JPH03228182A - Image threshold calculation processing method - Google Patents

Abstract

PURPOSE:To reduce the dependency of a threshold on pattern distribution and defect distribution by separating a gradation image to a pattern area and a background area based on a reference image, converting the image by multiplying the frequency of each density stage by a coefficient so as to set all the frequency of a histogram at the same degree, and setting the zero cross point of histogram difference as the threshold. CONSTITUTION:When the density distribution of a targeted image is found, total two density histograms are found by adding respective picture element value on either the histogram for pattern area or the histogram for background area based on whether the position of each picture element is shown as 1 or 0 on the reference image, and furthermore, corrected density distribution is found by multiplying the frequency of each density stage by the coefficient for at least the density distribution on one side so as to set the ratio of all frequency of the density distribution at equal, and subtraction is performed on the corrected density distribution mutually, and the zero cross point is identified, and a density value is set as the threshold. In such a way, it is possible to find the threshold hard to fluctuate for the distribution of the pattern area and the size of a defective area.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an image threshold value calculation processing method for determining a threshold value for converting nine images into 2tE in an image processing process in pattern matching, pattern inspection, etc.

[Conventional technology]

Conventionally, regarding binarization methods related to pattern matching and pattern inspection, as a method based on density distribution.

There are P-tile method and mode method. P tile method.

This is a method of calculating the density at which the cumulative density distribution obtained in order from the lowest density using the density distribution of the input image is P percent, and setting this density value as the threshold value T)If.Hereinafter, this method will be referred to as method ①. The mode method, as shown in FIG. 4, is a method of finding the valley of the fA degree distribution and using it as a threshold, on the premise that the concentration distribution exhibits bimodality. Hereinafter, this method will be referred to as method ■.

Further, the characteristics of human images will be described below, taking as an example the time of pattern inspection in the LSI manufacturing process. A human image of an object input using an electron microscope or the like is degraded by noise mixed in from the fibre-air system. Assuming that this noise intensity distribution is a Gaussian distribution with a variance of σ2 regardless of the signal intensity, the density distribution of the input image from the pattern area Y
The density distribution Y2 of the input image from 1 and the background area is expressed by the formula (
1) It can be expressed by 4 equations (2).

Ya=Ca exp (-(x-sa)"/
σ”) −=−Equation (1) Yb=Cb ex
p (-(x-mb) 2/ a 2) ---
Set (2) where, Ca is the peak height of the density distribution in the pattern area of the input image, -8 is the average value of the image density in the pattern area, cb is the peak height of the density distribution in the background area,
is the average value of image density in the background area. Each pic height Ca
, Cb is proportional to the area of each region.

From the above, the density distribution Y of a general input image consisting of a pattern region and a background region can be expressed as the sum of both head regions by equation (3).

Y=Ca exp(-(x-ma)"/et
2) +Cb exp (-(x-B)"/a
z) Equation (3) Problems encountered when the conventional method is applied to an input image having such characteristics will be specifically explained below.

[Problem to be solved by the invention]

(1) Problems with the noise dependence of the target pattern.

When method ① is applied, the threshold value THI must be obtained from the reference image. If an image defect exists in the target image, the density value at which the cumulative density distribution becomes P percent will change because the target image and the reference image are different. Resulting in.

Moreover, the target image has a pattern (
If there is no noise in the area (hatched in the figure), use formula (3).
The concentration distribution is as shown in 51ffl(b).
: If there is noise in the pattern area, Equation (4)
The concentration distribution is shown as

(Ca-Ca')exp(-(x-ma)"/σ"
)+Ca'exp(-(x-ms)"/+Cb e
xp (-(x-mb)"/a")
--Set (4, where Ca is the peak value of the density noise in the pattern area when there is no noise, Ca' is the peak value of the density distribution in the noise area, and Cb is the peak value of the density distribution in the background area when there is no noise. and Ca+C
a'+Cb is proportional to the area of each region.

When method (2) is applied to this concentration distribution equation (4), it can be easily shown by numerical calculation that the threshold value depends on the amount Ca' which is proportional to the area of the noise region.

As described above, both method (1) and method (2) have the problem that the threshold value fluctuates when noise is present in the target image.

(2) Regarding the dependence on the pattern of the target image and the background area distribution.

As shown in FIG. 6(a), the object has an area 200 that is sufficiently large with respect to the input image, and three small areas 201, 202, 202, and 202 with different area ratios between the pattern and the background.
203 is the target listing. It is assumed that each region has a concentration distribution shown by equation (3).

The concentration distributions of each region 201, 202, and 203 are shown in graphs 2), (C), and (d), respectively.
04.205.206) has a distribution in which the peaks on the high concentration side and the peaks on the low concentration side have different heights. When method ① is applied to this concentration distribution, the threshold value 207.
208.209 can be easily found to have the relationship of 2 (5) by numerical calculation.

(Threshold 209)<(Threshold 208)<(Threshold 207)
---Set (5) From the above, method (2) has a problem in that the threshold value varies depending on the pattern of the target image and the background area distribution.

