JP2015200658A - Defect inspection method and apparatus therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a defect inspection method and an apparatus therefor capable of detecting defects by discriminating the defects or a foreign matter generated on various circuit patterns to be repeatedly formed on inspection object substrates from normal circuit patterns, without depending on optical conditions.SOLUTION: A defect inspection method comprises the steps of: setting respectively a target local domain near the periphery of a target pixel where the periphery includes the target pixel, and each of a plurality of corresponding local domains near the periphery of a plurality of pixels corresponding to the target pixel where the periphery includes each of the plurality of pixels corresponding to the target pixel, against image signals acquired from inspection object substrates; searching for degree of similarity between an image signal of the set target local domain and the image signals of the plurality of corresponding local domains set in a local domain setting step; determining a plurality of image signals of the corresponding local domains similar to the image signal of the target local domain, by using searched similarity information; and comparing the determined image signals of the plurality of corresponding local domains and the image signal of the target local domain to determine defects in a comprehensive manner.

Description

The present invention relates to a method and apparatus for inspecting defects generated when manufacturing an FPD substrate such as an LSI, a liquid crystal display or a plasma display.

In Japanese Patent Laid-Open No. 2000-105203 (Patent Document 1), regarding the inspection of foreign matters or defects generated when manufacturing an LSI or a liquid crystal substrate, a repetitive pattern of circuit patterns existing on a substrate to be inspected (dense repetition) It is also described that a diffracted light pattern from a non-repetitive die is shielded by a spatial filter. Furthermore, in the defect determination process, the variation is calculated based on the signal originally obtained at or near the place where the same circuit pattern is formed, and the determination standard (threshold value) is set based on the calculated variation. However, it is described that a signal indicating a defect such as a foreign substance is extracted.

Further, in Japanese Patent Laid-Open No. 10-90192 (Patent Document 2), in order to search for a defect existing at an inspection point of the first pattern on the sample, at least one known inspection response of the second pattern of the same design is obtained. A defect inspection method and apparatus characterized by reference are described. In the inspection, it is important to use observation points that are equivalent to each other on the first and second patterns on the sample, and at least one search is performed to generate at least two inspection responses. The two responses (typically darkfield and brightfield response signals) are detected separately by photoelectric technique, compared separately and individually differential signals (between the first and second patterns) The technology to form is described. That is, the first and second responses in the first pattern are detected, and the result is compared between the two responses from the same inspection point in the corresponding second pattern, respectively, and the result As a result, a first difference signal and a second difference signal are formed.

The differential signals produced individually are processed in order to determine a list of first pattern defects in a centralized manner. Specifically, the first and second differential signals can be collectively processed to determine a unified defect list of the first pattern. Alternatively, data processing is subsequently performed on the first pattern defect list. Then, false defects found on the known and harmless sample surface are extracted and eliminated. On the other hand, such known harmless false defects are provided to the user for reference. In addition, various inspection searches are added to increase the inspection response so that more than one optical response is obtained from the sample for processing. This further improves the inspection accuracy. In addition, in the case of a transmissive sample, a photoelectric detector can be installed behind it to collect inspection responses due to transmitted light and improve the accuracy of the pattern defect list. It is also possible to search for defects buried in the surface.

JP 2000-105203 A JP-A-10-90192

T. Hiroi, S. Maeda, H. Kubota, K. Watanabe, Y. Nakagawa `` Precise Visual Inspection for LSI Wafer Patterns Using Subpixel Image Alignment '', Proc. Of 2nd IEEE workshop on Applications of Computer Vision (1994) K. Takeda, S. Kaneko, T. Tanaka, K. Sakai, S. Maeda, Y. Nakagaw `` Interpolation-based Absolute Gradient Matching and Its Application for LSI Wafer Inspection '', QCAV 2005, May18-20 (2005) Yayoyo Kita “Extracting changes between images using two-dimensional density histogram” IEICE Transactions Vol. J90-D, No. 8, pp. 1957-1965 Michi Sano, Takekazu Kato, Toshikazu Wada, Satoshi Sakai, Shunji Maeda “One Class Classifier Based on Pattern Proximity and Density Estimation”, MIRU2008, IS3-6 (2008) "Semiconductor wafer inspection technology using statistical threshold method according to background" by Satoshi Hamamatsu, Hisae Shibuya, Hidetoshi Nishiyama, Yoshimasa Oshima, Shunji Maeda, Satoshi Noguchi, View2004 (2004) Tetsuya Asami, Takekazu Kato, Toshikazu Wada, Satoshi Sakai, Shunji Maeda “Abnormality Detection Algorithm for Appearance Inspection Based on Saliency”, MIRU2008, IS5-2 (2008) Shunji Maeda, Satoshi Sakai, Takashi Okabe: LSI wafer thin-film multilayer pattern comparison inspection algorithm using scatter diagram information, Theory of Science, J88-D-II, 7 (2005) 1173.

When manufacturing an LSI or FPD substrate, as a pattern formed on a workpiece (for example, a semiconductor wafer), a repetitive pattern (dense repetitive) represented by a DRAM (Dynamic Random Access Memory) part, There are random patterns (non-repetitive patterns) as represented by logic. Then, a die that is a unit to be cut later to form one memory element or processor element is formed on the wafer substrate. Such LSI and FPD (
In the manufacture of a flat panel display) substrate, if foreign matter adheres to the surface of a workpiece or a defect occurs, it may cause a defect such as a wiring insulation failure or a short circuit. With the miniaturization of circuit patterns, it has become difficult to distinguish a pattern (non-defect portion) formed on an object to be processed from fine foreign matters and defects. For example, foreign matters and various defects ( Optical conditions (illumination wavelength, illumination polarization, illumination azimuth, illumination elevation angle, illumination beam numerical aperture, detection azimuth, detection elevation angle) that can detect wiring shorts, disconnections, pattern thinning, pattern thickening, scratches, hole non-apertures, etc. If the detected polarization and the detected numerical aperture are different for each defect, it is necessary to change the optical conditions for each defect type. Furthermore, it may be difficult to discriminate from non-defects under a single optical condition. In this case, it is necessary to inspect the same type of defect under a plurality of optical conditions, and the inspection speed ( There has been a case where the throughput has to be lowered.

Furthermore, due to variations in pattern roughness and film thickness, the difference between normal parts is relatively large compared to defects, making it difficult to set threshold values for defect determination even when optical conditions are selected appropriately. It is becoming. In particular, it is necessary to set a threshold value for each optical condition, and it takes time to create an inspection recipe.

