WO2003079292A1

WO2003079292A1 - Defect inspection method

Info

Publication number: WO2003079292A1
Application number: PCT/JP2003/003174
Authority: WO
Inventors: Shunji Maeda; Kaoru Sakai; Takafumi Okabe
Original assignee: Hitachi High-Technologies Corporation
Priority date: 2002-03-18
Filing date: 2003-03-17
Publication date: 2003-09-25
Also published as: JP2003271927A

Abstract

A defect inspection method for comparing an image to be inspected with a reference image to obtain a difference, according to which a defect can be found. In order to prevent false information caused by luminance difference between the images, conventionally, a large threshold value is set to lower the sensitivity. In order to solve this problem, the difference between the image to be inspected and the reference image is reduced by performing scatter diagram decomposition and comparison or adjustment of the object to be compared. This allows a difference between images caused by thickness difference in a wafer and can prevent generation of false information without lowering sensitivity.

Description

Details

Defect inspection method Technical field

The present invention compares an image of an object, such as a semiconductor wafer, a TFT, or a photomask, obtained using a lamp light, a laser light, or an electron beam, with a reference image stored in advance, and uses the difference. The present invention relates to a defect inspection method for inspecting fine pattern defects and foreign matter. In particular, the present invention relates to a defect inspection method suitable for performing a visual inspection of a semiconductor wafer. Technology background

As a conventional technique for performing defect detection by comparing an inspection target image with a reference image, a method described in Japanese Patent Application Laid-Open No. H05-2646467 is known.

In this method, a sample to be inspected in which a repetitive pattern is regularly arranged is sequentially imaged by a line sensor, compared with an image with a time delay of the repetitive pattern pitch, and the mismatched portion is detected as a pattern defect. . However, the actual position of the two images is not always the same due to the vibration of the stage and the tilt of the target object.Therefore, the image captured by the sensor and the delay of the repetitive pattern pitch The amount of positional deviation of the obtained image is obtained. Then, after aligning the two images based on the obtained positional deviation amount, a difference between the images is obtained, and when the difference is larger than a specified threshold value, a defect is determined, and when the difference is small, a defect is determined. Defects, that is, normal.

The problem of the above-described conventional technology will be described. For example, in FIG. 1, 11 is an image to be inspected, 12 is an example of a reference image, 1a is a uniformly bright background area, and lb is an area having a dark pattern on a bright background. In addition, inspection target image 1 1 has defect 1c. In the image of this example, the waveform of the brightness at the position 1D-1D 'is as shown in FIG. 1 (b). Here, the misregistration amounts of 1 1 and 1 2 are obtained, and the difference image after the alignment of 1 1 and 1 2 is as shown in FIG. The difference image is an image that is displayed in a gray scale according to the difference value at each corresponding position between the inspection target image and the reference image. If a portion where the difference value is equal to or larger than a specific threshold value TH is determined as a defect, only the defect 1 c of 11 is detected in FIG.

When the inspection target is a semiconductor wafer, if there is a difference in the film thickness within the wafer, the inspection target image and the reference image are the same as shown in (a) 4a and (b) 4b in FIG. A difference in brightness occurs in the pattern of (1), and the value of the difference increases as shown by 4c in FIG. 4 (a). This is a false alarm, and the threshold T H must be increased to prevent detection. Alternatively, the threshold value must be set separately for the area with uneven brightness and the area without brightness.

As described above, in the case of a semiconductor wafer, in the sense that the positional accuracy of the pattern is high and the positional information is reliable, how to process the ambiguous brightness information with respect to the relatively accurate spatial (position) information. But this is a big issue. Disclosure of the invention

The present invention provides a comparative inspection in which an image to be inspected is compared with a reference image and a defect is detected based on a difference between the images, and at the time of brightness comparison, data is voted for in a scatter diagram which is one of a multidimensional space. The problem of the conventional inspection technology was solved by decomposing the scatter diagram based on the features and suppressing the spread of data on each decomposed scatter diagram so that a low threshold value could be set.

Here, the scatter diagram, which is one of the multidimensional spaces, is obtained by taking the brightness of the inspection object image and the reference image on the vertical and horizontal axes, respectively. As a result, false alarms due to color unevenness are reduced, and a highly sensitive defect inspection method and apparatus are provided. To provide. In particular, in the inspection of semiconductor wafers, pattern brightness unevenness caused by differences in film thickness, etc. is checked by adjusting the brightness between images without increasing the threshold value TH. The aim is to reduce false alarms caused by uneven brightness and realize highly sensitive defect inspection. Note that, in the comparative inspection, the brightness is described here as a comparison target, but when the brightness is a target other than the brightness, the brightness is taken on the vertical axis and the horizontal axis of the scatter diagram. Also, three or more feature values may be selected to make the scatter diagram multidimensional.

