JPH03177040A - Pattern recognition - Google Patents

Abstract

PURPOSE:To reduce influence of a sampling error by registering a pattern on the basis of the pixel unit and then registering it in the precision of a subpixel. CONSTITUTION:After respective parts are initialized by a command from an overall control part 21, a two dimensional pattern is detected by photoelectrically converting the pattern on a wafer 1, illuminated by a light source 3 and a illuminating optical system 4, in synchronism with the scanning of an XY stage 2 with a one dimensional image sensor 6 via an objective lens 5, and it is digitized by an A/D converter 7 to obtain a two dimensional detected image 10 and this detected image is stored in an image memory 8. An image takeout part 12 takes out a reference image 11 referring to predetermined addresses for the image memory 8. Then, the pattern is registered on the basis of pixel as the unit by using the detected image 10 and the reference image 11 and the pattern, registered on the basis of the pixel unit beforehand, is registered in the precision of a subpixel, and the errors in said two patterns is extracted and compared to recognize the defect of the pattern.

Description

[Detailed Description of the Invention] [Field of Application of Togami] The present invention is a method for recognizing defects by comparing patterns of LSI wafers, TPTs, etc., and is particularly applicable to pattern recognition methods based on accurate alignment. I feel excited.

[Conventional a toughness]

The conventional pattern recognition method is disclosed in Japanese Patent Application Laid-Open No. 57-196577.
Detect the target pattern as described in the publication, and! lt
Extract the pattern that was created, align the pattern that was stored in memory and the pattern that was released in units of m, and extract and compare the difference between the two aligned patterns. This resulted in g-weaving with defective patterns.

This target is the pattern of a semiconductor wafer such as a memory LSI as exemplified in Fig. 2 (g), (j) (c),
These include patterns for P.transistor), KM plate patterns, ceramic substrate patterns, and patterns for masks and reticles used in the machinery used to manufacture them. Here, M5 is used as an example for semiconductor wafer patterns, but other patterns (
b) The same holds true. The pattern of the semiconductor wafer is finally cut out and the chips that become individually dyed crystals are called AlOM1.
They are placed on two wafers, and they have the same pattern as each other. The principle of recognizing such pattern defects is shown in FIG.

&, 21NN (A3 to (C) are illustrations explaining the principle of a pattern defect inspection method using a conventional general pattern comparison method, m2 (α) is a memory pattern, and 21NN (b) is a detection pattern. (c) is the pattern difference.Noting that each chip has exactly the same pattern, pattern '4I of Figure 2 (α): is detected and stored.
Figure 2 (C) Next, detect another pattern that should be the same as that in C, align the two patterns pixel by pixel, and extract the error between the two aligned patterns in Figure 2 (c). and compare.

If there is no defect in any of the patterns, there is almost no amount of patterns, but if there is a defect in any of the patterns, for example, the detection pattern 211 (b), as shown in Figure 2 (C). Since there is a difference in the pattern at the defective portion, the pattern defect can be identified as M* by comparing the patterns and detecting the location where the dA difference occurs.

Here, if the patterns are compared and there is a difference, it can be said that one of the patterns has a defect, but it is not possible to determine which pattern has the defect. There are several methods to determine this, but the technology according to the present invention is F! I especially want to read it here because it doesn't have any value.

[A task that Akio is trying to decide on]

In the above-mentioned conventional technology, since the pattern is digitalized and input, only the information of the sampling point is ignored, and the occurrence of sampling error 4 difference can be avoided. There is a problem. This will be explained in Sections 5-(a) to 0).

The 3rd-(α) to (-) are the modification buttons for the patterns of Raku 2 (α> to (C)), and the 5th-(α) is the detection gI waveform of the 2nd-(0 memory pattern). , the fifth J (j>) is 1K2N1
．． Detection polarization technique of 61 detection patterns, 5th m < c
> is i 21W (a
) detection when there is no sampling error for the Ctt pattern I! Figure 5 (m) is the difference between Figure 5-(α) and Figure 5 (with sampling error).
Figure 3 (-) is the difference sign technique between @ 5 [(l and the 5th i (C') without sampling error, and the .
The mark indicates detection number 11 at the sampling point. Fifth
-(Season.

(As shown in item 5, it is not possible to set the sampling points the same for each pattern even if the pattern is originally the same, so the detected waveform at the sampling point will be different, and the detection An error occurs in the pattern, and this wA difference is called a sampling error.If there is no sampling error in the detected pattern to be compared, the defective part number is sufficiently larger than the normal S difference signal as shown in No. 5-0>. It is easy to gWR the defect, but if there is a sampling error in the detection pattern, the defect S difference signal becomes Pi & degree with the normal part difference signal, making it difficult to detect the #llt fiber of the defect. .

