the sequence of contour candidate points extracted by the extreme point extraction process is not interrupted. , it is sufficient that background noise can be removed.

Since this process can be performed in parallel with the extreme point extraction process, the speed can be increased by arranging the extreme point extracting section 14 and the threshold value processing section 15 in parallel. As a result, at the horizontal, vertical 45° diagonal and 135° diagonal poles,
In addition, points above a certain slice level can be extracted as contour candidate points as a binary image.

By the way, in the threshold processing in the threshold processing unit 15, the slice level was set so that the contour candidate point sequence was not interrupted, so some whisker-like noise and isolated noise of several pixels or less remained as shown in FIG. . Therefore, assuming that the contour candidate point sequence extracted as binary pixels constitutes a closed loop with at least 8 connections, whisker-like noise and isolated Perform noise removal. Removal of each of these noises is performed by the noise removal section 16.

Before explaining the detailed structure of the noise removal section 16, its operating principle will be explained.

As shown in FIGS. 19 to 19, if the pixel of interest E at the center of the 3×3 window W is not noise, it is always a closed rape image. Since the closed-loop image is continuous, the closed-loop image necessarily crosses the window W, and it can be said that at least two of the surrounding pixels within the window W must be pixels on the closed-loop image. Considering the above preconditions, it can be said that the two peripheral pixels on the closed loop image must be in a non-adjacent positional relationship. From these facts, in order for the pixel of interest E to be a pixel on a closed-loop image, when the pixel of interest E is in state 1, at least two neighboring pixels that are not adjacent to each other must be in state 1. There are 16 cases that meet this condition, and each case is represented by a window W in FIG. 19. In this case, in order to simplify the explanation, it is assumed that if the closed-loop image has a corner, the pixel corresponding to the corner is treated as not a pixel on the closed-loop image.

Thereafter, a pattern in which the basic pattern does not occur within the window W shown in FIG. 19 is examined and set as a noise removal condition for removing noise of several pixels or whisker-like noise. In this case, the first noise removal condition is that when the pixel E of interest is in state 1 in the window W of FIG.
The surrounding pixels along the two adjacent sides of the window W are in the state 0, that is, the surrounding pixels A, B, C, D, and G along the upper side _I1 and the left side _I2 of the window W are in the state 0, and the left side of the window W is in the state 0. Peripheral pixels A along I ₂ and the lower side I ₃ ,
D, G, H, I are in state 0, bottom edge of window W I ₃
The surrounding pixels G, H, I, F, and C along the right side _I4 are in the state 0, and the surrounding pixels I, F, C, B, and A along the right side I4 and the top side _{I1 of} the window W are in the state 0. One example is that. The second removal condition is as shown in FIG. Yes, and the surrounding pixels located at the center of the opposite side of the window W are in state 1;
For example, peripheral pixels A, B, and C along the upper side _I1 of the window W and peripheral pixels D and F adjacent thereto are in state 0, and peripheral pixel H located at the center of the lower side _I3 of the window W is in state 1. There are certain things that can be mentioned. Note that the pixels marked with an x in FIG. 20 indicate that they may be in either state 1 or state 0.

Therefore, if the pattern within the window W in which binary pixels are moved matches either of the first and second noise removal conditions described above, the pixel E of interest is in state 1.
Even so, the pixel E of interest is not a pixel of a closed-loop image, but is detected as noise, and by setting this to "0", this noise can be effectively removed. For example, as shown in Figure 3, noise is caused by clusters of 4 or fewer pixels.
If N ₂ is present, the whisker-like noise that is removed by the first denoising condition and extends from the closed-loop image
If N ₁ exists, it will be removed under the second noise removal condition.

An example of a circuit for implementing the above-described noise removal method is shown below.

As shown in FIG. 21, this noise removal section 16
It is comprised of a data shift circuit 70 to which binary pixels from the spatial filter/threshold processing section 69 are input, and a logic operation circuit 71.

First spatial filter calculation unit 12, second spatial filter calculation unit 13, pole extraction unit 14, threshold processing unit 15
The output of the spatial filter/ _threshold processing unit ₆₉ consisting of
The data is input to a data shift circuit 70 composed of 3-bit shift registers 72 and 73, and thereby pixels A to I are output to each register R ₁ ' to R ₉ ', and input to a logic operation circuit 71.

Here, the logic operation circuit 71 is configured as shown in FIG. 23, and sets the level of the target pixel E to the state 0 when the first noise removal condition or the second noise removal condition shown in FIGS. 20b and 20c is satisfied. At other times, the level of the pixel E of interest is output as it is in state 1, and the logical operation circuit 71 sequentially obtains binary pixel data from which noise has been removed. Specifically, the logical operation circuit 71 is configured to satisfy the following logical expressions (a) to (d).