That is, according to the present invention, in the comparison between the inspection object image and the reference image, a scatter diagram is created by voting (by plotting), the obtained scatter diagram is decomposed based on the feature amount, and each of the decomposed scatter diagrams is decomposed. There is a method that can set a low threshold value by suppressing the spread of data on the diagram. Furthermore, even if there is a difference in brightness between the same patterns between images due to differences in the target film thickness, etc., by adjusting the brightness in advance, regardless of whether there is uneven brightness, A method is provided for performing high-sensitivity defect inspection with a threshold. In general terms, by providing a method for matching target features such as brightness, inspection can be performed with high sensitivity without being affected by mismatching of normal parts, and the occurrence of false alarms can be reduced.

By providing these methods, a highly sensitive defect inspection method that does not generate a false alarm even at a low threshold value for all inspection target areas is provided. In addition, it provides a method for classifying defects and a method for compressing image data. BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 is a diagram showing an example of an inspection target image and a detection waveform of brightness at that time.

Figure 2 shows an example of a conventional threshold setting method after alignment. is there.

FIG. 3 is a diagram showing an example of comparison results of images (a) and (b) of different brightness.

FIG. 4 is a diagram illustrating an image to be inspected when brightness unevenness occurs between the comparison chips and an example of a conventional threshold value setting method.

FIG. 5 is a block diagram showing an example of a schematic configuration of the inspection device. FIG. 6 is a plan view of a semiconductor wafer to be inspected.

FIG. 7 is a flowchart showing the order of defect extraction and classification in the image comparison unit.

FIG. 8 is an image diagram of pixels of a detected image showing an example of a contrast calculation method at a target pixel.

FIG. 9 is an image diagram of pixels of a detected image showing an example of a contrast calculation method at a target pixel.

FIG. 10 is a diagram showing an example of an approximate straight line obtained in a brightness relationship (scatter diagram) between a detected image and a reference image.

FIG. 11 is a diagram showing an example of a feature space as a set of a two-dimensional space for each category.

FIG. 12 is a diagram showing an example of a two-dimensional feature space in both (a) and (b). FIG. 13 is a diagram illustrating an example of a relationship between feature amounts between a detected image decomposed for each category and a reference image.

FIG. 14 is a diagram showing an example of the relationship between the feature amounts of the detected image and the reference image when the detected image is slightly shifted from the detected image and the reference image decomposed for each category.

FIG. 15 is a diagram illustrating an example of a category map, in which a detected image and a reference image are decomposed by contrast.

Fig. 16 shows (a) to (e), where the brightness It is a figure which shows an example of the behavior of fitting.

FIGS. 17 (a) to (c) are diagrams illustrating the operation of brightness correction on a scatter diagram. BEST MODE FOR CARRYING OUT THE INVENTION

Hereinafter, an embodiment of the present invention will be described in detail with reference to FIGS. 1 to 17.

As an example, a defect inspection method in an optical appearance inspection apparatus for a semiconductor wafer will be described as an example. FIG. 5 shows an example of the configuration of the device. 51 is a sample (inspected object such as a semiconductor wafer), 52 is a stage on which the sample 51 is mounted and moved, 53 is a detector, and a light source 501 and a light source for irradiating the sample 51 Illumination optical system 502, which collects light emitted from 501, Illuminates sample 51 with illumination light collected by illumination optical system 502, and forms an optical image obtained by reflection The objective lens 503 includes an image sensor 504 that converts the formed optical image into an image signal according to the brightness. Reference numeral 55 denotes an image processing unit, which detects a defect candidate on a wafer as a sample based on the image detected by the detection unit 53.

The image processing unit 55 converts an input signal from the image sensor 504 of the detection unit 53 into a digital signal.The AD conversion unit 504 performs shading correction, dark level correction, and the like on the AD-converted digital signal. Pre-processing unit 505 that performs image correction of the above, a delay memory 506 that stores the digital signal to be compared as a reference image signal, and a digital signal (detected image signal) detected by the detection unit 53 and a delay A displacement detection unit 507 that detects the amount of displacement from the reference image signal stored in the memory 506, creates a scatter diagram by voting (by plotting), decomposes it further, and further decomposes the scatter diagram. Scatterplot creator 508 b that calculates the correction coefficient based on Spatial information extraction unit 508 c that extracts spatial (position) information such as the spatial relationship (on the image of the pattern to be inspected), image comparison processing unit 508 a, coordinates and feature amounts of defect candidates (Area, size, etc.), detect the final defect, and classify the defect. The image comparison processing unit 508 a includes a positioning unit that performs pixel-by-pixel positioning of the inspection image and the reference image using the positional shift amount calculated from 507, and each pixel of the image after the positioning. (Brightness, contrast, color information, texture information, difference in brightness (shading difference) between the detected image and the reference image, or the frequency domain), and sends it to 508b A feature amount calculation unit, a registration correction unit that corrects registration between the two images using a correction coefficient obtained from 508b by scatter plot decomposition, and a correction image of the detected image and the reference image after the correction. An image comparison unit that calculates a difference at each corresponding pixel, and outputs a portion where the value of the difference calculated for each pixel is greater than a specific threshold value as a defect candidate; and a defect obtained based on the decomposition scatter diagram. Spatial information obtained from 508 c for the candidate (e.g., (Position information on the inspection pattern) and a spatial information for linking with the spatial information, and an image memory for temporarily storing an image.