SUMMARY OF THE INVENTION An object of the present invention is to provide a pattern introduction method in which the influence of sampling errors can be eliminated by adjusting the pattern position and setting the f'Iv degree to be less than the detection pixel size.

[Means to solve the problem]

In order to achieve the above object, the pattern line fiber method of the present invention aligns patterns in pixel units, and then aligns them with precision less than pixel units. In other words, the target pattern is detected, the detected pattern is memorized, the previously stored pattern and the detected pattern are aligned pixel by pixel, and the ll1
1lil for the pattern aligned to II elementary unit
+Extract and compare the errors of two patterns that are aligned with an accuracy of less than an elementary unit and that are aligned with an accuracy of less than a millisecond.

This is a defective MITI of the pattern.

For example, the following method of least squares is used as a means for aligning with an accuracy of less than the pixel unit. In other words, when the two patterns are / (J# y) -i (Jca y), t, (dJ, d
fii (dx6, dyo) that minimizes y)
Align in 1iII cumulative units, and both J and y loci are 0
Alignment is made so that there is a position (δ'Oaδyo) where the difference between the two patterns is the small distance between the 0th and 1i111iXth patterns.

gtCctxedy)=gcdx*dy)+#(d
J +1*'y) 10g (dJ:, dy+1) 10g (
da+1 , dy+1 ) (1) t (d, x, dy) = ΣΣIj (x, y) -1Cx+
dx, y+dy) I 121 Here, J and y are the counter-spots of the pixel-by-pixel pattern, dx.

ty is the pixel-by-pixel alignment of the two-sheet υ pattern ↓ (
δJ, alignment amount s less than δyt'zIIliIs
``'on '3'o is the alignment amount da; dy for each rtii cord that minimizes ε!.

Minimize dX6, dyo beam'ii! The alignment weights δX, δy, and ΣΣ below ll are the X of the circumference to be aligned.

The sums related to y-locus similarity are shown respectively.

In order to align the rays in pixel units, the following (1 shown in formula 51 is shifted) is used.

j's CIJ) = ! I(J + dX −J +
dy) (The intermediate distance between 51 pixels and the IIkl line is defined by the formula (41, (5)) jtcx+δX, J+δ
y)=fCx, y)+δx*(cX+1.J)-f(x
, ysu◆δy books (Hax, y◆1) - Hax, y)
(4)lldc'+δx,y−δ! ＞=1t C
xey) 10 δx*cyl<x-1,7>-! A<x,y
))◆δy*(,y, (J, y-1> -yx t
a a y)) (532 meters error can be defined by equation (6).

cd(δX, δy)=ΣΣ(7'i(J◆δJ, 31+
δy)−! s4(J-dx, y-ay))**
2 16) Differentiate equation (6) with dJ and ay and set it as =Ω, rearranging it as 171, 1131
despise the ceremony.

81 Here, co= fcxay)-tt (J#J)CL=(j
CJ+1, y) −j < JCmJ Lee(y, t
, r-t,y)-yX(,r,y))t91C)=C
j (”+ y+1)-ha, r, y))-(y, (J
cay-')-9Hc'ca'/)) - Calculate the alignment temporary pattern ft, 5/ from the alignment electric quantity δJC6eδyO under the elementary block using the following equation (1Ω) and (11).

f@ (J@y) "j4<2+δ"o+y+δ3'o
) (1(1)!lx ('#3')
=lrd<'-δX6ay-δy(1)
(11) [Operation] The operation of the above pattern recognition method will be explained with reference to WI4 diagrams (1 to ldl and 3145 diagrams).

Group 4 (g) to (d) are waveform diagrams of an operation example of sub-pixel machining of the pattern shown in Figure 2 according to the present invention, Figure 4 (α) is a stored waveform, and Figure 4 (b) is a The constructed waveform without defects, the fourth figure ((') is the simple difference technique, and Figure 4 (4) is the sub-vixel alignment and difference technique. Figure 5 is the figure 4 (α)
In the numerical table diagram of ~(d), the sampling position tILΩ ~15
Memory shaping, detection waveform, simple difference waveform, and f* −! after sub-vixel alignment. lx-IIs-! The numerical value of h l is shown respectively. For example, the memorized waveform and constructed waveform are 74a(3)
(α), (Ha and yLs diagram).