The density threshold that originally separates the pattern area from the background area depends only on the physical properties of the object and the conditions of the detection system, and is preferably independent of the pattern of the object image and the distribution of the background area. In addition, the main purpose of image processing for pattern matching and pattern inspection is to detect differences between the target image and the reference image as defects. It is desirable that the threshold value remains unchanged even for

References Method ■ J, Doyle: 0 operations
usefulfor 51m1rarity-in
variant pattern recognition
on, JACM, 9. pp, 259-267(1
962) Regarding method ■: J, M, S, Prewit
t andM, L1Mer+delsohn:The
analysis of cell image.

Ann, N, Y, Acad, Sci,, 1
28 pp, 10351053 (1966) The present invention solves the problem that conventional threshold value determination methods are highly dependent on the pattern distribution and defect distribution of the target image, and provides an image threshold calculation processing method that is less dependent on each distribution. is intended to provide.

(Means for Solving the Problems) In the present invention, when determining the density distribution of a target image, depending on whether the position of each pixel is 1 or 0 on the reference image, the pixel value is calculated in the histogram for pattern area or in the background. By adding to any of the region histograms,
A total of two concentration histograms are obtained, and a modified concentration distribution is obtained by multiplying the frequency of each concentration level by a coefficient for at least one of the concentration distributions so that the total number ratio of the concentration distributions is approximately the same. subtract each other,
A zero-crossing point is identified, and its concentration value is used as a threshold.

As described in the conventional technique, there was a drawback that the threshold value determined by the conventional method was dependent on the pattern distribution and noise distribution.9 However, in the case of the present invention, the threshold value determined by the pattern distribution and defect distribution is dependent on the pattern distribution and the defect distribution. Less dependence.

[Example] (1) Example 1 FIG. 1 is a flowchart illustrating a first example of the present invention. In particular, the case where the present invention is applied to pattern inspection in an LSI manufacturing process will be described below. do. In this case, the noise region corresponds to a defective region of a pattern different from the design data.

(A) First, input the target image 100 from the outside (Sl
).

(B) Also, a binary setting image of an area corresponding to the target image 100 is input as the reference image 101 (S2).

(C) Next, for each pixel of the target image 100, the reference image 10
Determine whether the same location on 1 belongs to the pattern area or the background area, and create a density histogram 1 for each area.
02.103 is created (S3).

(0) Next, for each of the density histograms 102 and 103, the total number belonging to each area is determined (S4).

(E) Next, the total number ratio is calculated as (total number of locations for the histogram 102)/(total number of locations for the histogram 103) (S5).

(F) Next, the frequency of each density level of the density histogram 102 is multiplied by the inverse ratio of the total number ratio to create a new modified density histogram 10.
Create 7 (Sc).

Note that when creating the corrected density histogram, the area ratio between the pattern area and the background area in the binary reference image may be used instead of the total number ratio of the histogram obtained from the target image. In this case, step S4. In S5, the area of the pattern area, the area of the back 1911 area, and the area ratio thereof are calculated. Furthermore, since the binary reference image is used for each pattern matching, if only the area ratio is calculated in advance and stored in memory, step S4. It goes without saying that S5 is unnecessary.

(G) Next, the corrected density histogram 107 and the density histogram 103 are subtracted from each other to create a density histogram difference 10B (S7).

(H) Next, the zero-crossing point of the density histogram difference 108 is identified and output as the threshold value 109 (S8).

The above algorithm is applied to an input image mixed with Gaussian distributed noise of 1 variance σ1. The image density histogram of the pattern area and background area is expressed by equation (1).

Assume that equation (2) is followed.

☆1. Regarding the dependence of the target image on the pattern area distribution.

When the density histogram 102 shown by equation (1) is multiplied by the inverse ratio C2/CI of the total number ratio in step S6, the corrected density histogram 107 becomes equation 1 (6).

Y1=C2exp (-(x-ma)"/a"
) - Equation (6) Therefore, from Equation (6)1 Equation (2), the zero crossing point of the density histogram difference of density histogram 107°103 is.

(lIlb+II+,)/2-m- Formula (7) is obtained. Therefore, it can be seen that it does not depend on the total number ratio.

☆2. Regarding the dependence of the target image on the defect pattern.

Pattern area 12 as shown in FIGS. 5(b) and (C)
When there is a defective area 121 in 0 and the background area 122
Consider the case where there is a defective area 123 inside. Here, it is assumed that the image density in the defective area 121 has the same density distribution as that of the background area, and that the image density in the defective area 123 has the same density distribution as that of the pattern area.

As a result of the process in step S3, each defect area 121,
If the peak value on the concentration histogram of 123 is ClC2', each histogram will be as shown in Equation 1 (8) and Equation 1 (9) and Figures 3 (a) and (b), and the contribution of defects α and β is affecting the distribution.

Y1=(C1-C1')exp(-(x-mm)2/
a ”)+C1'exp(-(x-mb)"/
a ”) formula (8) % formula %) formula (9) Here, in steps S4 and S5, formula (8), 2 (9
), the density histogram 102 (Equation (8)) is multiplied by the inverse ratio of the total number ratio (area ratio), and the density histogram 107 (Equation 00
)).