An object of the present invention is to solve the above-mentioned problems, in manufacturing an LSI or FPD substrate, a defect or a foreign matter generated on various patterns formed on a workpiece is not dependent on optical conditions, and a normal circuit pattern. It is an object of the present invention to provide a defect inspection method and apparatus for inspecting defects by distinguishing them. In particular, in the inspection of semiconductor circuit patterns, the size of detection pixels is reduced for high sensitivity (this causes a reduction in throughput), the wafer diameter is increased from φ300 mm to φ450 mm, and the margin for exposure and etching is increased. With the limited inspection of the problem location represented by the hot spot, which is a shortage of location, a partial inspection is required instead of the entire wafer surface.

Accordingly, another object of the present invention is to provide a defect inspection method and apparatus capable of stably inspecting even when there are many optical conditions and there are few image signals to be compared.

In order to achieve the above object, the present invention acquires an image signal from a substrate to be inspected on which a circuit pattern repeated by a die is formed, for example, and exists on the substrate to be inspected based on the acquired image signal. A defect inspection method and apparatus for inspecting a defect to be performed, wherein a local region of interest to be cut out around the acquired image signal including a pixel of interest, and each of a plurality of corresponding pixels corresponding to the pixel of interest A local region setting step for setting each of a plurality of corresponding local regions to be cut out in the vicinity thereof, and between the corresponding local regions based on the image signal cut out from each of the corresponding local regions set in the local region setting step A forming step for forming a joint histogram feature space, and a threshold value is determined based on the joint histogram feature space formed in the forming step. Using the threshold value determination step and the threshold value determined in the threshold value determination step, the defect existing in the local region of interest is detected based on a difference image between the local region of interest and the corresponding local region And a defect detection step.

Further, the present invention is characterized in that, in the forming step, the corresponding local region is classified by a feature amount for each pixel, and a joint histogram feature space is formed for each category.

According to the present invention, in the forming step, the joint histogram feature space between the corresponding local regions is formed around a difference and brightness obtained from an image signal between the extracted corresponding local regions. Features.
In the threshold value determining step, the present invention calculates a convex hull from a joint histogram feature space between the corresponding local regions formed in the forming step, and determines the threshold value from the calculated convex hull. It is characterized by doing.

Further, according to the present invention, in the threshold value determining step, a maximum difference value (maximum grayscale difference) from the joint histogram feature space between the corresponding local regions formed in the forming step (
Diff) and the brightness (Bs) when the maximum value of the difference is calculated, and using the calculated maximum value (Diff) of the difference and the brightness (Bs) at that time, the local difference is calculated. The threshold value is determined based on brightness (Bf) obtained from an image signal cut out from the target local region set in the region setting step.

Further, the present invention is characterized in that in the threshold value determination step, threshold value addition in consideration of image collation accuracy is performed.
Further, the present invention further includes an image quality improvement step for restoring deterioration of the image signal acquired from the inspection target substrate using a point spread function representing deterioration of the optical system before the local region setting step. It is characterized by having.

Further, the present invention is a defect inspection method for acquiring an image signal from a substrate to be inspected on which a repeated circuit pattern is formed, and inspecting a defect present on the substrate to be inspected based on the acquired image signal. Each of a local region of interest that is cut out around the acquired image signal including the pixel of interest, and each of a plurality of corresponding local regions that are cut out around that includes a plurality of corresponding pixels corresponding to the pixel of interest. A local region setting step for setting a local area, and a search for searching for the similarity between the image signal of the local region of interest set in the local region setting step and the image signals of a plurality of corresponding local regions set in the local region setting step And a determination step for determining a plurality of image signals of the corresponding local area similar to the image signal of the local area of interest using the step and similarity information searched in the searching step And a defect determination step of comparing the image signals of the plurality of corresponding local regions determined in the determination step with the image signals of the local region of interest and comprehensively determining defects using multi-value logic or the like. It is characterized by.

In the defect inspection method, the present invention further forms a joint histogram feature space between the corresponding local regions based on the image signals cut out from the corresponding local regions set in the local region setting step. A threshold value determination step for determining a threshold value based on the joint histogram feature space formed in the formation step, and the threshold value determined in the threshold value determination step. A defect detection step of detecting the defect present in the local region of interest based on a difference image between the region and the corresponding local region.

According to the present invention, fine defects can be detected with high sensitivity without being affected by pattern roughness and film thickness unevenness in semiconductor wafers and FPD substrates that are becoming finer and larger in diameter. In particular, it becomes possible to detect a defect by discriminating it from a normal circuit pattern regardless of various optical conditions. Furthermore, even a small number of dies and a small number of data can be inspected, and furthermore, limited inspection of problem locations represented by hot spots in the die can be performed.

In addition, according to the present invention, since a pattern shape difference is expected for a die far from the target die, the die is limited to the vicinity, and unnecessary information is excluded. By defining many, it is possible to judge defects in a nearby area with similar meanings and properties, and comprehensively judge defects with high sensitivity, and also inspect systematic defects. Become.

It is the schematic which shows one Embodiment of the defect inspection apparatus which concerns on this invention, and the top view which shows a semiconductor wafer. It is a figure which shows the pixel projected on the to-be-inspected board | substrate, the linear sensor in the defect inspection apparatus which concerns on this invention, the illumination area on a to-be-inspected board | substrate. It is a flowchart which shows the process sequence which is 1st Embodiment of the signal processing part in the defect inspection apparatus which concerns on this invention. It is explanatory drawing of the example (method called Richardson-Lucy) of the image quality improvement performed in 1st Embodiment of the signal processing part in the defect inspection apparatus which concerns on this invention. In the first embodiment of the signal processing unit in the defect inspection apparatus according to the present invention, (I) image verification using three dies under a predetermined optical condition as a comparison target and (II) a target die having a target pixel It is a figure which shows the corresponding | compatible local area | region R set to the die | dye on both the next. In the first embodiment of the signal processing unit in the defect inspection apparatus according to the present invention, (III) is an explanatory diagram of a joint histogram feature space created between corresponding local regions, and (IV) is between corresponding local regions It is the 1 explanatory drawing of the procedure which determines a threshold value from the joint histogram feature space of. In the first embodiment of the signal processing unit in the defect inspection apparatus according to the present invention, each of (IV-I) and (IV-II) determines a threshold value from a joint histogram feature space between corresponding local regions. It is explanatory drawing of the 1st simplified version of a procedure. In the first embodiment of the signal processing unit in the defect inspection apparatus according to the present invention, each of (V), (VI), and (VII) is a second explanatory diagram of a procedure for determining a threshold value. It is explanatory drawing of the 1st process sequence in the case of the wafer edge die (incomplete die) and complete die performed in 2nd Embodiment of the signal processing part in the defect inspection apparatus which concerns on this invention. It is explanatory drawing of limitation of the test | inspection area | region performed in 2nd Embodiment of the signal processing part in the defect inspection apparatus which concerns on this invention. It is explanatory drawing of the 2nd process sequence of the in-die comparison system performed in 2nd Embodiment of the signal processing part in the defect inspection apparatus which concerns on this invention. It is explanatory drawing of the 3rd process sequence of the in-die comparison system performed in 2nd Embodiment of the signal processing part in the defect inspection apparatus which concerns on this invention. It is explanatory drawing in the case of setting an aperture shape only in the vicinity of a defect portion in order to detect the scattered light distribution of only the defect on the irregular pattern according to the present invention. It is a flowchart which shows the setting procedure of the optical condition in the defect inspection apparatus which concerns on this invention. It is explanatory drawing of the defect determination block corresponding to the some optical condition setting which concerns on this invention.