Here, the detected waveform of No. 41J (A) is equivalent to the waveform obtained by taking the average of #2 strokes of the stored waveform of 0, and shifted by approximately 0.5. Applying the above least squares method to the waveform of
11) Positioning pattern No. 4 using formula. Find h. At this time, the difference between -V''1 pixel with and without rainbow alignment is shown in Figure 4 (tt), (
d) and Figure 5, the residual error has been halved.

As a result, the pattern difference due to the sampling error becomes sufficiently smaller than the defect, so that the defect can be clearly identified.

[Fruitful example]

The present invention will be explained below with reference to FIG. 1, FIG. 6, and FIG. 7, which are practical examples of the present invention.

The first example is the first example 2A of the pattern recognition method according to the present invention.
1 is a diagram showing the pattern recognition device of an LSI Quafer shown in ilfil. In this example, LSI wafer tD /<
/ -7'x: HK tDt, but it cannot be applied to patterns such as TET.

This pattern - the organ is parallel to wafer 1
, a light source 5 illuminating the stage 2 and the wafer, and an illumination light string 4
and an objective lens 5 that detects the optical emission of the illuminated wafer.
1) Image input s consisting of a converter 7 and an image memory unit 8
9, and the detection image 10 and the comparison image 11 inputted to the image input section are taken out from the six-temple memory 8 by the engraving section 12 and the detection image 10 is compared with the image 1I! From ll image 11 (2
) Calculate the image difference expressed by the formula and calculate the comparison tOW by (5
Li moves and aligns like the 1st set! 11 from the il elementary matching unit 15 and the pixel unit matching unit
Iil elemental unit position correction completed! i-image 14 and detection-images 10 to 17), a sub-pixel matching section 15 that tints the positioning process δJg mδ3'o below the pixel expressed by the ta1 formula, and a positioning summation based on 7゛ pixel matching. (10) The position alignment unit 16 performs the position correction expressed by equation (11), and the shadow extraction unit 18 extracts the um* left image 17 after alignment, and the difference mt* is 2. A 11111 image processing unit 20 consisting of a defect determination unit 19 that extracts each yuzu characteristic amount in a place where there is a difference and makes a defect determination, and an image processing unit 20 that controls the XY stage 2 and outputs from the image electrode agglomeration unit 20 It consists of an overall control section 21 consisting of I'', which performs the maintenance and display of defect records and the WS of the entire sequence.

The operation of detecting pattern defects using the above configuration will be described next. First, after initializing the 6 evil by a command from the overall irt+a section 21, there is an interval between the scanning of the XY stage 2,
The pattern on the wafer 1 illuminated by the light source 5 and the illumination optical system 4 is photoelectrically converted by the one-dimensional image sensor 6 through the objective lens 5Y to detect two original patterns, and
A two-dimensional detected image 10 is digitized by the D converter 7, and the obtained detected image is stored in the image memory s8. The image area extractor 12 extracts the comparison image 11 by referring to a fixed address in the image memory section 8. Here, the operation of pixel unit matching @IS will be explained using m6(3).

Xie 6 Yu is Audience 1 (3) - Explanation of the operating principle of the search unit matching unit 15. Above detection-image 10 and comparison image 11
Therefore, the comparison image is ΔX, ΔY-)j, and the positional deviation permissible type δ pixel (this will be explained in 6-21 in this paper). Determined 1
1iL, and the required short value is set appropriately), calculate the wheat of the detected image 10 and the comparison image 11 by L2) formula, ss
) is the minimum of dx6 a Qo CW46□□□Δ
X=-1. ΔY=0) is calculated, and the comparison image 11χ(51
The position is corrected using the formula, and the image 14 whose position has been corrected in units of one pixel is output.

The sub-vixel matching s15 is performed using the image 14 that has undergone the pixel unit fill correction from the pixel unit matching unit 15 and the detected image 10.
! Calculate alignment values δxO, δy0 of i or less.

The alignment capital 16 performs position correction expressed by equation (10) and (It) based on the alignment violet from the sub-pixel matching unit 15. The difference m-body extracting unit 18 extracts a difference image 17 from the image for which the furisode is completed using the following equation (12).

S(&,)) = 171 ('-))-h C
(12) The defect determination unit 19 converts the difference thl image 17 to 2 at the defect determination threshold Vth.
11! Defects are determined by extracting the area of the location where the quantity exists, the convergence, the projection length, etc.