E^ ₁ = {(A∪B∪C∪D∪G)∩(A∪D∪G∪H∪
I)∩(G∪H∪I∪F∪C)∩(A∪B∪C∪F
∪I)}∩E = {(A∪B∪C∪A∪D∪G)∩(A∪D∪G∪G
∪H∪I)∩(G∪H∪I∪C∪F∪I) ∩(A∪B∪C∪C∪F∪I)}∩E = {(X ₁ ∪X ₂ )∩(X ₂ ∪ X ₃ )∩(X ₃ ∪X ₄ )∩(X ₄ ∪X ₁
)}∩E……(a) E^ ₂ = [{(A∪B∪C∪D∪F)∪H−}∩{(A∪
D∪G∪B∪H)∩F−}∩{(D∪F∪G∪H∪I)
∩
B-} ∩{(C∪F∪I∪B∪H)∪D-}]∩E = [{(X ₁ ∪D∪F)∪H-}∩{(X ₂ ∪B∪H)∪
F−}∩{(X ₃ ∪D∪F)∪B−}∩{(X ₄ ∪B∪H)
∪
D−}〕∩E ……(b) E^=E^ ₁ ∩E^ ₂ ……(c) =X ₁ A∪B∪C, X ₂ = A∪D∪G, X ₃ = G∪H ∪I,X
₄ =C∪F∪I...(d) In equations (a) to (d), A to I represent data of pixels A to I, and B-, D-, F-, and H- represent pixels B and D. ，
Indicates the inverted data of F and H, E∧ ₁ and E∧ ₂ are index values indicating whether the first noise removal condition and the second noise removal condition are satisfied, respectively, and E∧ is whether the attention level E is noise. This is an index value indicating whether or not.

A specific configuration of the logical operation circuit 71 that satisfies this logical formula is shown in FIG. In Figure 23,
OR-1' to OR-14' are OR circuits, AN-1'
AN-3' to AN-3' are AND circuits, NT-1' to NT
-4' is a knot circuit.

For example, if the pixel of interest E and other pixels are
In the case of the condition shown in , "0" is output from OR-1' and OR-2', so OR-7'
also outputs “0”, resulting in AN−1′, AN−
3' outputs "0", therefore the pixel of interest E is "0"
is output as follows.

In addition, in the case of the conditions shown in Figure 20C, OR
-1' and OR-5' each output "0", and the NOT circuit NT-1' also outputs "0", so the OR circuit OR-11' outputs "0". As a result
Since AN-2′ and AN-3′ output “0”,
The pixel of interest E will be output as "0". Such control, like the other circuits above,
This is performed sequentially in synchronization with the horizontal synchronizing signal output from the scanning signal generator 20 shown in FIG.

In this way, the logic operation circuit 71 performs the second
The conditions shown in Figure 0 b and c, that is, when there are no pixels on the top and right sides of the window W, on the right and bottom sides, on the bottom and left sides, or on the left and top sides (noise removal condition 1); , there are no pixels on the left side and above and below and it is on the right, there are no pixels on the bottom and left and right and there are pixels on the top, or there are no pixels on the right side and above and below and it is on the left (noise removal condition 2). This can be considered as noise and removed. In this way, the contour image of the object from which noise has been removed can be stored in the image memory 19.

The changeover switch section 17 connects the first spatial filter calculation section 12 and the second spatial filter calculation section 13 in series or in parallel.For example, the operator may actively switch this from the panel, or the switch may be switched automatically. It may also be controlled. The state in FIG. 4 shows a case of series connection.

If the first spatial filter calculation unit 12 and the second spatial filter calculation unit 13 are connected in series, for example, the load coefficients shown in FIG. Among them, for example, the first spatial filter calculation unit 12
The loading coefficient for the smoothing filter is sent to the second spatial filter calculation unit 13, and the loading coefficient for the second-order differential filter is sent to the second spatial filter calculation unit 13 to perform Laplacian calculation H=∂ ² h(x,y)/Δx ² +∂ ² h(x,y)/
By performing Δy ² , it is possible to perform processing that makes the outline of the object appear raised.

In addition, the first spatial filter calculating section 12 and the second spatial filter calculating section 13 are connected in parallel to the absolute value calculating section 18, and, for example, the loading coefficient (B-1) in FIG. 13 is applied to the first spatial filter calculating section 12. , by sending the weighting coefficient (B-2) to the second spatial filter calculation unit 13, gradient calculation H=|∂h(x,y)/∂x|+|∂h(x,y)
/∂y|, which is effective when you want to extract a directional straight line. For example, you can easily extract diagonal lines.

In addition, according to the present invention, the interrupted portion in FIG.
The reason why the generation of N ₀ can be suppressed is as follows.

That is, as shown in FIG. 2, when three or more areas intersect at one point, a level higher than the intersecting pixel level may appear in surrounding pixels. For example, in FIG. 24, pixels D or H may be at a higher level than E. In this case, pixel E should have a higher level than pixels A and I (that is, pixels A and I are internal pixels), and such a case can be covered by performing polar point extraction in the diagonal direction.

