JPS5927206A - Defective pattern detecting method - Google Patents

Abstract

PURPOSE:To detect the defects in a correct form and to detect only the true defect, by performing shade conversion and distance conversion in a light and shade image, from which a suspected defect pattern of the surface of an object material is extracted. CONSTITUTION:A suspected defect pattern of a surface 101 of an object material to be checked is detected by a TV camera 102 and the image signal is stored in an image memory 104 through an AD converter 103. The output of the image memory 104 and the picture element values from an image memory 105, wherein a reference pattern is stored in advance, are inputted to a difference detecting circuit 106. The differences between the absolute values of both picture element values are sequentially outputted. The picture element signals, whose shade values are fij are sequentially inputted to a shade value converting circuit 401, a judging circuit 402, an addition circit 403, and a shade value circuit 404. A shade level corresponding to a noise level is removed in the shade image, which has three or more kinds of the shade values, through the circuits 401 and 404 without changing the configuration of the defective pattern. The defects, which are not fatal, are removed through the circuits 402, 405, and 403. Thus only the true defects are detected in the correct form.

Description

DETAILED DESCRIPTION OF THE INVENTION Object of the Invention The present invention relates to a method for detecting defect patterns, and in particular detects fatal defects with high density and large shape from a video signal of the surface of an object inputted by a television camera or the like. jE U
Let's move on to the defect pattern detection method 7μ for weighing M.

Prior art print > k KN or I C, L S l-1-
1r, Kll et al.'s mass inspection furnace i (formerly: 1'B-1 7t)] was carried out, but the
Publication Inclination 11 Suzu '41' Time ``Hachiotsu, in order to reduce the Me cost, etc., there was a request for Yuhaku to be made into a private company.

Conventionally, a method for inspecting defects on the surface of an object, such as an LSI, is to use an image of the object's surface input using a pre-vision camera, etc. ■Start 2, when the target lnf surface is burnt, the defect candidate pattern extraction image (871, 1,000σ) is detected (fatal (H'1j7 defect))
j possible image), and object #i+'+'l
If there is a bump in the blue, the difference image between the pattern design data, standard pattern data, and adjacent bump pattern IW11 is displayed as the defect f: sleeve pattern IW11 image, respectively. From these images, the defective lock image, defective position 1 f,
1, which calculates defect feature values such as defect area,
In this method (:L1 following two drawbacks 1)・M・Ru7,
(1) When calculating the feature quantities of a defect, it is common practice to binarize the grayscale image of the defect candidate pattern using t, l, μ, and an appropriate tc 61 or 11Δ.
When P. Shaojong of f1α was disgusted and refused to buy candy that was too expensive,
If the correct shape of the defect is not correct, the incorrect shape will be ejected and set to a value too low for 4', 1 = 2 pairs! If you are smart, you will be able to extract defects such as noise and undulations of the target harmony, as well as f late IJ and distortion.

52) Defect (1) It would be fatal (!'
Defects of a size that cannot be J (r7 tfit pattern) are also extracted at the same time.

When selecting a defect candidate pattern, a certain size L
If you exclude those below J to the midline, you will be eliminated by the fatal f, r defective p god pa-pat players.

In other words, linear fire caused by scratches, cracks, etc.
:fii contains fatal defect candidate patterns because the τj method is large even if the area is small.

Therefore, the area is used for single #Ii as a criterion for size. 2. and 4. j i+;:ft is not correct.

Object of the Invention Summary of the present invention
! To provide a defect pattern detection method capable of selecting a defect with a correct shape by removing surface undulations and the like that cannot be regarded as defects due to the work p1 having a low lightness value equivalent to a noise level. The purpose of the lead 2 of the present invention is to extract defect candidate patterns on the surface of the object to be inspected. It is an object of the present invention to provide a defect pattern detection method that can determine fatal 7j defects even if they are small in area in both images and select only correct defects.

The pattern detection method is as follows: a) In the defect pattern detection method in which a defect candidate pattern is extracted and a defect is selected from a light image containing at least three types of gray values,
? A defect pattern having an internal value of lj:% fj>value greater than or equal to the second FjA value lower than the line 1 in a six-complementary pattern is characterized by outputting only the portion having a value of fj> as a defective pattern. have Further, in the defect pattern detection method of the present invention, (2> After performing the process (1) above to generate a 2fiα image, all pixels of the defect candidate pattern are shielded from the target screen. A so-called distance conversion process is performed in which a value obtained by adding "l IT to the maximum gray value of multiple pixels is written to the target rrr!" in a fixed running direction, and the obtained distance conversion image is pre-written. The feature is that binarization is performed using a set third threshold value.

Embodiment 1 FIG. 1 is a block diagram showing the pretreatment process leading up to the present invention.

In FIG. 1, 1ol is the surface of the object to be inspected which has a certain pattern, defect, etc. on the surface, 102 is a pre-vision
A detector for an optical system such as a camera, 103 is an A/D converter, 1
04 is image memory, 105 is also image memory, 106
is a difference detection circuit, and fy indicates all pixel values.