Reference numeral 56 denotes an overall control unit, which has a display unit and an input unit that accept changes in inspection parameters (such as threshold values used for image comparison) from the user and display detected defect information. It comprises an interface unit 5110, a storage device 511 for storing feature amounts and images of detected defect candidates, and a CPU for performing various controls. Reference numeral 512 denotes a mechanical controller that drives the stage 52 based on a control command from the overall control unit. Although not shown, the image processing unit 55, the detection unit 53, and the like are also driven by a command from the overall control unit 56.

In the configuration shown in FIG. 5, the AD conversion unit 54 is provided on the image processing unit 55 side, and the detection unit 53 is provided with the image processing unit 55, the overall control unit 56, and the user interface. It may be configured to be installed separately from the evening space section 5110, the storage device 5111 and the like. In this case, the output from the image sensor 504 is A / D converted by the A / D converter 54, and the converted digital signal is input to the image processor 55 via the communication means, and the pre-processor 5 The processing after 05 is performed. In this case as well, the detection section 53 and the image processing section 55 are controlled by the overall control section 56.

As shown in FIG. 6, a semiconductor wafer 51 to be inspected has a large number of chips of a pattern that should be identical, which are regularly arranged. In the inspection device shown in FIG. 5, images are compared at the same position of two adjacent chips, for example, between the area 61 of FIG. 6 and the area 62 of the adjacent chip, and a difference between the two is obtained. Is detected as a defect.

Explaining the operation, in the overall control unit 56, the semiconductor wafer 51 as a sample is continuously moved by the stage 52 in, for example, the direction opposite to the scan A direction shown in FIG. In synchronization with the continuous movement of the stage 52, the optical image of the sample 51 is sequentially detected in the direction of scan A by the image sensor 504 of the detection unit 53, and the image of the chip is detected by the detection unit 5. Imported from 3. The image sensor 504 of the detection unit 53 outputs the input signal to the image processing unit 55.

In the image processing unit 55, first, the input analog signal is converted into a digital signal by the AD conversion unit 54, and shading correction and dark level correction are performed by the preprocessing unit 505. If necessary, perform processing to improve S / N by removing noise and enhancing edges. However, the image quality improvement processing by S / N improvement can be performed later. The position shift detecting section 507 has an image signal (detected image signal) of the chip to be inspected outputted from the preprocessing section 505 and a stage inputted from the delay memory 506, and the stage moves by the chip interval. Image signal delayed by an The chip image signal (reference image signal) is input as a set.

These two-chip image signals sequentially input in synchronization with the movement of the stage do not become signals at the corresponding locations when the stage vibrates or the wafer set on the stage is inclined. For this reason, the displacement detection unit 507 calculates the displacement between two images that are continuously input. At this time, the detected image signal and the reference image signal are continuously input, but the calculation of the positional deviation amount is performed sequentially for each processing unit with a specific length as one processing unit. It is important to select this length smaller than the period that affects the image, such as the stage or vibration of the optical system.

Here, the calculation target of the positional deviation amount may not be all the images but a part thereof, and the position may be determined by judging from the image of the first chip of the scan shown in FIG. If the behavior of the stage has some reproducibility, the amount of misalignment obtained in the first scan A may be referred to, and the amplitude of the subsequent misalignment calculation in scan B may be determined. Further, the displacement amount may be obtained by matching such as a normalized correlation of the image, or may be calculated in a frequency domain. In the latter case, even if attention is paid only to the phase, the mouth is bust due to the difference in brightness, which is preferable.

In the following processing, each processing unit is determined and performed for each unit. The image comparison processing unit 508 a aligns the images using the calculated positional shift amount, and creates an exploded scatter diagram described later in 508 b based on this information. Then, the detected image and the reference image are compared by the image comparing unit 508a, and an area where the difference value is larger than a specific threshold value is output as a defect candidate.

The feature extraction unit 509 performs editing such as deleting small defects as noise for each of the plurality of defect candidates, merging neighboring defect candidates as one defect, and so on. Area, size, other Calculate features for real-time ADC (Defect Classification) and output them as final defects. These pieces of information are stored in the storage device 5 11. Also, it is presented to the user via the user interface section 5 10. The feature quantity in 509 may be the scatter plot axis or the feature quantity used for decomposition. In that case, the defect judgment and the classification can be realized simultaneously in one shot.

Here, when the image comparison processing unit 508a obtains defect candidates from a mere difference value, not all of them are true defects. Examples are described below.