  Embodiments of a defect inspection method and apparatus according to the present invention will be described with reference to the drawings.

  In order to simplify the description, the same components are denoted by the same reference numerals.

FIG. 1A is a diagram showing a schematic configuration of a dark field defect inspection apparatus which is an embodiment of a defect inspection apparatus according to the present invention, and FIG. 1B is a plan view showing a semiconductor wafer. . The dark field defect inspection apparatus performs dark field illumination from a predetermined azimuth direction in a direction having an angle with respect to the normal direction of the semiconductor wafer W (hereinafter referred to as an elevation angle) on the surface of the semiconductor wafer (substrate to be inspected) W. A dark field illumination optical system (not shown) and a detection optical system 20. Dark field illumination optics (
(Not shown) may be configured to illuminate, for example, P-polarized light or S-polarized illumination light. The detection optical system 20 includes an objective lens 22 that collects reflected light (including scattered light and diffracted light) from a semiconductor wafer (substrate to be inspected) W that has been illuminated with dark field, and a space installed on a Fourier transform plane. Filter 28 and reflected light (including scattered light and diffracted light) that has passed through spatial filter 28
And a linear sensor 30 that receives and photoelectrically converts a light image formed by the imaging lens 29. The detection optical system 20 may include an analyzer that adjusts the rotation angle around the optical axis so that polarized light of a specific component can be detected by the linear sensor 30.

By the way, for example, the surface of the semiconductor wafer W scanned in the X direction is dark-field illuminated from a predetermined azimuth direction in an elevation angle direction having an angle with respect to the normal direction of the semiconductor wafer W by the illumination light 200. For example, a linear beam 300 whose longitudinal direction is in the Y direction is formed.
The light scattered or diffracted from the foreign matter, defect, or pattern on the semiconductor wafer by the linear beam 300 is, for example, from the direction parallel to the normal direction of the semiconductor wafer (that is, from above) from the objective lens 22.
It is condensed by. When the pattern formed on the semiconductor wafer W has a dense repetitive shape, the diffracted light generated from the repetitive pattern (a circuit pattern having periodicity) is regularly spaced from the exit pupil of the objective lens 22, that is, the Fourier transform plane. Since the light is condensed as a diffraction image, the light is shielded by the spatial filter 28 placed on the Fourier transform plane. Diffraction and scattered light from other than a repetitive pattern, that is, foreign matters, defects, and patterns on the semiconductor wafer W are reflected on the spatial filter 28.
, Is guided to the imaging lens 30, and is imaged on the linear sensor 30, for example.

FIG. 2A is a perspective view showing an embodiment of a linear sensor for imaging, and FIG. 2B is a diagram showing pixels of the linear sensor for imaging. As shown in FIG. 2A, the pixels 31 of the linear sensor 30 are arranged one-dimensionally. As shown in FIG. 2B, if the detection magnification is set so that the sensor pixel 53 projected on the semiconductor wafer W falls within the spot size range of the linear beam 300, the detection occurs from the linear region. Diffraction and scattered light can be detected together. The semiconductor wafer W is placed on an X, Y stage (not shown), and the semiconductor wafer W is scanned in the X and Y directions by the X, Y stage, whereby a two-dimensional image of the semiconductor wafer W is obtained. obtain. At this time, the main scanning direction is a direction perpendicular to the longitudinal direction of the linear beam 300, that is, X
The entire surface of the semiconductor wafer W can be inspected at high speed by moving in the Y direction in steps in the Y direction by the length of the sensor pixel 53 projected onto the semiconductor wafer W. The diffracted and scattered light received by the linear sensor 30 is photoelectrically converted and sent as an image signal 400 to the signal processing unit 100 where the signal processing characteristic of the present invention is executed.

Next, a first embodiment of the signal processing unit 100 that characterizes the present invention will be described. The first embodiment of the signal processing unit is configured by a computer having a memory (storage device), and a first image signal acquired from the target die and an adjacent die (adjacent to the target die) (
A comparison test between the actual objects with the repetitive circuit pattern) as a comparison target is executed. That is, as shown in FIG. 3, the signal processing unit 100 (1) improves the image quality by restoring the optical system and stores it in the image memory (not shown) according to the program. (S31). Next, image collation (image alignment) is performed between the target die cut out from the image memory and adjacent dies (S32), and then (2) the target local region is included in the vicinity of the target die including the target pixel. The corresponding local region R is set in the vicinity of the surrounding die including the corresponding pixel at the position of a plurality of corresponding pixels corresponding to the target pixel in the adjacent die or the target die (S33), (3
-5) For each of the set corresponding local regions R, the pixels are classified by the feature amount (each pixel is classified into categories) (S34), and a joint histogram feature space is created for each category ( S35), a threshold value (Th1) is calculated in the created joint histogram feature space (S36), (6) threshold addition is performed in consideration of image matching accuracy (S37), and the target local area of the target die is determined. In comparison between the image signal of the region and the image signals of the corresponding local regions of the adjacent die or the target die, the defect 1001 is detected by binarization using the threshold value (Th1) (
S38) The procedure is executed.

The signal processing unit 100 is further connected to an overall control unit 101 having an external input / output system 102 so that parameters and the like can be set and changed from the outside. The overall control unit 101 is connected to an external device 103 such as a mouse, a keyboard, or a printer. Therefore, part of the signal processing in the signal processing unit may be executed by the overall control unit 101.