After magnifying an image within a set field of view with a television camera 102 or the like on the surface 101 of the object to be inspected such as a printed circuit board or LJ, IC, LSI, and their masks,
It converts into a lightning and air signal and outputs it to the next A/D converter 103. The A/D converter 103 zams and holds the input video signal at fixed time intervals using a clock (LK), converts it into digital data, and stores it in the image memory 104. The center is composed of a shift register, etc. that stores a two-dimensional iit+i (Q+ image).On the other hand, a shift register that stores a pattern (reference pattern) that is the same as the pattern on the surface 101 of the object to be inspected as a two-dimensional image image.・Register empty・Fm to the image memory 105 consisting of
.

All the pixel values of the pair b3 of the image memories 104 and 105 are sequentially outputted to the difference detection circuit 106, and the difference detection circuit 106 calculates the difference between the input corresponding image values in the form of absolute values. The results are output in the 11th monthly serial order in the same %i3 order as the input order of the memories 104 and 105.

FIG. 2 is an overall block diagram of a video signal processing block for defect pattern detection showing an embodiment of the present invention.

In FIG. 2, 40] is a circuit that inputs the defect/complement image data and performs a logical operation to convert the grayscale value, and 402 is 2
A size determination circuit 403 for a valued image
Addition circuit 4 ('l) performs the same logical operation as 4-01 on the 5-valued image obtained by addition, A circuit for converting values; 405 is a shift register;
It is a QIIF circuit.

The video signal processing device shown in FIG. 2 includes a processing circuit according to the first invention and the invention of Rod 2, and includes a grayscale value conversion circuit/L01.
and AQ4 are circuits using the defect pattern detection method invented by Kan 1. Further, the processing process from the determination circuit/1.02 + delay circuit/LO5 to the ticket addition circuit 403 is a part that performs defect pattern detection processing including criticality determination according to the invention of Fue 2.

The :/a light value conversion circuits 401 and 404 of the 1st I in 1 calculate the shape of a defective pattern in a grayscale image that has at least three types of grayscale values from which defective candidate patterns are extracted based on differences in grayscale levels. To remove the low gray level flI region corresponding to the noise level without deforming the image. 1. In other words, this method has a threshold value that is as low as 1J, which is considered to be a defect in advance (although the density level is low, it reflects the correct shape of the defect), and a concentration level of 1-
We set two relatively high-level thresholds that can remove most of the noise, but only extract part of the correct shape.

Using these two thresholds, ? above the lower threshold? ! '%
Among the defect candidates having 15 values, only defects having a gray value H equal to or higher than a high threshold value are extracted.

This eliminates noise without changing the shape of the defect.
Defect candidate patterns corresponding to the level can be removed on the image.

Next, the second invention, the fatal defect candidate fI!
Describe the method for removing the J region.

First, a grayscale value-converted image is generated using the 1i11i image converted to 2 fla using the aforementioned low threshold value. An image generated by distance conversion according to a fixed scanning direction, or an image generated by performing two types of distance conversion depending on the scanning direction, and a composite of the two conversion results, are used for the grayscale Wi conversion image. Next, the distance-converted image is binarized using a threshold value that determines the size.

In this way, it is possible to extract a defect ft-complementary pattern ζJ with a correct shape (the size of the shape is larger than a certain value, but the area is not).

FIG. 5 is a procedure explanatory diagram of a video signal processing method showing an example of the present invention.

The 31st m (a) is a grayscale image in which at least three types of grayscale values are extracted from which defect candidate patterns are extracted based on the difference in grayscale of one novel. Here, the white background A is the area with the highest gradation value, the black point B is the area with the next highest gradation value, and the diagonal line C is the area with the highest gradation value.
is the 9A region with the lowest gradation value.

In this grayscale image, two threshold values t□ and t that satisfy the European style are set.

Δ≧t,)B≧t,〉C-Tj(,1) Fig. 5 (PL)
First, for the grayscale image of , the threshold value t.

More than 71! Out of the defect candidate pattern areas with light and dark areas, only areas containing shading values greater than or equal to barplant t□ are extracted to generate 2-1 to 1-Φ (see Figure 3 (1)))
. That is, in the Pb0 diagram <a>, the upper right and lower right areas are t.

1-σ) Shading value [3 and tl or more shading jjfi
A, whereas the lower left region contains t, pJ, 17
．． 4・, h7;? Value in ζ1? contains grayscale values greater than or equal to tl? j), extracted 2 and binarized image (' right 1- and lower right area's fiber qj /7 + t
, the processing of
It will be held on 01.

Next, in the binarized image of the 31st kalpa (b), <r=I L −
A binarized image is generated by extracting only the area whose size is larger than a predetermined reference value (Fig. 5(c)).
reference),.