If the film thickness of the semiconductor wafer 51 is not uniform, a difference in brightness (luminance) occurs between the inspection target image and the reference image. For example, the three crosses 4a and 4b in Fig. 4 are the corresponding patterns in the inspection target image 11 and the reference image 12, but the brightness differs greatly due to the difference in film thickness. (Hereinafter, it is described as uneven brightness. In addition, only the detected image 11 has a defect 1c. FIG. 4 (a) shows that these images are correct by the misregistration detection unit 507. This is an image of the difference at each corresponding position when the positional shift amount is calculated and the alignment is performed by the alignment unit of 508a, but in the case of the same pattern, there is a part with uneven brightness. The value of the difference increases.

FIG. 4 (b) shows the waveform of the difference at position 1D-1D '. If an area having a difference value equal to or larger than the threshold value TH is regarded as a defect, a cross pattern in which the difference value increases due to uneven brightness is detected in addition to the defect l c. These are false information.

To avoid detection of false alarms due to such uneven brightness, increase the threshold from TH to TH2, and perform inspection with low sensitivity as a whole, or set the threshold to TH in areas with uneven brightness. Second, in areas where there is no uneven brightness, adjust the sensitivity using multiple thresholds, such as setting the threshold to TH. For example, there is a method of performing an inspection.

In the present invention, before the difference between the detected image and the reference image is calculated, the adjustment of the brightness between the images is corrected in advance by the adjustment correction unit of the image comparison processing unit 508a. Then, the difference between the two images is calculated in the image comparison unit using the detected image and the reference image in which the brightness adjustment between the images has been corrected.

FIG. 7 is an example of the outline of the processing. First, on the basis of the displacement amount calculated by the displacement detection unit 507, the registration unit 508a performs registration of the inspection image and the reference image in pixel units (70).

The feature amount of each pixel is calculated by the feature amount calculation unit of 508a for the image after the alignment (71), and the target image is decomposed into a plurality of pieces according to the feature amount in 508b (7). 2). As a result, a plurality of scatter plots, one kind of multidimensional space, are created in 508b. The group of pixels after decomposition is hereinafter referred to as a category (which may be referred to as a class). Furthermore, a category with a high frequency is detected, and this is regarded as a normal category.

Next, a correction coefficient for matching the brightness of the detected image and the reference image for each category is calculated using 508b while referring to the normal category (73). The fitting correction unit of 508a corrects the brightness of one image so as to be close to the brightness of the other image for each category using the above-described correction coefficient, and thereby adjusts the brightness. Correction is performed (74). Then, the image comparison unit of 508a calculates the difference between the corresponding pixel of the corrected detected image and the corresponding pixel of the reference image (75), and the one that is larger than the threshold calculated for each pixel is regarded as a defect candidate. Extract (76).

Finally, at the connection with the spatial information, as described later, for the mismatch (defect candidate) obtained from the image comparison unit based on the scatter diagram, the spatial information obtained from 508 c (for example, Position) A check is made as to whether or not it belongs to the match (defect candidate) (77), and finally a defect is output (78), and at the same time, a defect classification is performed, for example, at 509 (7.9). The defect classification may be performed at 508a based on the decomposition scatter diagram. Next, an example of a processing procedure for adjusting the brightness from 71 to 74 will be described in detail. Here, the brightness (selected feature amount) is corrected for one of the detected image and the reference image to be compared. First, the feature amount of each pixel is calculated using the detected image and the reference image that are aligned in pixel units. The feature amount includes statistics such as brightness, contrast, color information, texture information, brightness difference (shading difference) between the detected image and the reference image, variance of brightness, or feature amount in the frequency domain. Although there are a number of such cases, a case where contrast is used as a feature will be described below as an example.

First, the contrast is calculated for all pixels in the target area. There are various types of contrast operations in contrast operations, and one of them is a range filter. As shown in Fig. 8, the contrast at the pixel at the coordinate position (i, j) in the target area is defined as the difference between the maximum value and the minimum value of the brightness in the neighboring area. If the fill size is 2 × 2, if the brightness at (i, j) is A and the neighboring brightnesses are B, C, and D, the formula is (1). Of course, the size may be set according to the target, such as 3 × 3.

Contrast (i, j) = Max (A, B, C, D) — Min (A, B, C, D)

• (1) Also, depending on the image quality, a percentile fill that reduces the effect of noise may be used instead of a range fill. Also, the contrast at the pixel at the coordinate position (i, j) in the target area may be calculated by the second derivative. In this case, as shown in FIG. 9, the brightness A to I near 9 is used, and the word calculation by the equation (2) is completed. D x = B + H—2 XE,

D y = D + F- 2 xE

Contrast (i, j) = Max (Dx, Dy) · · · (2) In addition, various calculation methods can be used to determine the amount of luminance change in the neighborhood. In this way, the contrast F c (i, j) for each pixel in the detected image and the contrast G c (i, j) for each pixel in the reference image are calculated, and the correspondence between the detected image and the reference image is calculated. Equation (3), difference (4), or the larger one (5) is used for the pixels to be processed.

j

The contrast of each image is integrated, and the contrast at each pixel is uniquely determined. Then, the image is decomposed in several steps according to the contrast value C (i, j) c.