Next, embodiments of signal processing that are characteristic of the present invention and are executed in the signal processing unit 100 will be described in sequence with reference to FIGS. Note that the image signal obtained by the embodiment shown in FIG.

The first image quality improvement (S31) is to correct (improve) the image quality of the acquired image signal. FIG. 4 shows an example of processing. Point spread function (deterioration of optical system)
The image degradation is restored using a point spread function. For restoration, Richardson-Lucy method using Bayes' theorem, Iterative Back Projection (IBP) method, Bilateral Fil
There is a Total Variation (BTV) method using ter, and FIG. 4 shows the result by the Richardson-Lucy method. When using this restored image to determine the defect, it was confirmed that the defect signal intensity of the defect that shines in a dot-like manner in the dark-field image is particularly improved, and also has the effect of noise removal, so that the defect detection sensitivity is improved. did.

Next, as shown in FIG. 5 (I), a target die (46b) is selected as a comparison target under predetermined optical conditions.
And a die located in the vicinity of the target die (46b), for example, adjacent dies (46a on both right and left sides)
46c). Of course, adjacent dies may be arranged side by side instead of left and right. Here, three dies (46a, 46b, 46c) including a target die and left and right adjacent dies are set as comparison targets. The image signals of the three dies are represented by f1 (x, y), f2 (x, y),
Let f3 (x, y). In step S32, for these image signals, first, AG
Image matching is performed by M (Absolute Gradient Matching), NCC (Normal Cross Correlation), etc., and image signals are aligned. For example, T. Hiroi, S. Maeda, H.
Kubota, K. Watanabe, Y. Nakagawa `` Precise Visual Inspection for LSI Wafer Patterns U
sing Subpixel Image Alignment '', Proc. of 2nd IEEE workshop on Applications of C
The position of the image signal can be adjusted with subpixel accuracy by the method described in omputer Vision (1994) (Non-Patent Document 1). Also, K. Takeda, S. Kaneko, T. Tanaka, K. Sakai
, S.Maeda, Y.Nakagawa `` Interpolation-based Absolute Gradient Matching and Its Appl
The method described in “ication for LSI Wafer Inspection”, QCAV 2005, May18-20 (2005) (Non-Patent Document 2) is also applicable.

Next, in step S33, as shown in FIG. 5 (II), it is considered to determine a defect in the local region of interest near the periphery of the pixel of interest. Next, for example, for image signals f1 and f3 of adjacent dies. A corresponding local region R is set near the periphery of the corresponding pixel corresponding to the target pixel. The corresponding local region R includes, for example, corresponding pixels corresponding to the target pixel in the image signal f2 of the target die with respect to the image signals f1 and f3 of the adjacent die adjacent to the target die, and is set in the vicinity of the corresponding pixel. The size of the corresponding local region R depends on the technology node of the target pattern, the optical magnification, and the detection pixel size, but various settings such as 7 × 7 pixels, 13 × 13 pixels, and 33 × 33 pixels are possible.

The focus here is to limit the target data and not include much data. Inclusion of a large amount of data has the advantage of reflecting variability between dies, but also leads to having excessive information other than the target die to be compared. A die is limited to a neighboring die (for example, an adjacent die), and a region in the die is also set as a corresponding local region R. For example, the dies are, for example, dies on both sides so that the target die can be pinched. Of course, it may be on the target die so that the target local region is pinched and shot against a plurality of corresponding local regions R. It is assumed that the corresponding local region R does not include a defect (this assumption is established with high probability). Even if it contains a defect, there is no substantial adverse effect.

Next, in steps S34 and S35, as shown in FIG. 6 (III), each of the corresponding local regions is subjected to LIM (Shunji Maeda, Satoshi Sakai, Takashi Okabe: Scattering). LSI wafer thin-film multi-layer pattern comparison inspection algorithm using figure information, theory of science, J88-D-II, 7 (2005) 1173. (Method described in Non-Patent Document 7))
And categorize to form a joint histogram feature space. The joint histogram feature space is a scatter diagram, and the joint histogram feature space of the local region is, for example, the difference and brightness obtained from the image signals f1 (x, y) and f3 (x, y) of the corresponding local region (
It is a scatter diagram of data formed with gradation values f1, f3) as axes. In FIG. 6, as a difference, f
The absolute values of 1-f3 and f1-f3 are shown by changing the size of the local region. The axis increments may be grouped in a certain unit, not a gradation unit. It can be seen that if the size of the corresponding local region R is too small, the data tendency cannot be grasped. It is desirable to set the size of the corresponding local region so that the boundary of data becomes clear to some extent. The joint histogram feature space often refers to a two-dimensional density histogram, but may be a two-dimensional or more multi-dimensional.

Furthermore, when normalized by the total frequency, it can be treated as a probability density function. Considering the joint histogram feature space as a probability density function, entropy, mutual information, Cullback liber information, or Cullback liber divergence (Kullback-Leibler div
ergence) and other concepts can be applied.

In general, the magnitude of the variation of the probability density distribution can be evaluated by the information amount (entropy) of the distribution. The entropy is maximized when the variation is a uniform distribution, and the entropy is minimized when the distribution is concentrated only at a certain position. When the two images are A and B, the information amount H (A, B) (joint entropy) calculated from the joint histogram of these images is the maximum H (A ) + H (B)
It becomes smaller as the correlation becomes higher.

The Cullback-Liber information amount is a measure of the difference between two probability distributions (true probability distribution P and arbitrary probability distribution Q) in probability theory and information theory. Information divergence (Informatio
n divergence, Information gain, Relative entropy
).

In addition, it is more preferable that the joint histogram feature space of the corresponding local region R is formed for each pixel having similar features by dividing each pixel by the feature amount. Here, existing features such as brightness, differential value, standard deviation, etc. can be used.

When a joint histogram feature space is formed around the brightness of the corresponding pixel obtained from the image signals of the plurality of corresponding local regions R that are cut out, the joint histogram feature space is regarded as a probability density function, and the mutual information is calculated as a numerical value. It can also be determined how similar the image signals in the local region are. Of course, positional deviation between image signals can also be evaluated. That is, the joint histogram feature space represents the existence probability of the corresponding pixel between the image signals, and the distribution changes due to the deformation between the image signals. It is also possible to perform registration between image signals by modifying the image signals so that this distribution is small. The accuracy of the above-described image matching method can also be evaluated based on the mutual information amount.