That is, in the determination circuit 4-02 of FIG. Extract only the missing +<r iri''/supplementary Bakun that has a fatal shape/dimension.

Defects with extremely large areas are of course dangerous, but defects that are longer in size than in area are more dangerous! Hq (a
ii. Distance transformation according to constant scanning direction? ru.

According to this method, the defect candidate region I'i/n, i';
4: L Card Game 4 The degree of fatality is expressed by the sum of each length. In this way, the determination circuit 402 determines the area (1') and the size (size). The distance conversion will be explained in detail in Fig. 5. This is performed by writing a value obtained by adding 1 to the maximum value of a plurality of pixels adjacent to each other into the target pixel.

Next, with the binarized image of the whistle, ``51\il in'' (, J
1. The shape of the lower right region is different from that in FIG. 3(b), and is not the correct defect shape. Therefore, the binarized image of pjf in Figure 3 (b) showing the correct defect shape and Figure 5 (c)
By adding the binarized image of M31M(d)
The ternary image shown in is generated.

In this case, since the processing (() of the binarized image in Section 31(b) has already been completed, as shown in FIG. How ←
-J''< ly, it is necessary to input the image Φ after the distance conversion from the size 11 circuit 402 and 11'f in correspondence and to perform the processing.

In the absence of f1 ~ 21';:l's strong chivalry value split
-04, 2! V3 diagram (binarization of bj jl+i
Perform the same processing as when generating the image.

In other words, the rut is also the high part A' of the shade 11, then the high part A' of the shade 1i.
ii (high part B of 11'4, (j) Uνdfi/It2 of L dark f
Threshold G/j among the defect candidate Pakun Ryohaku with k value of 41
b, C1f-shade 11! Generates a binarized image in which only 11 regions are extracted (including 61 in Figure 3(d),
Since the upper right region does not contain A', it is removed, leaving only the lower right region, as shown in Figure 3(e).

By performing the uniform processing in 1- below, from among the defect candidate pattern openings 9, the size is less than or equal to the base i%r% value,
and;;; the defective reactor Ml whose seawater value is equal to the noise level tl;
Is there a binarized IIF+I image that reflects the correct defect shape? (See candy 3 (θ)).

FIG. 1 is an explanatory diagram of a grayscale value conversion method showing an actual example of the present invention.

% for grayscale images with at least 3 maximum grayscale values of 1 or higher.
t t, i to have a gray level value equal to or higher than the low threshold t2
In tA, the high threshold is 4 or more! , including the il light value tr
A method for extracting only a region will be explained.

The following execution is performed for fl,1 on a pigeon coop where the gray scale value of the picture in the 1st row and jth column of the grayscale image is fij.

〉 tx :(fl:, +2)△(('i-1,j-1
≧t□)” j,, j−1≧11)(fi+1.j
-'1≧11)■> (fl-], j -tl)■(f□+11.!≧1
1)〉 (fi, -1, j, -70)\” '1.., i+
, ≧1. ) (f□ or 0.++, ≧70)) fu: Its extension Where, 11> 1. It is.

...(2) The above equation 2) means that the gray value of a certain pixel is t, hereinafter J-, and at least one of the pixels surrounding that pixel has a gray value t1.
This means that the pixel with the above value is converted into a gray level value of 〇. In addition, other pixels are left as flj.

Then, if the above formula (2) is applied to all images, among the pixels adjacent to the pixels of W41+^t1 or more, those that do not have an f-light value of not less than the threshold value t2 will all have a gray-scale value of t1 is converted to

After repeating this process for all PTII values, the entire screen is binarized with the toy value t1, so that the shading (n\\medo discussion I'a'JJ / Ni-( High combat A 1f+, t +jtQ 4-extraction-d is 7n・ζ'.

I:, Get, the general dark 7′1°b art line phantom 97I
-, Takeshi Ri] is the i-method. Unexploded 1st It-J %
l"V l'b picture j'<,': 6>匝・"()Jeach firj', r゛One day, it is trigonized by the following formula.

Here, a1) an) a8.

After converting to δ value, use the following formula to calculate β capital (1')
Do p-principle.

J: (fi-j4 L2)△((f,,1-0≧8
．． )X/〉 (f, j-0-al) (fj, +1., i-1≧P,
x) (fi-1,,1≧a1)V(f1+1.-28,
) ■ (f knee 11.1, ≧・1) (fl, , i+]
Knee 1■ (f, 1.1 publication ≧81)) f Engineering j: Other 11. (4-) The above equation (4) is the gradation value power of a certain pixel・a 2 L'J,
, l- for each stroke), and the world around that pixel + the i element's l, one of which has a value of 7.7>light value a1 or more, is converted to gray (+fi 淘). This means that other than that (iIτ;
．． Leave it alone.