I do. Hereinafter, what is decomposed into several stages is referred to as a contrast category. As a result, the area where the brightness is uniform (low contrast area) like area 1a to the area where brightness j changes sharply like the pattern edge of area 1b (high contrast area) Are gradually decomposed.

C (i ₅ j) = (F c (i, j) + G c (i, j)) / 2

(3)

(4) c (j) = Max (F c (

(5) Next, a correction coefficient for adjusting brightness (selected feature amount) is calculated for each contrast category.

An example will be described with reference to FIG. First, for pixels belonging to the same contrast category, create a scatter plot with the horizontal axis representing the brightness of the detected image (the selected feature) and the vertical axis representing the brightness of the corresponding reference image (the selected feature). . Then, an approximate straight line is obtained from the scatter diagram. Where the frequency is small Since the category may be defective, substitute the frequently used normal category data. A linear approximation is performed by using the nearest neighbor decision rule or by using the data of the neighboring normal category or the data including the focused category and the neighboring normal category.

10 in FIG. 10 is an approximate straight line obtained from a scatter diagram of pixels belonging to a certain contrast category. There are various methods for calculating the approximate straight line. One example is a method of least squares approximation (a method of finding a straight line that minimizes the sum of distances from each point). Then, the calculated slope a and the Y intercept b of the approximate straight line are the correction coefficients of the contrast category. The brightness (the selected feature) of the detected image is corrected using the correction coefficient thus calculated, and the adjustment of the brightness (the selected feature) is performed.

Actually, assuming that the brightness of the detected image is F (i, j), the corrected detected image: F '(i, j) is calculated from the slope a of the approximate straight line and the Y intercept using equation (6). . Then, the difference between the corrected brightness F ′ (i ₅ j) of the detected image and the brightness G (i, j) of the reference image is obtained from the difference D (i, j) in equation (7), and the value of the difference Are larger than the set threshold TH as defect candidates.

F, (i, j) = a xF (i, j) + b · · · (6)

D (i, j) = F '(i, j)-G (i, j) · · (7) The correction of the brightness (selected features) of the detected image is scattered as shown in equation (6). Rotate the brightness (selected feature amount) of each pixel in the scatter diagram (rotation amount: gain), shift (shift amount: Offset). FIG. 17 shows the operation. Then, the difference value D (i, j) is the distance and circumference from the converted straight line. This means that the closer the distance to the straight line, the smaller the value of the difference after correction.

Also, the threshold value TH for defect detection is set outside the converted scatter plot. (Fig. 17 (b)). For this reason, in order to set the threshold TH low and perform highly sensitive inspections, it is necessary to make the converted scatter diagram slim as shown in Fig. 17 (c). Therefore, select a feature value that minimizes the spread of data on a scatter diagram (for example, the minimum variance).

Here, the reason why the scatter diagram can be reduced by the method shown in FIG. 7 will be described. In the case of an LSI wafer, the variation in film thickness occurs not only at the flat portion of the pattern but also at the edge portion. However, in bright-field detection, most of the specularly reflected light at the edge does not reach the image sensor. What is observed is mainly the diffracted light. Therefore, the influence of the film thickness variation is smaller at the edge portion than in the flat portion. Therefore, even when comparing two adjacent chips on the wafer, the influence of the mismatch due to the film thickness variation is small at the edge. For this reason, as shown in FIG. 11, if the scatter diagram of the two images is decomposed based on the contrast of the edge portion, a scatter diagram with small variations can be obtained in the edge portion. Therefore, if the threshold value is set appropriately, it is possible to detect even a minute defect at the edge portion, that is, a fine shape defect at the edge.

The threshold may be two values with a positive or negative sign, or may be an envelope (such as a polygonal line) that envelops the scatter plot. Further, if the gradation is converted based on the expression (6), a comparison with higher sensitivity can be realized.

In this way, if feature values are selected by focusing on the behavior of the diffracted light, contrast is one of the leading feature value candidates in bright-field detection. In addition, the contrast may be divided at equal intervals or unequal intervals, and each may be assigned to a different category. In Fig. 7, each pixel of the image is decomposed into contrast categories. This means that as shown in Fig. 11, scatter plots are prepared for the number of contrast categories.

In the actual machine, the contrast calculation type, fill size, number of divisions to the contrast category, step size, etc. are determined by the lookup table. By doing so, the configuration can be flexibly changed.

In addition, even if layer information based on CAD data that has the edge information equivalently is used instead of the pattern edge contrast, the spread can be suppressed and the scatterplot decomposition can be performed because of the separability. In this case, the area where the layers overlap should be regarded as another layer. Note that this idea does not reduce the feature value by selecting the axes as in the principal component analysis, but similar ideas can be applied.