In addition, the joint histogram is Yayoyo Kita, “Extraction of change between images using two-dimensional density histogram”, IEICE Transactions Vol. J90-D, No. 8, pp. 1957-1965 (Non-patent Document 3).
There is a description.

Next, in step S36, as shown in FIG. 6 (IV), in the joint histogram feature space, the relationship between the brightness (f1, f3) and the difference (| f1-f3 |) is a convex hull (envelope). By expressing. Then, the difference value (| f1-f3 |) for each brightness (f1, f3) is adopted as a threshold value (Th1) for determining a defect in the local region of interest.

The focus here is to actively determine the boundary without determining the variation of the data. Therefore,
・ Determine the envelope (convex hull, R-concave hull),
Apply a local subspace method to the feature space (eg, centroid near k),
・ Categorically connect points with a small number of frequencies (for example, frequency 1),
The boundary is obtained by such a method.

In addition, the boundary may be determined using a generation type discriminator, or the boundary may be set using an identification type discriminator.

Here, the convex hull refers to a convex polygon including all points of the point set and all vertices of the polygon and having the smallest area. For the calculation of the convex hull, the packaging method, Graham scanning method, sequential construction method, divide-and-conquer method, etc. are known. The above-mentioned R-concaves are Michi Sano, Takekazu Kato, Toshikazu Wada, Satoshi Sakai, Shunji Maeda “One Class Classifier Based on Pattern Proximity and Density Estimation”, MIRU2008, IS3-6 (
2008) (Non-Patent Document 4), which is a method based on a convex hull.

The centroid in the vicinity of k means that regarding the maximum value of the difference, those centroids are set as threshold values for k pieces of data around the centroid. This method is effective when the pixels in the corresponding local region cannot be assumed to be normal. The parameter k is set to 3, for example. As with the R-concave hull, it is possible to add a margin to the threshold according to the density of data near the maximum value.

If the comparison method is based on the difference between dies, the difference and brightness obtained from the image signal are appropriate for the axis of the joint histogram feature space of the corresponding local region, and it is not based on the difference between dies. For example, in the case of a method based on the difference between the differential values, the difference and the differential value obtained by differentiating the image signal are appropriate for the axis of the joint histogram feature space of the corresponding local region. The axis of the joint histogram feature space is determined depending on what is compared to detect the defect.

By the method described above, a threshold value (Th1) for defect determination in the local region of interest is calculated.

Next, a simplified version of threshold calculation will be described with reference to FIGS. 7 (IV-I) and (IV-II). That is, in the joint histogram feature space of the corresponding local region R, linearity is assumed, and first, a maximum absolute value of differences (maximum density difference) (Diff) is calculated. Next, the threshold value Th1 is set according to the following (Equation 1) or (Equation 2).
Th1 = Diff × Bf / Bs (Equation 1)
Bf indicates the brightness of the adjacent dies (f1, f3) that produce the maximum contrast difference Diff, and Bs
Indicates the brightness of the pixel of interest (position of f2).
Th1 = Diff × f2 (x, y) / (max (f1 (x, y), f3 (x, y)))
Th1 = Diff × f2 (x, y) / (f1 (x, y) + f3 (x, y)) / 2 (Equation 2)
If the boundary data has linearity in the difference and brightness obtained from the image signal, it means that the calculation is simplified.

Next, in step S37, finally, as shown in FIG.
Considering the remaining misalignment with respect to h1 (in consideration of image matching accuracy), the threshold value Th
Let 2 be the following (Equation 3).
Th2 = Th1 + k2 × grad (f2 (x, y)) (Equation 3)
Here, grad (f2 (x, y)) is a differential value of the image signal. k2 is a weight parameter, for example, a value of about 0.3.

Further, as shown in FIG. 8 (VI), the threshold value is propagated to the vicinity of the surrounding eight pixels, for example, by one pixel to obtain the final threshold value.

As shown in FIG. 8 (VII), the die near the outer periphery of the wafer (edge die) has a poor pattern, and the brightness variation σ of the corresponding local region R for the purpose of allowing this to some extent.
And the threshold value is corrected with this variation (standard deviation or the like) σ. That is, in the edge die, the threshold value is corrected in consideration of the edge factor (multiplying the σ ratio).

As described above, in the first embodiment of the signal processing unit 100, which is a feature of the present invention,
The inspection using three dies (f1, f2, and f3) has been described. This indicates that the periphery of a wafer with a small number of dies can be inspected reliably. As a conventional example, Hamamatsu
Tsuji, Hisae Shibuya, Hidetoshi Nishiyama, Yoshimasa Oshima, Shunji Maeda, Satoshi Noguchi, “Semiconductor wafer inspection technology using statistical threshold method according to background”, View 2004 (2004) (non-patent document 5) Yes, in order to stably determine the threshold value, it is necessary to acquire brightness information from many dies. However, in the first embodiment, the inspection is surely performed even if the number of dies is small. As shown in FIG. 9, the edge die around the wafer and the die adjacent thereto can also be inspected. That is, even an incomplete die in which the circuit pattern is not completely formed can inspect a region where the circuit pattern is formed and a die adjacent to the incomplete die.

That is, in the second embodiment of the signal processing unit 100 that is a feature of the present invention, the first described below.
The processing procedure is taken. In FIG.
(S41) Search for similar corresponding local regions in adjacent dies, for example, sequentially in the image in units of pixels,
In each case, find the similarity,
(S42) Rearrange based on the obtained similarity,
(S43) In the adjacent die, a corresponding local region that is most similar to the target pixel (target local region) is determined,
(S44) The target pixel (target local region) and the similar local region Ω are sequentially compared.
The comparison is performed through, for example, S34 to S38 in FIG. Variations in the procedure are possible.

The similarity is the sum of absolute values of differences between pixels in the local region, the root of the sum of squares, or the correlation coefficient. A concept called saliency can be applied. For example, Tetsuya Asami, Takekazu Kato, Toshikazu Wada, Satoshi Sakai, Shunji Maeda “Abnormality Detection Algorithm for Visual Inspection Based on Saliency”, MIRU2008, IS5-2 ( (2008) (Non-Patent Document 6) describes the saliency, and this method can be applied. The mutual information amount described above can also be used.

It should be noted that, when calculating the degree of similarity, it is possible to simultaneously detect the positional deviation of the image, but if it is performed at the sub-pixel level using interpolation or the like, the sensitivity is further improved.