The shading (ii'j variation processing of (4) is generally 1.
Compared to the 11th image, the memory of the shading %fi is small, so the circuit f,!
,'/li! This can be achieved by

Furthermore, if the following equation Ijp(1) is performed, the grayscale fM conversion process and the ternarization can be performed at the same time.

p, t, : f 1.1(tB !J, 1. : (f ≧t)△((f□−Lj ]
" t1 ) i, j +1 (' j, , ' j-1≧1. ) (ft+t, , 1
−1 = tl)V (f, −,,1≧11)■(f
≧t)1+1. j ] V (fJ, -1,,,1ya,go11)zo(fl,,
++, ≧11)V(f,,~l,;1.□, 2t))
When a, :Other...(δ)
In addition, here, al:>e,') a,, t,>7.

In FIG. 4 1, ++011, +“l”,
Figure 4 (
Assume that the δ-valued image in b) is subjected to the following scanning direction indicated by the IJ line arrow in FIG. 4(8):

To '41 in l (a) G-1, fj to 4 Figure (b) ゛(, tl (or 'l) = 21 ''-2(
t or a, ) = 3, and the cutting allowance for the entire image of one mountain j (2) Ota IJ, j) i + formula (4) %4 ir, 'i
When t is used, σ) is a direct-to-f direct conversion connection KL. Also, 汀<4.1! 1 (di G-1, hill"j bow f"'
? 1, which is the result of performing the 2+2 value conversion process, and ``1'' of the flute cow diagram (θHt, 4th)
Mouth 1 image, L value 1. This is the result when binarized using (or A).

FIG. 5 is a 1q diagram of a size determination method using distance conversion, showing an actual example of the present invention.

In FIG. 5, the defect candidate pattern area is set to "+1 u", and the other areas are set to "0".
The binary image f extracted from the defect candidate Bakun containing only the value
・For *, a method for extracting only defect candidate pattern regions whose size is larger than a certain value is shown.

First, in accordance with the raster scanning direction (not shown by the solid line in Figure 51(a)), the binary image in Figure 5(c) is subjected to the following transformation. That is, 2 (1/+t of image
rJ, + column (1) Find the distance transformed image (di, , ) using the following equation.

(](t)j-(m a X (d 1 book), i,
,□,? 1. ．． ．． , , , , ni l , -le-
,.

Xflj...(6
) In the above equation (6), (1) the direction of the keira anpricot is from left to right, ? , y5, m a y is the maximum value % d-
11,j+ jLl, , 11 rdi, , i-1+,
t sowa each object and mail ii'IIi rice: j'f1・
, left I-1 left P? Indicates the density value of the adjacent pixel, f J, ,
1 is the original 7-illness value of 1-Ning ("1" or 0")
) is shown.

As shown in FIG. 51(a), while scanning the end of the target image wheel y and r with a mask of 4 pixels from 10 to 10 in the direction of the actual bond arrow, the formula (6) above is performed. 'y and rewrite the value of the target town Mugisue. Now, 4 strokes σ) mask dl
-L j i-1,, i -1' 飄, 1-11
dl, 1" 1+ 0+ 1+, d When reaching the area of 1'', 6=J. Since the maximum value of the adjacent pixels is 1, the value 2 obtained by adding 1 to this is set to the target pixel d4.I. Write in.The next moment you are 6 years old, l iir+
i ``The mask shifts to the right by 4g, so ``2.1
．． 1. ], ”, and 2 in this area is the target pixel that was rewritten in the nt+ process, but this time it is an adjacent pixel. At this time, the eye, the adjacent Ptt Since the maximum value of is 2, the value 3 obtained by adding ] to this is written to the target pixel.In this way, the calculation is performed while red apricot, and Q first II I I +
The value of the pixel that was 11 is increased by 11. Then, as shown in FIG. 4(d), the defect candidate area (vertical direction) → (horizontal direction), that is, L QJ, is (horizontal direction) → (vertical direction), that is, the opposite, L
In J's group 1 order, add values ``~9'' to 1~7.

In addition, when scanning in only one direction, the processing is biased in the 1st and 5th
As shown in FIG.
The 1ilti image in c) is multiplied by χ·1 to create the image in Brush 5 (e).

Then, set the values of Cl1 and (e) to the corresponding 1ir
If we compare each other and take the larger value, we get the pattern of values shown in Fig. 5 (f). As a way of getting along,
Both (+Ii.

It is also possible to obtain a pixel value by adding . Next, the 51st >
By binarizing each 1iIIIi element of the grayscale image Φ in l (f) with a value of 9iT in advance, here a convex value, 21 straight lines 1 peak 1f1 in Fig. 5 (g) are converted to I1. There's nothing I can do.

The distance change according to the mask direction indicated by the solid line arrow in Figure 5(a) can be expressed as follows.

(1) written in the L side of the equation indicates scanning from left to right and \, and Pk indicates the number of years of pixels within the defect candidate area.

Indicates that it represents distance.