In the case of DUV light or VUV light, a laser light source is mainly used. However, the present inventors use an objective lens as described in Japanese Patent Application Publication No. 2001-194432. The above idea holds for coaxial epi-illumination and bright-field detection

喑 In the case of visual field detection, the scattered light from the flat part is not detected and it is stable, but the brightness fluctuates due to slight differences in the shape of the pattern edge, etc. growing. Therefore, the point is how to reduce the scattered light from the pattern edge by blocking it with the Fourier transform surface of the sample. For this purpose, a light-shielding filter called a “spatial filter” corresponding to the frequency of the target pattern is inserted into the optical path to reduce the scattered light from the pattern. This can reduce the spread of the night on the scatter diagram. Therefore, a scatter diagram can also be used to evaluate the conformity of the spatial fill to the target pattern.

Features such as contrast differences, shading differences, brightness (information), texture information, frequency information on scatter diagrams, etc., are used to correctly decompose the scatter diagram using criteria such as the minimum variance criterion. Various features are used depending on the target and the detection method. In any case, if a sparse area with no data is secured on the disassembled scatter diagram, defects mapped to this area can be detected and inspection sensitivity will be improved. In other words, sparse The feature amount is selected so that the area can be secured. A sparse area is a category whose frequency is less than or equal to a defined threshold. The more this category, the better the defect detection sensitivity.

Here, an example using frequency (the number of pixels) will be further described. As a general feature, the color unevenness (normal area) is large in frequency due to its features such as widespread, repeated occurrence, and occurrence over the entire surface of a certain pattern. Of course, the normal part concentrates on the scatter diagram, so the frequency increases.

On the other hand, defects (abnormal regions) are less frequent. Even large defects are often scattered on the scatter plot, and the frequency of each category is small. Using this, the defect and the color unevenness are identified. Here, a category whose frequency is equal to or higher than the set threshold value is searched in the feature space, and this category is regarded as normal. Then, the distance from the normal category is added to the mismatch information, or the value itself is output. This distance may be a short distance or a Mahalanobis distance normalized by a covariance matrix. Normally, in the field of pattern recognition, as shown in Fig. 12 (a), the discrimination plane (hyperplane) that separates different categories (also referred to as classes) in consideration of the pause error probability is trained. Ask by

On the other hand, in the present embodiment, as shown in FIG. 12 (b), only the normal category (data on the right side of the figure) is correctly obtained from the training data, and the defect data is expressed as the distance from the normal category. The user can obtain the desired result by binarizing the distance. Depending on the threshold, it is a logic that prevents oversight (false alarms are likely to occur, but can be controlled by thresholds).

The surface for identifying the normal category may be a straight line or a curve (including a broken line approximation). Here, the normal pattern limit is given as frequency data, and in the case of a linear discriminator, the weight and bias are learned. What Note that the scatter diagram may store the normal range in the data table and compare it with this data table.

In the above, an example has been described in which the scatter diagram (image) is decomposed by contrast or the like to make the scatter diagram slim, but the brightness, color information, texture information, variance of brightness, etc. of the detected image and reference image have been described. Scatter plot (image) decomposition may be performed based on statistics or features in the frequency domain. In short, if the image is decomposed for each region having the same characteristic amount and the scatter diagram can be made slim, it is within the scope of the present invention.

In addition, the axis of the scatter diagram may be selected as the above feature or the calculation result thereof (for example, in the case of a feature amount of brightness, difference in brightness). In other words, as shown in Fig. 13, a scatter plot is created around the feature values selected in advance, and the adjustment of the target feature values is corrected to be affected by the mismatch of the normal part. Inspection can be performed with high sensitivity and the occurrence of false alarms can be reduced.

By the way, the feature amount is a comparison target.

When the feature value is brightness, brightness correction is performed, and when the feature value is contrast, contrast correction is performed. The same applies to other feature amounts. As shown in Fig. 14, a plurality of reference images whose positions are shifted little by little in the x and y directions, for example, in increments of 0.1 pixels, are created by techniques such as interpolation, etc. If you create a scatter diagram, decompose it, and select one with a slim data spread, a pair of aligned images is automatically selected, and you can perform alignment at the same time. Fig. 14 shows, for example, that a reference image is created by shifting the position of the detected image little by little, then a scatter diagram is created according to the respective positional shifts, the scatter diagram is decomposed, and then the combination is performed. FIG. 4 is a diagram showing that a slim scatter chart is selected in FIG. Using the above features, ifthen rules, fuzzy voting (voting), NN method (k-NN method), etc. It is also possible to classify defects by a pattern identification method. In this way, image alignment, defect judgment, and defect classification can be realized with one shot.o

Furthermore, creating or decomposing a scatter plot, which is one of the multidimensional spaces, means that the detected images are stored with greatly reduced capacity, and as a compression method for image data, It is suitable. Furthermore, it is effective in preventing the explosion of the scale of image processing hardware, which has been improved in speed and has become more complex with larger functions. The scatter diagram reduces the data volume by eliminating the spatial information, so the spatial information is kept to a minimum by selecting the features.