If the degree of similarity is low and the degree of similarity is lower than the set threshold value, the comparison inspection is impossible and the data is output as uninspected. The threshold value is determined in advance. If unchecked, the pixel is analyzed in detail later. For example, processing is performed in which inspection results for each die are overlapped and the frequency thereof is checked.

In addition, as shown in FIG. 10, even when it is desired to perform a limited inspection of a part of the wafer, in the second embodiment, it is not necessary to inspect many dies, and only the periphery of the die to be inspected is targeted. If so, there is an advantage that it can be realized. For example, inspecting a wafer in a cross shape,
It is also possible to inspect the upper and lower five points on the wafer. For example, Non-Patent Document 6 described above
Then, for one swath (obtained with an image sensor while scanning the stage in one direction,
Although a threshold is determined for a plurality of dies in a horizontal row covering both ends of a wafer), it is necessary to acquire a horizontal row image covering both ends of the wafer even when some dies are inspected. In the second embodiment, no matter how the inspection area is limited, inspection can be performed on a minimum number of dies, and throughput does not decrease. Furthermore, limited inspection within a die can be performed for a problem portion represented by a hot spot, which is a portion where a margin at the time of exposure or etching is insufficient. That is, information for determining the threshold value regarding the limited area may be obtained from the minimum die, and unnecessary image detection and stage movement need not be performed.

Next, a second processing procedure in the second embodiment of the signal processing unit 100, which is a feature of the present invention, will be described with reference to FIG. Second in the second embodiment of the signal processing unit 100
This processing procedure is another embodiment of the in-die comparison, and is as follows.
(S51) A similar corresponding local region is searched with an adjacent die,
(S52) sort based on the similarity,
(S53) In the adjacent die, the corresponding local region R1, which is similar to the target pixel (target local region),
Determine R2, R3, ...
(S54) The target pixel (target local region) is compared with the corresponding local regions R1, R2, and R3,
(S55) The defect is determined by multi-value logic such as majority vote.

According to the second processing procedure, one-to-many comparison and multi-valued logic can cope with systematic defects common to each die (defects caused by the exposure apparatus, etc.), and many counterparts. Therefore, even if there is a roughness of the pattern, the inspection can be performed with high sensitivity. Here, the multi-value logic may be a majority logic or a logic determination with a threshold value.
Conventionally, it is a one-to-one comparison, and such a one-to-multiple comparison has not been made. The number of dies is limited, and the variation originally required for die comparison is obtained and unnecessary information is eliminated, but in the neighboring dies, many comparison partners to be compared are defined, and in a sense, nearby areas with similar properties The defect determination is performed on the object, and a highly sensitive defect extraction is performed by comprehensive determination. This concept of one-to-multiple comparison is considered advanced in the sense that it appropriately handles the information necessary for comparison.

One-to-many specific comparison logic is
(1) Binary logic such as determining a defect if at least one mismatch occurs, determining a defect if a number of mismatches greater than or equal to a threshold value occurs, and determining a defect if all do not match,
(2) In the state of the difference image before the defect determination, the difference is added to determine a defect if the difference is equal to or greater than a threshold value, and if the minimum value of the difference is detected to be equal to or greater than the threshold value, the defect is determined Then, a multi-value logic such as determining a defect when the maximum value of the difference is detected and becomes equal to or greater than a threshold value can be considered. Here, each difference value may be weighted according to the degree of similarity. Moreover,
(3) A plurality of difference signals can be handled as multidimensional data. In other words, the defect is determined based on the length and direction of the vector as a vector.

Note that if the processing steps S51 to S54 are executed before the inspection, the inspection throughput will not be reduced. Furthermore, if the adjacent die is an adjacent die on the wafer inner side, at most 2
A die can be inspected.

In the second processing procedure, if the comparison within the same die may be included, the variation in the pattern is small, and the inspection with higher sensitivity can be realized. FIG. 12 shows a third processing procedure. That is, the third processing procedure executed by the signal processing unit 100 is as follows.
(S61) Search for similar corresponding local regions in the adjacent die and the target die,
(S62) sort based on the similarity,
(S63) In the adjacent die and the target die, corresponding local regions R1, R2, R3,... Similar to the target pixel (target local region) are determined,
(S64) The target pixel (target local region) is compared with the corresponding local regions R1, R2, and R3,
(S65) The defect is determined by multi-value logic such as majority vote.

The specific comparison logic of one-to-multiple is the same as that described above, but in the comparison within the target die,
It is expected that the similarity between local regions is high. For this reason, for example, when the weight is set, the weight for the difference is large.

The first and second embodiments in the signal processing unit 100 described above are all directed to the dark field inspection image shown in FIG. Optical conditions (illumination wavelength, illumination polarization, illumination azimuth, illumination elevation angle, numerical aperture of illumination beam) that can detect foreign objects and various defects (such as wiring shorts, disconnection, pattern thinning, pattern thickening, scratches, hole non-openings, etc.) , Detection azimuth, detection elevation angle, detection polarization, detection numerical aperture, etc.) are different for each defect, so it is necessary to change the optical conditions for each defect type. The second embodiment is applicable.

Next, an example of a technique for optimizing various optical conditions will be described. FIG. 13 shows a plan view of the wafer. Defect 4 is a defect to be detected. The case where the defect 4 exists in a non-periodic pattern area | region is shown. In order to optimize an aperiodic spatial filter (not shown), it is necessary to grasp the scattered light distribution of defects. Since the illumination light is illuminated over a wide area, defect 4
In order to detect only the scattered light distribution, it is necessary to set the aperture shape 26 only in the vicinity of the defect portion by a field stop (not shown) arranged in the detection optical system 20, as shown in FIG. The size of the opening changes according to the size of the defect and the positional relationship with the peripheral pattern. In the LSI pattern, the size of the aperture of this diaphragm is that the size of one side is a semiconductor wafer (substrate to be inspected) W
About 1 to 10 mm above is an appropriate size (in the case of a circular opening, it corresponds to the outer diameter).
The scattered light distribution detected by an observation camera (not shown) that observes the Fourier transform image formed on the Fourier transform plane by the field stop of this detection optical system is mainly the scattered light distribution from the defect. By setting the aperture of the aperiodic spatial filter where the brightness of the light is high,
It becomes possible to detect scattered light from the defect and suppress scattered light from the normal pattern.