As an example, if the defect candidate area has a rectangular shape, the above distance number indicates the sum of the horizontal length and vertical length of the rectangle. Therefore, by using such a distance variation, it is possible to determine the size using the length of the side of the defect candidate as a criterion. On the other hand, if the 100 shields are used as a standard for size determination, linear defects may be excluded from defects even if one side of the shield is long. However, as described above, cotton-like defects have the possibility of being fatal defects such as scratches and cracks. Therefore, if the length of the side is used as a criterion as in the present invention, it is possible to eliminate some highly fatal defects and remove defect candidates with a surface area below a certain level.

Next, the distance transformed image (di, , ) of interest+8'<Q(11) desired in the mask direction of 51r; 4 (b) is obtained by the following equation.

...(el) The maximum pixel value of the defective electrical field shown in cutting allowance (7) ('', if the defective shape 1 is a convex or concave shape, the horizontal length and vertical length of the defect ◇11 In order to supplement this, we add the 1111 result of running in the opposite direction.

In other words, 2 given by art formula (6) and the above formula, etc.
(1) (B) Distance transformation image (dlj,) , (,cl,,, ,) of Yuzu
For each corresponding p111 ゛, select the larger one (lj) using the following equation.

(1) (2) d1j= max(d, a) −
--(9, 1i, 1 ij Distance 1G transformed W11 image (d,
,,1), the maximum pixel value "k rra,
Compared to k mFlx, t[t] more accurately reflects the sum of the horizontal and vertical lengths of the defect shape. However-)
So, we only perform distance conversion of l PI+ type! Compared to 1 go, the judgment of the magnitude is more effective.

- to 6 (1″4c, l, configuration 1 of the gray value conversion circuit showing the embodiment of the present invention) IJ.

In FIG. 6, 501 is a convex line shift register that stores 5 lines of two-dimensionally arranged grayscale image data and sequentially shifts the data, and 5o3 is pixel data f to be processed by the 3-line shift register 501. Comparison circuit 502 which compares the magnitudes of s and t with a preset M@ts and provides a binary logic output according to the comparison result.
is the peripheral pixel data ft-1,j-0t+□ of the 3-line shift register 501 adjacent to the pixel data f i,+
, j-1-0 and a preset threshold value t□, and provides a 2m logic output according to the comparison result. 504 is a selection circuit for the eight logic outputs of the comparison circuit convex 02. OR circuit for calculating rational sum, 505 is comparison circuit 5
03 and the output of the main circuit 504; 506 is an AND circuit;
', te, I# value t, or current pixel data fu
Select one of these to create a new pixel footer after n-dark 11t4 conversion (this is also expressed as fij for simplicity)
The ratio of the size of the output value of the pixel data output value ll'n % 50 'γ to the threshold value t1 is calculated as J
1 comparison circuit.

The circuit shown in FIG. 6 performs gray scale inverse processing according to the old method 2), and its operation will be explained below in groups # and ttl.
I will clarify.

Now, it is assumed that the image data of the dark and dark images are sequentially input to the 3-line shift 1 register 501 and are sequentially shifted in synchronization with the pixel data transfer lock. In this case, pixel data f that shadows a predetermined area of the original image arranged in 2υζ elements in the 3-line shift register 501, that is, a 3×5-bit stop shadow grid (not shown in FIG. 6), and its surrounding pixel data. Satj rifi−1,,j −11”tj −111+],,j
-11'i-1. j lfL, 1st edition' fi-1
,, i+l l fi, j+I T fi+l, j
+1 is pulled out from the 5-line shift register 501 every transfer l clock. Pixel data f1. Regarding j, the comparison circuit 503 compares it with the rjd value t, and regarding the surrounding pixel data, the J'f comparison circuit 502 compares it with the threshold value t, and determines whether each pixel data is thicker than the threshold value. l′
If j, outputs °゛1°, and if smaller, outputs 0. When a logical operation is performed using the OR circuit 504, AND circuit, and 505 as output values, it is determined that at least one of the peripheral pixel data is larger than the threshold t1 and the center pixel data is F#.
(When it is larger than m t !, a logic output of "1" is obtained, and for other values, a logic output of "°°0" is obtained. This output value is led to the selection circuit 506, and when the logic output is "1", a
When the current value is t0, the original ii1+i decompressed data f□ is selected and output as new pixel data.The above operation is performed on the entire original image, that is, on all input pixel data.

By performing such processing, the old style can be realized. In case V<, this process is repeated multiple times, so
A plurality of stages of the above processing circuits are provided. Further, the finally determined S surface element data is binarized by a comparison circuit 507 using a threshold value t□, and the grayscale value conversion process is completed.

′? FIG. P7 is a configuration diagram of a determination circuit showing an embodiment of this circuit.

In FIG. 7, there are 61 memory units. Here, for the image data that has not been binarized by the gray value conversion processing circuit,
do.

(11) σ after distance conversion according to each scanning direction)
First, when comparing the two repair results, the pixel positions are made to match.