Also, an example has been described in which an image is decomposed using one feature amount and the brightness is adjusted, but the scatter diagram may be slimmed down from two or three or more feature amounts. The exploded scatter plot will have multi-dimensional axes. For example, the two axes of a scatterplot are brightness, which can be further broken down into four dimensions by contrast and brightness. The processing described in this embodiment is performed in this four-dimensional box.

Fig. 15 shows an example (category map) of scatter plot decomposition by contrast for pixels whose brightness falls within the set gradation range. The vertical axis divides categories by contrast, and the horizontal axis divides categories by brightness difference. It expresses the frequency for each category. The contrast here is given by a value that does not take the absolute value in equation (4). Multiple category maps are created according to the brightness category. Although it is described that the brightness is adjusted naturally, a feature amount other than the brightness may be adjusted.

Next, a description will be given of an embodiment in a connection section with the spatial information in the decomposition scatter diagram creation section 508 b, the spatial information extraction section 508 c, and the image comparison processing section 508 a. As a result of the above-described scatter diagram decomposition, the inconsistency output at 508a that is larger than the set threshold value on the scatter diagram is finally output as a defect candidate. One defect candidate tends to be scattered rather than concentrated in one place on the scatter plot. This is because the position in the feature space is determined by the defect and the background pattern (the position corresponding to the defect on the reference image side) where the defect is present, so that it is not always concentrated on the scatter diagram.

For example, the defect 1c in Fig. 3 (a) is concentrated on one place on the scatter diagram because the background is uniform, but if there is a defect 1d on the edge of pattern 4a, it is Since it is applied to both the uniform part and the edge part, it is scattered in at least two places on the exploded scatter diagram. For this reason, the inconsistency information (defect candidate information) dispersed on the scatter diagram also depends on the spatial proximity (on the image of the pattern to be inspected), which is its occurrence position, and the proximity (distance) in FIG. In the case of defect 1, since it corresponds to the position where defect 1d occurred on the image, checking the spatial distance (spatial information) at that point allows checking whether or not the point belongs to the same defect candidate, resulting in higher reliability. The degree of defect can be evaluated in degrees.

Conversely, even if it is determined that there is a mismatch (candidate defect), a certain spatial condition, for example, a spatial condition (spatial information) such as a maximum brightness, is satisfied. Can also be determined. In addition, the order statistics in the neighboring area (for example, in a 3 × 3 pixel, the value obtained by multiplying the brightness order assigned to each pixel by the brightness max-min: the brightness at each pixel It is also possible to judge a defect using a value close to a maximum or a minimum). As described above, the defect candidate may be determined using the order statistics in the local space. In each case, both the scatter diagram information and the space information on the image are used to determine whether a defect exists (77, 78 in Fig. 7).

As described above, in the present invention, two images are compared, and the difference In the inspection to detect defects from the scatter plot, comparison using scatter plot decomposition is performed and brightness adjustment is performed. FIG. 16 (d) is a waveform of the difference after brightness adjustment according to the present invention when there is uneven brightness in FIG. Brightness adjustment is performed, and the difference value decreases. Therefore, in the past, the threshold was set to TH2 for the entire area, or two thresholds, TH and TH2, were set to avoid the occurrence of false alarms. It is possible to avoid false alarms due to uneven brightness without dropping the light.

In addition, it is possible to perform a high-sensitivity inspection using only the low threshold value TH3 and easily adjust the sensitivity. Furthermore, in an optical visual inspection device using ordinary visible light as a light source, the detection sensitivity is limited to 100 nm, but according to the present invention, a detection sensitivity of 50 nm can be realized. is there.

As described above, the embodiment of the present invention has been described with reference to an example of a comparative inspection image in an optical appearance inspection apparatus for a semiconductor wafer, but an electron beam pattern inspection for detecting an image using an electron beam, and the like. It is also applicable to optical visual inspection using DUV light (ultraviolet light), VUV light (vacuum ultraviolet light), and EUV light (extreme ultraviolet light) as light sources. In this case, a detection sensitivity of 30 to 70 nm can be realized. In addition, the inspection target is not limited to a semiconductor device, but can be applied to, for example, a TFT substrate, a photomask, a print plate, and the like, as long as defects are detected by comparing images.

As described above, according to the present invention, by using information of a scatter diagram, which is one kind of multidimensional space, and comparison using decomposition information of a scatter diagram, high sensitivity can be obtained without being affected by mismatching of normal parts. Can be inspected. In addition, the occurrence of false alarms can be reduced by adjusting target features such as brightness. As a result, a low threshold can be set, and a high-sensitivity test can be stably realized.