Next, the procedure for optimizing and setting the optical conditions of the defect to be detected will be described with reference to FIG. That is, the optical condition to be tried first is selected in the GUI (Graphic User Interface) of the overall control unit 101 (S81). Next, in the GUI of the overall control unit 101, the aperture shape of the field stop of the detection optical system 20 is set near the defect based on the screen shown in FIG. 13 (S8).
2). Next, fixed optical conditions (a wavelength band of illumination light, an illumination light amount, an imaging magnification of the imaging lens 29, etc.) are set in the GUI (S83). Next, the optical condition to be tried is set in the GUI (S84).

As trial condition candidates, for example,
1. Illumination optical system (1-1) Elevation angle of illumination light 200 (angle formed by optical axis of illumination light from wafer surface)
(1-2) Illumination direction of the illumination light 200 (for example, an angle formed with respect to the notch direction of the wafer or an angle formed with reference to the traveling direction of the X stage of the apparatus)
(1-3) Polarization of illumination light 200 (P-polarized light, S-polarized light, etc.)
(1-4) NA of illumination light 200
2. Detection optical system (20)
(2-1) Rotating angle of analyzer (not shown) for detecting polarized light of specific component (2-2) Opening shape of spatial filter 28 and birefringent material (selected after the above conditions are determined)
and so on.

While appropriately changing these conditions for each defect, optimization is performed in the following loop according to various defects.

That is, the dark field image of the defect portion observed by the first observation camera (not shown) of the detection optical system 20 and the Fourier transform image of the defect portion observed by the second observation camera (not shown). Obtained and stored in the image memory of the signal processing unit 100 (S85). Next, the semiconductor wafer W is placed on the adjacent die.
Is moved (S86). Even in this adjacent die, a dark field image of the normal part observed by the first observation camera of the detection optical system 20 and a Fourier transform image of the normal part observed by the second observation camera are acquired and a signal processing unit is obtained. The image is stored in the image memory 100 (S87). Then, the signal processing unit 10
0 calculates a difference image between the dark field image and Fourier transform image of these defective portions and the dark field image and Fourier transform image of the normal portion of the adjacent die, and stores them in the image memory of the signal processing unit 100 (S88).
). In step S89, the signal processing unit 100 or the overall control unit 101 determines whether or not the trial optical condition has ended, and repeats this if it has not ended. The first and second observation cameras are configured so that the optical path can be switched so that the dark field image and the Fourier transform image can be observed when the optical condition is set.

Next, the signal processing unit 100 or the overall control unit 101 determines whether the scattered light distribution of the defective part is normal based on the difference image between the dark field image of the defective part and the dark field image of the normal part under various trial optical conditions. The optical conditions with different distributions are calculated by comparing with the scattered light distribution of the part, and the optical conditions are output to the GUI or the like (
S90). At this time, the signal processing unit 100 or the overall control unit 101 also utilizes a difference image between the Fourier transform image of the defective part and the Fourier transform image of the normal part, and the like, so that the optical conditions differ in the scattered light distribution between the defective part and the normal part. And the optical condition is output to a GUI or the like. Note that the difference in the scattered light distribution is an optical condition in which the scattered light from the defective portion is relatively strong relative to the normal portion, or an optical condition in which the scattered light from the defective portion is present in an optical condition in which the scattered light from the normal portion is not emitted. Etc. Thereby, the signal processing unit 100 or the overall control unit 101 grasps an optical condition that allows more scattered light from the defective part and can be detected efficiently, and selects this (S91). There may be a plurality of selection conditions instead of one.

By these optical condition setting procedures (S81 to S91), it is possible to grasp the optical conditions that can detect more scattered light from the defective portion and the position on the Fourier transform plane where the scattered light from the defective portion is generated. Next, a region that overlaps the scattered light distribution from the opening of the spatial filter 28 and the defect portion on the Fourier transform plane is calculated, and an optical condition having a wide region is selected (S92).

As described above, the signal processing unit 100 or the overall control unit 101 finishes narrowing down the optical conditions (S93). When there are a plurality of narrowed-down conditions, test inspection is performed under the narrowed-down conditions, and a condition with high defect detection and normal pattern classification performance is selected from the result of the test inspection, and the setting of the optical conditions is completed.

Note that various combinations of the configurations, functions, and contents of the conditions described in the above embodiments are conceivable. Further, in order to detect various defects at high speed, it is necessary to simultaneously detect images under different optical conditions. In this case, the setting of each optical condition is the same procedure.

Further, when an image is detected under a plurality of optical conditions, the amount of the image becomes enormous according to the number of image sensors. In this embodiment, as shown in FIG. S35), threshold generation (S36), defect determination by comparison (S37, S)
38) is limited as much as possible to the target pixel (local area R of the die adjacent to the target pixel shown in FIG. 5). In this sense, even in the case of a plurality of optical conditions A, B, and C, the process is performed. There is an advantage that it does not become large-scale. That is, in the signal processing unit 100, each block 1001a (S33 to S35), 1001b (S36), 1001c (S37, S38) of image processing.
) Requires a small amount of memory capacity, and can be realized at low cost. Therefore, a programmable FPG, not to mention DSP (Digital Signal Processor).
A (Field Programmable Gate Array) has an advantage that it is easy to configure. That is, it can be said that the method of the present invention is suitable for simultaneous image detection and defect determination under a plurality of optical conditions using a plurality of image sensors.

In FIG. 1, the illumination light 200 is a laser, but a lamp may be used. Also,
The light having the wavelength of the illumination light source may be a short wavelength, or may be a light having a broad wavelength (white light). If the wavelength band is appropriately selected, further improvement in defect detection sensitivity can be expected. When using short-wavelength light, in order to increase the resolution of the image to be detected (to detect minute defects)
Light with a wavelength in the ultraviolet region (Ultra Violet Light: UV light) can also be used. When a laser is used as the light source, if it is a single wavelength laser, it is possible to provide means for reducing coherence (not shown) (temporal and spatial low interference). The image sensor 30 employs a time delay integration type image sensor (TDI (Time Delay Integration) image sensor) configured by two-dimensionally arranging a plurality of one-dimensional image sensors, and is synchronized with the movement of the stage. By transferring and adding the signals detected by the one-dimensional image sensor to the next-stage one-dimensional image sensor, it becomes possible to obtain a two-dimensional image with high sensitivity. This T
By using a parallel output type sensor having a plurality of output taps as the DI image sensor, the output from the sensor can be processed in parallel, and higher-speed detection becomes possible. Of course, a CMOS linear sensor may be used. In the case of CMOS, the dynamic range can be expanded.