Perform processing such as. As for the configuration, the force counters 601A to 601B, the force (1 force un etc. 601)
At the end of scanning one line of C1 image data, a pulse of 1 pulse is generated 4 [Scanning end signal VPI) S'' Slip as a strophe signal 70 tubes (302, in image data arranged in 2 lines) Line ] - line・
Memory 60, '5 A to 603 G, selection circuit 6
Consists of 04A, 604B, etc.

*T, -62G;! :, old style (6L (8)
) in 1 PM'KPo>t-,? ) is a calculation unit that performs calculations. The key to planning is the selection circuit 606, the comparison circuit 607, the addition circuit 608, etc.c)'1. 'rru.

As shown in 63 i:l, nit 2〇) in 7, regarding each painting method of the Shin 1 image of the two gods obtained with the old style 0) and (8), This is part 2 of creating a binary image that is easy to judge by binarizing it with a predetermined value and determining the thickness. Selection circuit 6101 ratio circuit 6], 1, 612
Consisted of.

Next, I will explain the previous work in detail.

First, we will explain the size printing based on distance variation that scans in the opposite direction to the rask scanning direction, shown in the half of the block. W11 image data is input serially. Now, the flip 70 tube 602 whose human power code is always l
The scanning end signal 7°8 is input when the signal reaches 1.1 - tC. , the output trace (夛σ)fl of the flip-flop 602
Invert loIo, 1. Now scan ends f't Kano V
Flip-flop 602 by g Q 8
The output Q of is 1 and 1. I feel tired. ko())iii force(
Numbers 7 to - are addition/subtraction counters 601 A, collar L,.../
) to I, )su:, if・1. , -:j:r is input via the inverter, and the counter (+ (1)A operates as a decrementing counter.Furthermore, this counter value becomes the address signal of the 1-line memory 603A.On the other hand, the flip - The output signal of the flip 602 is also input to the write enable terminal of the 1-line memory 6 (13A).
becomes writable f++'-like mF4. The input image data that is serially received by Taylor et al. is stored in the one-line memory 603A in order from the beginning of each clock. In addition, the selection circuit Q Q4.
A's gatey F (-! 10G", also flip 70
The output of the knob 602 is connected to ``r% f-(,! and indicates a value of 1. When >Q 4'F! The output of H t-y <4 is selected and output to the 1-line memory 003 (if is not selected and is not read out).

Next, the scan end signal ■o. 8 is (ξZhi1 and 7:('
2),'), the flip-flop OO21;1, and its output value is inverted by 171-2).Therefore, the addition ρ calculation counter 601A, now IIJ', operates as an addition counter. On the other hand, the l-line memory 603A is no longer writable and the output signal of the l-line memory ('IO3A) is selected in the ♂1φ selection cycle 11':j 604. The data II of the line memory 605A stored in the case is read out in order from the smallest address. Therefore,], the line memory 605A
The serial image data read from 03A is delayed by one line with respect to the serial image data input to the same memory, and the pixel transfer order is reversed. Furthermore, the main image data ζ1 is sent to the α processing section 62, and the distance conversion processing ' to be described later is carried out.
After passing through an AND circuit 609 that performs AND with the IJl result,
When the signal is sent to the one-line memories 603C to 603E, it is sent to the shift register 605B for one pixel.

Now, since the output signal Q of the flip-flop 602 is O, the output signal Q of the flip-flop f302 in the one-line memory described above is write enable! When the 1-line memory 603E connected to 4!4-r- is write-enabled*i<state, the addition p1γ counter 601 Hi:, which becomes the address signal of this memory.
Acts as a counter by subtracting the transferred image −
γ-Y is written into the 1-line memory 603E every clock in order from the one with the largest address.

In addition, the addition counter 60C is the address number 44 and the line memory 603C is always writable.
, il) are included in order from the smallest address.

Furthermore, a scan end signal vv. ,] is pi f@°]
-, the flip-flop inverts its output value and the value of Q becomes 1. At this time, the addition/subtraction counter 601B acts as an addition counter, and the selection circuit 604B selects the output signal of the 1-line memory aO3+> since the gate signal is 7 rip and the selection circuit 604B is connected to the output terminal Q of the flop 602σ. do. Therefore, the image data stored in the one-line memory 603E ('') are read out in order from the smallest address. On the other hand, the image data stored in the one-line memory 603C i=t, the value of the addition counter (]0]C As an address signal, the address signals are read out in order from the smallest one, and then transferred to the shift register (+05A) for one pixel. From the above, the serial image data read out from the first line 603E is as follows: There is a delay of two lines with respect to the serial image data when input to the 1-line memory 603A, and the pixel transfer order becomes the same as the transfer order when input.This image data is already transferred in the opposite direction to the scanning direction. 1F is sent to the determination unit 63 as the result of the distance/music suppression calculation when the signal is aged.In addition, the serial Prm data read from the 1-line memory 003C is sent to the 1-line memory 603A.
There is a delay of two lines with respect to the input serial image data, and the pixel transfer order is opposite to the input transfer order.