According to the unfortunate example, the two patterns, the detected image and the stored image, are both moved in the same but opposite direction to create an alignment error below the LIIII element.

Smoothing effect of m-image (e.g. δJO=Ω, 5, ax6
== If O, then j*<' e y) = (j(x”
m y) + fCx e y))/2) which is equivalent to shading the average filter) becomes a sharp contrast between the two patterns, and the resulting difference image! #4 It has the effect of minimizing the difference.

Modifications of this implementation include the following. That is, XY
Instead of detecting a two-dimensional pattern by performing photoelectric conversion with the IVC source image sensor 6 during the P1 period of the scanning of the stage 2, the XY stage 2 is moved step by step to detect the TV.
A two-dimensional pattern is detected by photoelectric conversion using a camera. Alternatively, instead of the one-dimensional image sensor 6, any type of sensor can be used, such as a point-type sensor such as a photomultiplier and a scanning structure.

In addition, instead of calculating the difference between the detected image and the comparison image using the equation (2) and outputting the difference between the images corresponding to each shift amount as a myching value, edges are The image difference k (calculated using equation 21 and the difference between the images corresponding to each shifted child) is extracted and output for the two centers of the edges. Alternatively, the detected 1iII image and the Binarize the edges and calculate the difference between the images using (2)
It is calculated using the formula and output as the difference between the m images corresponding to each shifted child. According to this modification, since edges are used, there is an effect of making it less susceptible to the effects of differences in pattern cutness between the detected image and the ratio 4!211111 image.

Also, (s), (91e) Instead of equation (10), the following (
15), (14).

(15) is used.

j'tCJ: +δx, y+δy) = fcx, y
) (15)'o"'j'cJa y)
−Is (” # y) Ci=−(y, (Jc−t,
y)-y, (z, y)) (14) CJ
= Knee (!I, (J ey-')-y>
CJ −y)) f* Cx a y) = jtC
x e y) (15) In this modification, only the comparative image is changed without changing the detected image, so the structure is simple.

Furthermore, instead of using the image that was stored after detecting the comparison 1ilII image, a previously stored non-defective item m image is used. According to this modification, there is a young fruit that can be recognized without missing any defects even if they have a tendency to be kneaded. Also, instead of using an image that has been detected and stored as a comparative image, another set of detection systems is provided and the six images detected by this are used. According to this modified example, there is an effect that the image memory size and unnecessary IL circuit requirements are reduced.

Figure K7 is a BG diagram of pattern recognition of an LSI wafer showing the second embodiment of the pattern recognition method according to the present invention. This pattern-weaving apparatus consists of an XY stage 2 that scans the wafer 1, a light source 5 that illuminates the wafer, an illumination optical system 4, and an objective lens 5 that detects the illuminated Kuehler image.
a detection unit consisting of a one-dimensional image sensor 6; an image input unit 9 consisting of an A/U conversion image memory unit 8 for digitizing and storing the QS of the one-dimensional image sensor 6; -1i#! 11 from the image memory section 8 & 1i12, the detected image 10, and the comparison image 11 (Calculate the difference of 1m11!f! expressed by equation 21 and output the alignment of the comparison image t 'JO@ ``io (16), (1
7) Alignment amount δX of 1thi or less expressed by the formula
6. Sub-pixel function ′J! to calculate dXo! # part 22
and a subpixel calculation unit 16 that performs position correction expressed by equations 153, (10), and (11) based on the alignment amount from the sub-pixel calculation fB22, and a difference image that extracts the M image 17 of the aligned image. The error extraction unit 18 converts the difference image into 21 pixels and extracts each tlth1% of the Mh location where the difference exists.
A 1IkI image processing unit 20 consisting of a defect determination N19 that extracts unfinished areas and performs defect determination, controls the XY stage 2, stores and displays defect information output from the image processing unit 20, and manages the entire sequence. It is composed of an overall control section 21 consisting of five computers.

The operation of detecting pattern defects using the above configuration will be described next. First, after initializing each part according to a command from the overall control unit 21, in synchronization with the scanning of the A two-dimensional pattern is detected by converting light with a one-dimensional image sensor 6, and AI
D': Two-dimensional detection surface 1j digitized by JLm device 7
R10, and the obtained detected image is stored in the 1lIiII image memory section 8. The image retrieval unit 12 retrieves a comparison image by referring to a fixed address in the 1lIII image memory unit 8. The pixel unit matching unit 15 is the detection surface II.
! 10 and comparison image 11, the comparison 1illI image is ΔX, Δ
±δ11ii11. of positional deviation tolerance in Y direction. @ Comparison with detected image 10 when using Dage Ding II! l111!11
Calculate the difference between the images by Wt using formula (2), minimize 6 of formula 11, and calculate dJc@ * 'y6 'l to calculate ε and cLxl
) d sub-pixel performance 314s2 that outputs e dyo
2 is ε and d−x@e from the pixel unit matching unit 15
From dy6, the sub-pixel alignment common δJ(1, δ'/) is given by the following equations (16) and (17).