For example, when there is original image data as shown in FIG.
Obtain data as shown in the figure. This is compared to the conventional horizontal
If the poles are extracted only in the vertical direction, the circled portions will be extracted. At this time, although 51 in the window W in FIG. 26 is originally a pole, it cannot be extracted as a pole when compared in the horizontal direction, but if the pole is extracted diagonally as in the present invention, (8), ( 51), (-58)
This can be extracted as an extreme point by comparing
Even where the three regions intersect, the extreme points can be extracted correctly without creating any discontinuities.

In the image processing device of the present invention, by synchronizing the data shift of the data shift circuit with the horizontal scanning signal of the TV camera, grayscale image processing for one frame can be completed in, for example, 16.7 ms. In addition, by providing an address generation section after the noise removal section, the pixel can be set at a certain slice level or higher at the horizontal, vertical, 45° diagonal, and 135° diagonal directions and satisfy noise removal conditions 1 and 2. If the address information is transferred to the planning machine only when the information is not satisfied, only the necessary contour portion can be stored in the memory, and the memory can be dramatically saved.

In the above description, a case has been described in which a 3×3 spatial filter is used, but by increasing the size of the filter to 5×5, 7×7, etc., detailed processing can be performed.

In addition, by making this image processing device a one-chip VLSI, it will be possible to recognize two-dimensional shapes using a microcomputer-level computer in fields that require visual recognition, such as object classification, sorting, and product inspection. Become.

〔Effect of the invention〕

According to the present invention, the contour of an object can not only be extracted in real time and at high speed in synchronization with the horizontal scanning signal of a TV camera, but also can extract polar points in diagonal directions, resulting in an uninterrupted contour. It becomes possible to extract lines. Furthermore, it is possible to remove whisker-like noise, isolated noise, and the like. In addition, the two spatial filter calculation units can be selectively connected in series or in parallel, so if they are connected in series, it is possible to perform processing that makes the outline of the object appear to be highlighted, and if they are connected in parallel, it is possible to perform processing that makes the outline of the object appear raised. It becomes possible to extract a certain straight line, for example, a diagonal line, and perform processing appropriate for the purpose of use.

[Brief explanation of the drawing]

Figure 1 is a schematic diagram of a conventional image processing device; Figure 2 is a schematic diagram of a conventional image processing device;
Fig. 3 is a diagram explaining the problem, Fig. 4 is a configuration diagram of an embodiment of the present invention, Fig. 5 is an illustration of an object and its contour extraction state, and Fig. 6 is an output when changing the slice level. An explanatory diagram, Fig. 7 is an explanatory diagram of the principle of the spatial filter, and Fig. 8 is an explanatory diagram of the spatial filter calculation section.
Fig. 9 is an explanatory diagram of the data shift circuit, Fig. 10 is an explanatory diagram of the multiplication circuit, Fig. 11 is an explanatory diagram of the addition circuit, and
Figure 2 is a diagram explaining the division circuit, Figure 13 is a diagram explaining the use of the spatial filter and its load coefficient, Figure 14 is a diagram explaining the basic operation of the pole extractor, Figure 15 is a configuration diagram of the pole extractor, and Figure 16 is a diagram explaining the basic operation of the pole extractor. The figure shows the data shift circuit, Fig. 17 shows the data comparison circuit, Fig. 18 shows the threshold processing section, and the first
9 is an explanatory diagram of contour data, FIG. 20 is an explanatory diagram of noise removal conditions, FIG. 21 is a noise removal section, FIG. 22 is a data shift circuit, FIG. 23 is a logic operation circuit, and FIGS. 24 and 25. , FIG. 26 is an explanatory diagram of the interrupted portion. In the figure, 10 is a TV camera, 11 is an A/D converter, 12 is a first spatial filter calculation section, 13 is a second spatial filter calculation section, 14 is a pole extraction section, 15
16 is a threshold processing section, 16 is a noise removal section, 17 is a changeover switch section, 18 is an absolute value calculation section, 19 is an image memory, 20 is a scanning signal generation section, 21 is a CPU, 22
3 is a load coefficient holding unit, 30 is a data shift circuit, 3
1 is a multiplication circuit, 32 is an addition circuit, and 33 is a division circuit.

Claims (1)

[Claims] 1. In order to process a grayscale image pixel by pixel and extract its outline, a data shift circuit that holds and stores a certain length of data, and a weight coefficient that is set in advance with the held data is provided. A spatial filter calculation unit includes a multiplication circuit that performs multiplication of In an image processing device that has a polar point extraction section that extracts polar points in a direction, the polar point extraction section extracts horizontal and vertical polar points as well as diagonal polar points by comparing the gray levels of adjacent pixels, and a polar point extraction section that removes background noise. Equipped with a threshold processing unit to remove whisker-like noise and isolated noise of several pixels or less,
Further, a first spatial filter calculating means and a second spatial filter calculating means are provided as the spatial filter calculating section, and the first spatial filter calculating means and the second spatial filter calculating means are selected to be in a series state or a parallel state. The present invention is characterized in that it is configured so that it can be connected to a TV camera, and extracts the outline of a grayscale image as a binary image in real time in synchronization with the horizontal scanning signal of a TV camera.