Here, the relationship between ], line memories 603A and 603B, and the relationship between ] line memories 603D and 603E will be described. First, 1 line memory f103A
Describe the relationship between this and 603B. Flip now
Assuming that the output Q of flop 002 is d+1,
As mentioned above, the 1-line memory 603
At A, (1, manual data) l is written to jlt'i from the address with the largest number (2). Tie, l → in memory 0
031 (Then, the stored data is the one with the smaller address or eP
It is read out to Iα. Now, if the output Q of the flip-flop 002 is O, then the 11-line memory 603A calculates p from the smaller of +;j and at;2.
'a, 1. Line memory (6 at 10313)
1. "Shikomi" +: negative, in order from the largest address to the smallest address
Ru. However, since the output of the C1 flip/]p knob 602 is inverted every time the - line scan is completed, the output image 7'-ta to the selection circuit 004 is stored in the 1-line memory C3031~ , 6031-3, "IJ Human Power 11. 1 line T of the data
The data 'l'l is always outputted in the I11I order, which is the reverse of the input Illλ order. On the other hand, the relationship between 1 line memory 603 L) and 603 toshi is the same as 603A and 603'FS li'+1o) (5/11 thread), and this memory I The data of one line removed is output in the direct order of input 11 and inverse order of 1.1.Here, the data input to each one line memory is the image data after the distance conversion result, and the pixel The input order is 1 line memory 603A, 603B.
input to ti! This is the opposite of the ii element's ^a order. Therefore, the data output from the selection circuit 604B is data two lines before the image data input to the one-line memories 603A and 603B, and the pixel transfer order is 1. This is the same as the input order to the L-distributed memory.

Now, the arithmetic unit 62A that performs distance conversion in the direction opposite to the rask scanning direction will be described. Manual data input to the calculation section, output data of the selection circuit 604A, 1-line memory 603
C output data, and shift registers 605A, 60
This is 14 pieces of 5B output data. As previously explained, selection circuit 604. Using the output data of A as a reference, the pixel value is set to f1. +, then fi, the output data of the shift register 605B is 1i
The distance conversion value delayed by the iIlli element (note that the transfer order of image data is reversed) indicates d
AI am a, , j, and the output data of the shift resistor θQ5A is the image 51'' i τ'r (F3 is the selection circuit 60'6, comparison circuit No. 007, input Well, the one with the largest pixel value is selected and output as Z.

Furthermore, this pixel data is sent to the addition 1"5 circuit (]C0,
I: the raft to which 1 is added, and by the AND times i & 60 Q I), the aforementioned 1iTli expansion data f s , l and the logic flIJ are taken, and it is calculated only when f j-j (2) = 1. Distance change F'<fr<+'
l, , 1 are output. As mentioned above, the result of this calculation is the minus side Zenith register 605B, 1 line
The message is sent to 003C and becomes a new wavemeter input...], line memory (403+), 60
After returning the pixel transfer order to the transfer order when manually operated in the memory section 01, the data is sent to the 111 treasure section 63.

Next, the size 1 (constant) based on the distance FJI conversion according to the rask scanning direction shown in the lower half of FIG. 7 will be explained.

In the distance transformation according to the rask scanning direction, there is no need to invert the 'Ijl'C sending order of the pixels as in the distance transformation which exists in the opposite direction in pairs 1-5 in the rask scanning direction.

However, if the determination unit 63 does not compare the distance conversion result in the reverse scanning mode with the same number of pixels, it is necessary to add a delay circuit (=J).

The flow of Shun's 11 works will be explained below.

The image data that has been converted into 21-digit data after passing through the digital conversion circuit is first processed as follows.
1 line memory (transferred to 303F).

Further, 7 addresses of this memory are determined by the addition counter 601C. Therefore, the image data is stored in the line memory 603F from the smaller address to the 14th line memory 603F.
be included. Further, reading from the main memory is similarly performed from the address with the smallest address, so the output data of the main memory is data delayed by one line from the time of input. Furthermore, this image data is sent to the arithmetic unit 02B, passes through an AND circuit that performs AND with the distance conversion result, and is sent to the shift register 6Q5D for 9 pixels.
] is sent to line memory 603G. l line memory 603 G's il'i+ ('+ is 1 line memory τ
- Wl of 603F) Exactly the same as the 1st work Pa;, Q, From 1 line memory 603G (:1, Sa r-, D15 I: / /J1t'fi fl j, '7゜After distance calculation result The 1ITii image 7-ta outputs λ., -'s output data is transferred to the sheet register 605C for one pixel at the same time as one of the input data of the calculation unit 62 (3σ).

Also, this output data is ], line menu 603F
For the PIi image data at the time of input, the 2-line delay i and distance calculation total', 1 result is ``t'11 Ube 63'',
sent.