Calculate.

”% = (' ””o * '31(1)+g Cd
x6 e'yo "1)-g Cd50+ 1, d
yo)-1(cLJ(1+1,Ql) ) )/
#@ (j6)δ'io= (' (
“On”! 10) Su (dxo, ct yo◆1
) 10g(dx0◆11αyo)-a<dX6 +1*'
yo ”1 ) )/a(gr (17
)Khan”g (dXor’yo)”g (
cLJ@a dy6 ” 1) + g (dx6 +
1+d3/6)+ε(ctx6+1, d
yl + 1) The alignment unit 16 uses +518 (1Ω) based on the alignment value from the sub-pixel calculation unit 22,
Position correction expressed by equation (11) is performed. The difference image extraction unit 18 has finished position correction! From l1111 (
The difference image 17 is extracted using equation 12).

Defect determination s19 uses the difference image 17 as w4 value Vth for defect determination.
Convert to 21 and calculate the area, width, and projection of the place where the wheat is located.
Defects are determined by extracting the characteristic quantities of each red oak. , δ
The calculation of y0 is simple and has the effect of reducing the -travel formation.

Modifications of this embodiment include the following. That is, (1
0), (It), the following (18), (
19) Use formula.

j', (J, y) = (1-δJO)*j(J,
y)◆δ, r0*(1-δ!io ＞*j (J+1
.

y)◆(1-δx0)*δyo*7(xsy◆1)◆δ
x0*δyo*7('◆1゜y◆1)
(1B)yl(,c,y)=(1
-δJ(1)*(1-ri@)*yl(,z,y)+δJ
O*(1-δy0n*31.<x◆1.y)+(1-δ
x0n*δyo*j't (xeJ◆1)◆δ, λ;
.

*δ! o*! h ('◆1.y◆1)
(19) [Effects of the Invention] According to the present invention, the influence of the sampling #A difference can be reduced by setting the pattern alignment at 8 degrees below the detection pixel size.1
This has the effect of easily detecting defects with a size of 81 degrees, which is the same as the pixel size.

[Brief explanation of drawings]

Figure 1 shows the first problem A3 of the pattern recognition method according to the present invention.
A configuration diagram of the device showing example i, Figure 2 (41 to tC+ is a diagram explaining the principle of pattern defect detection method using the conventional pattern ratio method, A3 Figures -1 to (-1 are Series 2 Figure 18) ~I
C) pattern technique example diagram, Theory 41V (cl to u1 are the patterns of Figure 2 according to the present invention) ゛Vixel matching operation example waveform-1 Figure 5 is Figure 4 (g) to -) FIG. 6 is an explanatory diagram of the operating principle of the pixel-by-pixel matching shown in FIG. 1, and FIG.
This is a configuration of an apparatus showing an embodiment of the present invention. 1...Wafer 2...XY stripe 3.
...Light source 4...Illumination #J optical system 5...
Objective lens 6... One-dimensional image sensor 7... A/D converter II & device 8... Image memory section 9... Image input section 10... Detected image 11
...Comparison I! Ii image 12... Image extraction unit 15... Pixel unit matching unit 14... - Image 15 after cable unit position correction...
Sub-pixel matching section 16...positioning section...difference image 18...
Difference image extraction unit 19...Defect determination unit 2Ω...Image processing unit 21...Overall control unit 22...Sub-pixel calculation unit representative

Claims (1)

[Claims] 1. Detect a target pattern, align the detected pattern with a pre-stored pattern or a separately detected pattern pixel by pixel, and compare the pixel with respect to the aligned pattern pixel by pixel. Align with the following accuracy,
A pattern recognition method characterized by recognizing defects in patterns by extracting and comparing errors between two patterns aligned with sub-pixel accuracy. 2. The pattern recognition method according to claim 1, wherein the method of aligning at or below a pixel uses a least squares method. 3. The pattern recognition method according to claim 1, wherein the method of aligning with an accuracy of less than a pixel is performed by estimating a matching value when aligning on a pixel-by-pixel basis.