In addition, since it is easy to simultaneously reduce the occurrence of false alarms and increase the sensitivity of defect detection, The detection sensitivity can be easily adjusted.

Furthermore, by applying the present invention to a comparative inspection in an optical appearance inspection apparatus, it is possible to stably maintain a detection sensitivity of 50 nm while maintaining a low false alarm rate.

In addition, by applying it to electron beam pattern inspection and appearance inspection using DUV as a light source, it is possible to stably maintain a detection sensitivity of 30 to 70 nm. Furthermore, the hardware scale of image processing can be reduced to a reasonable scale. Industrial applicability

The present invention relates to an appearance inspection for optically inspecting a defect generated during a process in a manufacturing process of a semiconductor wafer, a TFT, a photomask, and the like, and relates to a manufacturing process of a semiconductor wafer, a TFT, a photomask, and the like. It is used as a means to monitor the state of the process and maintain stable production.

Claims

The scope of the claims

1. A method for inspecting defects, including the following steps:

Imaging a sample to obtain an image to be inspected;

Imaging the sample to obtain a comparison target image to be compared with the inspection target image;

Obtain features such as brightness / contrast between the inspection target image and the comparison target image;

Voting the obtained feature amounts of the inspection target image and the comparison target image in a multidimensional feature amount space having a preset feature amount as an axis; and

The defect is detected by comparing the inspection target image with the comparison target image using the data voted in the multidimensional feature amount space.

2. The method for inspecting a defect according to claim 1, wherein the multi-dimensional feature amount space includes a predetermined number of contrasts such as brightness / contrast of the inspection target image and the comparison target image. It is a scatter diagram which consists of a feature quantity.

3. The method for inspecting a defect according to claim 2, wherein the scatter diagram composed of the predetermined plurality of feature amounts is decomposed into a plurality of scatter diagrams by other feature amounts, and the decomposed plurality of scatter diagrams are obtained. The defect is detected using the scatter plot of.

4. The method for inspecting a defect according to claim 3, wherein, using each of the decomposed scatter diagrams, matching is performed on a feature amount determined between the inspection target image and the comparison target image. .

5. The method for inspecting a defect according to claim 3, wherein each of the decomposed scatter diagrams is used to determine whether the scatter diagram or the frequency on each scatter diagram is normal. A category is obtained, and a distance such as a Euclidean distance or a Mahalanobis distance from a normal category is added to the mismatch information. ·

6. A method for inspecting defects according to claim 3, wherein a characteristic is selected so as to improve the separability of data on a scatter diagram, and the scatter diagram is decomposed.

7. The defect inspection method according to claim 3, wherein the defect is detected using both information obtained from the scatter diagram and spatial information obtained from the image.

8. A method for inspecting a defect according to claim 3, wherein feature values such as brightness are compared with each other, and an operation result such as those feature values or a difference between them is selected as an axis of the scatter diagram, and the defect is selected. Is detected.

9. The method for inspecting a defect according to claim 1, further comprising: classifying the detected defect using data voted in the multidimensional feature amount space.

10. The method for inspecting defects according to claim 9, wherein the classified defects are displayed on a screen separately for each classification.

11. The method for inspecting a defect according to claim 1, wherein the voting is performed in a multi-dimensional feature amount space having a feature amount defined as an axis such as brightness of the image to be inspected and an image to be compared and contrast. Thus, the capacity of the image data is reduced, and the defect is detected using the image data with the reduced capacity.

12. A method for inspecting a defect, comprising the steps of: imaging a sample to obtain an image to be inspected;; imaging the sample to obtain an image to be compared with the image to be inspected. Gain;

Determining a feature amount of each of a plurality of features for the inspection target image and the comparison target image;

Using the information on the obtained feature amounts to obtain a distribution of feature amounts for each of the plurality of features in the inspection target image and the comparison target image;

Correcting the difference in brightness between the image to be inspected and the image to be compared based on the obtained information on the distribution of the feature amounts for each of the plurality of features;

A defect is detected by comparing the inspection target image corrected for the difference in brightness with the comparison target image.

13. The method for inspecting a defect according to claim 12, further comprising: a step of correcting a positional shift between the inspection target image and the comparison target image in pixel units, wherein the positional deviation is corrected in pixel units. A feature amount for each of the plurality of features is obtained for the corrected inspection target image and the comparison target image.

14. The method for inspecting a defect according to claim 12, wherein the step of comparing the image to be inspected corrected for the difference in brightness and the image to be compared to detect a defect includes the step of inspecting the defect. A mismatch is detected by comparing the target image with the comparison target image, information on the distribution of the feature amounts of the plurality of detected mismatches is obtained, and based on the information on the obtained distribution of the feature amounts. To detect defects.

15. A method for inspecting a defect according to claim 12, wherein the inspection is performed. A defect is detected by comparing the target image with the comparison target image, and the detected defect is classified.