On the other hand, the number j1 shift register 60 to the performance n section 62 B
51) output data, 1 line memory (': (J
3 G output data, shift register (i (+5
C, output data, and 1 line memory 0 (13J
' output data is input. The input data to this is
]-The calculation in the line F32B is exactly the same as the calculation in the calculation unit 62A (1), and the performance village ρf': (1, , 11, l line・
If the data is not transferred to the memory 603 (1), the data is sent to the memory section (33).

To explain the determination section to Europe, as described above, the distance conversion results according to the running direction of fj are inputted in the claw direction and the reverse direction 2 In (the comparator circuit Q],
1. The selection circuit 610 selects the one with the larger value for each corresponding image quality? Select. In general, in the selection circuit 6J-2, if the obtained pixel value is larger than the threshold value t□, 1. When it is small, binary data of 0 and f (() is output.

In this way, it is possible to determine the criticality among the defects Fi=complementary of the binary 1itli image by distance conversion, and extract only correct defects.

As explained above, according to the present invention, the gradation value of Deform the shape of the defect by performing transformation!: 7j
In addition to removing areas with low levels equivalent to noise and noise, by performing distance transformation, small defects that cannot be considered fatal can be removed by considering the defect type. Therefore, it is possible to obtain a defect extraction rule (φ) with less false information.
Takatsuji processing is possible because T can be achieved by simply performing 1C deception and the next step.

[Brief explanation of the drawing]

Rod] The figure is a 70-piece figure of pre-processing VS up to the present invention, ? Figure A2 does not show an embodiment of the present invention, but is a schematic diagram of the entire processing equipment of Yinglin Ye No. 1, Figure 3 shows an example of the fabric of the present invention, and is a detailed explanation of the defect pattern detection method. Figure 5 is an explanatory diagram of the gradation value conversion method of the present invention, which is an example of the present invention. A: Figure 6 shows the actual fil'+i (+'l) of the present invention? This is a diagram. 401.404: Grayscale value explosion 4 circuit, 402: Size judgment circuit, 4.Q3: Assist circuit, 405: Delay circuit, 501 = Shift register, 502, 803゜50
8.607, 611: Comparison circuit, 61: Memory section, 62
: Arithmetic unit, 63: Ban! Department, 601A. 601B, 601C: Counter, 603A to 60δ (1
: Reinme Senri, 506. ,'+07. ri04A. 60A li, 6C)6. 608. 610.
(312! Selection circuit. 1 patent expert, 4 shares) Thank you!
Person Patent Attorney Toshi Iso 1st Figure LK 2nd Figure 02 05 3rd Figure 4th Figure 4

Claims (1)

[Scope of Claims] α) In a defect pattern detection method in which a defect candidate pattern is extracted and a defect is detected from a grayscale image containing grayscale values of at least three buttons or more, a grayscale value of a first threshold value or more is internally selected. A defect pattern detection method characterized in that, among defect candidate patterns, only a portion having a gray value equal to or higher than a threshold value of the rod 2 lower than the first is outputted as a defect pattern. e) Defect candidate patterns are drawn and include at least three types of grayscale values: (a) In a defect pattern detection method that selects defects from frowning images, defective candidate patterns that have internal grayscale values that are equal to or higher than the threshold value of Ml After generating a binary image by outputting only the portion having a gray level value equal to or higher than the threshold of the hole 2 which is lower than the first, all pixels of the complementary pattern of the defect I. A value obtained by adding °゛1'' to the maximum gradation value of a plurality of pixels adjacent to a pixel is executed in the target pixel according to the direction of travel of the vehicle. A defect pattern detection method characterized by binarizing the obtained high-speed conversion image using a preset scale of 3 and 6μ. (3) Defect candidates having a gray value equal to or higher than the second threshold value. The pixel values corresponding to each other are calculated for the image in which only the third defect having a gray value equal to or higher than the first threshold is extracted and the image binarized by distance-converted 2%, f4@. Generate an image of the added value, and again exceed the threshold of t*2^
Defects that increase the light value by 1゛ ('
A defect pattern detection method according to Claim Wl, Enclosure 2, Description 1, characterized in that only defects having a t4 light value are extracted. (4) The distance conversion is performed by changing the horizontal scanning direction by 2f; (
After doing the same, select the maximum value of the pixel values corresponding to nu f L, te, and hi for the two distance M conversion images, Gate (2 by fi
Claim 2 or 3 F! 8. Defect pattern detection method. (6) The distance conversion is performed by changing the horizontal scanning direction and
The feature is that an image of a value obtained by adding mutually corresponding pixel values to two distance-transformed images is formed by 71, and the image is binarized using a third threshold set in advance. Claims 'I42 or 3]'
J″f 41 defect pattern detection methods. (6) The distance conversion is performed by making the distance conversion image performed according to a certain scanning direction and fIt′ correspond to each other by 1 with respect to the initial binary image. The image of the sum of the image values is taken as a natural tooth, and the Pappa Ti1II image is binarized using a preset threshold value of the rod 3. Go to 1) Defect pattern detection method described in Tricho.