Yl' = (C2-C2*C1°/C1)exp(-
(x-mm) ”/e”)+(C2*C1'
/C1)exp(-(x-mb)"/a2)
--- Complete set〇〇) In step S7, the difference between the density histograms 103° and 107 is calculated and the histogram difference 108 (
201)) is required.

Y3= (C2-C2°-C2*C1°/C1)exp
(-(x-m,)2/a 2)(C2-C2°-C2
*C1'/CI)exp(-(x-mb)2/er
”)-(C2-C2'-C2 CI'/C1) (
exp(-(x-m, )27 az)exp(-
(x-tab)"/σ")) -11200 In step S8, when finding the density value X that satisfies the formula (1υ-), it is clear that this density value X is a value that does not depend on the area of the defective region. , can be used as a threshold.

As described above, according to the method of the present invention, it is possible to obtain a threshold value that does not easily fluctuate with respect to the distribution of pattern areas and the size of defective areas.

(2) Embodiment 2 FIG. 2 is an overall configuration diagram of an apparatus explaining a second embodiment of the present invention. A case will be described in which the present invention is implemented in a system that inputs a target image from a submicroscope and design data from an external storage device, binarizes the target image, and monitors the target image for the purpose of pattern inspection in the LSI manufacturing process. .

Reference numeral 50 is an image signal source such as an electron microscope that takes an image of an object, and outputs nine image signals 54.

Reference numeral 52 denotes an external magnetic memory that stores design data 49 of the object, which can be referenced through one communication path 51 from a computer or the like on the communication path 51.

Reference numeral 20 denotes a target image input unit which inputs an image signal 54 from the outside and outputs a target image 55. 21 is.

A reference image input unit inputs design data 49 corresponding to the same area as the image signal 54 from an external magnetic memory 52 via a communication path 51, and inputs design data 49 corresponding to the same area as the reference image 54.
Outputs 6. 22 is the target image 5 with reference to the reference image 56.
5, a histogram creation unit that creates density histograms 57.58 for each of the designed pattern area and the designed background area; 23 is a density histogram 57.58;
Find the total number of pattern areas and background areas based on.

Furthermore, a total number ratio calculation unit that calculates a total number ratio 59.

24 is a histogram conversion unit that multiplies the frequency of each density level by the reciprocal of the total number ratio 59 for the density histogram of the pattern area to obtain a new corrected density histogram 61; 26 calculates the difference between the density histogram 58 and the corrected density histogram 61; This is a histogram difference calculation unit that calculates a density histogram difference 62. Reference numeral 29 denotes a zero-crossing point identification unit that determines the zero-crossing point of the fA degree histogram difference 62 and outputs it as a threshold value 63. Zero cross point identification section 29
Now, if the density histogram difference 62 contains noise exceeding a certain level, the neighborhood is approximated by a linear equation or a cubic equation to find a zero-crossing point. 27 is the threshold (1
28 is an image binarization unit that binarizes the target image 55 based on 63 and outputs the 2(! This is an image output unit that outputs images.

(3) Embodiment 3 The total revenue ratio calculation unit 23 in Example 2 can be replaced by a total revenue ratio calculation unit that uses the reference image 56 as an input and calculates the frequency ratio between a similar pattern area and a background area. be.

[Effects of the Invention] As described above, according to the present invention, it is possible to provide a processing method for calculating a threshold value using a reference image with less dependence on pattern distribution and noise distribution. Note that in this case, it is not strictly necessary that the area ratio of each region becomes 1 as a result of the histogram conversion, and a certain degree of deviation is acceptable.

Therefore, if the calculation speed is an issue, the histogram conversion can be omitted if the total power ratio is not far from 1.

[Brief explanation of drawings]

FIG. 1 is a flowchart of an algorithm according to an embodiment of the present invention, and FIG. 2 is an overall configuration diagram of an embodiment of a system for binarizing and monitoring microscopic images according to the present invention. Fig. 3 is an explanatory diagram of the histogram difference of the present invention, Fig. 4 is an explanatory diagram of the density histogram of the input image, Fig. 5 is an explanatory diagram of the dependence of the density histogram on defect patterns, and Fig. 6 is an explanatory diagram of the threshold value pattern. An explanatory diagram of dependence on distribution. In the figure, S3 represents creation of a density histogram, S4 represents calculation of total number of hits, S5 represents calculation of total number of ratios, S6 represents creation of corrected density histogram, S7 represents creation of density histogram difference, and S8 represents zero crossing point identification.

Claims (1)

[Claims] In calculating a threshold for binarizing a grayscale image that has a binary reference image and includes noise, the grayscale image is divided into a pattern area and a background area based on the reference image, A density histogram is created for each region, and at least one of the density histograms is transformed by multiplying the frequency of each density level by a coefficient so that the total frequencies of the two density histograms are approximately the same. If only the converted density histogram is converted and the unconverted density histogram, and if both are converted, subtraction processing is performed between the converted density histograms to find the histogram difference, and the histogram difference is calculated. An image threshold calculation processing method characterized by using a zero-crossing point of a histogram difference as a threshold.