The application object of the present invention can be applied to an image of a bright field inspection in addition to an image of a dark field inspection. In the case of bright field inspection, the illumination light source may be a laser or a lamp. Also,
Needless to say, the present invention can also be applied to an electron beam inspection image. In either case, the threshold value can be determined from a small number of data without being affected by the difference in the optical system, and a highly sensitive defect determination can be performed. In addition, by one-to-many comparison and multi-valued logic, it is possible to deal with systematic defects common to each die (defects caused by the exposure apparatus, etc.), and to compare with many opponents, Even if there is roughness, inspection can be performed with high sensitivity.

As described above, in various inspection apparatuses, for a plurality of image sensor outputs corresponding to a plurality of optical conditions, a die on the wafer is limited to the vicinity of the target die, and a local region is selected in each selected die. Set and extract brightness data from it, extract effective information necessary for defect determination from the extracted brightness data using the joint histogram feature space, and obtain the defect determination threshold The original purpose of performing a highly sensitive comparison determination is achieved.

Needless to say, the present invention can be applied to masks, reticles, FPD substrates such as liquid crystal displays and plasma displays other than semiconductor wafers. Any substrate can be applied as long as it has a repeating pattern such as a die.

The present invention can be used as a foreign matter / defect inspection apparatus in semiconductor manufacturing.

4 ... Defect, 20 ... Detection optical system, 22 ... Objective lens, 26 ... Opening shape,
28 ... Spatial filter, 29 ... Imaging lens, 30 ... Linear sensor (one-dimensional image sensor), 31 ... Pixel, 46a, 46b, 46c ... Die (also called chip), 53 ... Pixels of linear sensor projected on wafer, 100 ... Signal processing unit,
101 ... Overall control unit, 102 ... External input / output system, 1031 ... External display device,
DESCRIPTION OF SYMBOLS 103 ... Peripheral device, 200 ... Illumination light, 300 ... Linear beam, 400 ... Sensor output (image signal), 1000a ... Corresponding local area joint histogram feature space formation part, 1000b ... -Threshold generation unit, 1000c ... local region comparison unit, W ...
Wafer (substrate to be inspected).

Claims (14)

A defect inspection method for acquiring an image signal from a substrate to be inspected on which a repeated circuit pattern is formed, and inspecting a defect present on the substrate to be inspected based on the acquired image signal,
For the acquired image signal, a local region of interest to be cut out around the pixel including the pixel of interest,
A local region setting step for setting each of a plurality of corresponding local regions to be cut out around each of the plurality of corresponding pixels corresponding to the target pixel;
Forming a joint histogram feature space between the corresponding local regions based on the image signals cut out from the corresponding local regions set in the local region setting step;
A threshold value determining step for determining a threshold value based on the joint histogram feature space formed in the forming step;
Using the threshold value determined in the threshold value determination step, a defect detection step for detecting the defect existing in the local region of interest based on a difference image between the local region of interest and the corresponding local region; A defect inspection method comprising: The joint histogram feature space between the corresponding local regions is formed around the difference and brightness obtained from the image signals between the extracted corresponding local regions in the forming step. Described defect inspection method. In the threshold value determining step, a convex hull is calculated from a joint histogram feature space between the corresponding local regions formed in the forming step, and the threshold value is determined from the calculated convex hull. The defect inspection method according to claim 2. In the threshold value determining step, the maximum difference value (Diff) and the brightness (Bs) when the maximum difference value is calculated from the joint histogram feature space between the corresponding local regions formed in the forming step. And using the calculated maximum value (Diff) of the difference and brightness (Bs) at that time, it is obtained from the image signal cut out from the local region of interest set in the local region setting step. 3. The defect inspection method according to claim 2, wherein the threshold value is determined on the basis of the brightness (Bf). In addition, before the local region setting step, the image processing method further includes an image quality improvement step of restoring deterioration of the image signal acquired from the inspection target substrate using a point spread function representing deterioration of the optical system. The defect inspection method according to claim 1. A defect inspection method for acquiring an image signal from a substrate to be inspected on which a repeated circuit pattern is formed, and inspecting a defect present on the substrate to be inspected based on the acquired image signal,
For the acquired image signal, a local region of interest to be cut out around the pixel including the pixel of interest,
A local region setting step for setting each of a plurality of corresponding local regions to be cut out around each of the plurality of corresponding pixels corresponding to the target pixel;
A search step for searching for the similarity between the image signal of the local region of interest set in the local region setting step and the image signals of a plurality of corresponding local regions set in the local region setting step;
A determination step of determining a plurality of image signals of the corresponding local region similar to the image signal of the local region of interest using similarity information searched in the search step;
A defect inspection method comprising: a defect determination step for comprehensively determining defects by comparing the image signals of the plurality of corresponding local regions determined in the determination step with the image signals of the local region of interest. Furthermore, a forming step of forming a joint histogram feature space between the corresponding local regions based on the image signal cut out from each corresponding local region set in the local region setting step;
A threshold value determining step for determining a threshold value based on the joint histogram feature space formed in the forming step;
Using the threshold value determined in the threshold value determination step, a defect detection step for detecting the defect existing in the local region of interest based on a difference image between the local region of interest and the corresponding local region; The defect inspection method according to claim 6, further comprising: A defect inspection apparatus that acquires an image signal from a substrate to be inspected on which a repeated circuit pattern is formed, and inspects a defect present on the substrate to be inspected based on the acquired image signal,
For the acquired image signal, a local region of interest to be cut out around the pixel including the pixel of interest,
Local region setting means for setting each of a plurality of corresponding local regions to be cut out around each of the plurality of corresponding pixels corresponding to the target pixel;
Forming means for forming a joint histogram feature space between the corresponding local areas based on the image signal cut out from each corresponding local area set by the local area setting means;
Threshold determination means for determining a threshold based on the joint histogram feature space formed by the forming means;
Defect detection means for detecting the defect existing in the local region of interest based on a difference image between the local region of interest and the corresponding local region, using the threshold value determined by the threshold value determining unit. A defect inspection apparatus comprising: In the forming means, the joint histogram feature space between the corresponding local regions is formed on the basis of the difference and brightness obtained from the image signals between the extracted corresponding local regions. The defect inspection apparatus according to claim 8. The threshold value determining means calculates a convex hull from the joint histogram feature space between the corresponding local regions formed by the forming means, and determines the threshold value from the calculated convex hull. The defect inspection apparatus according to claim 9. In the threshold value determining means, the maximum value (Diff) of the difference and the brightness (Bs) when the maximum value of the difference is calculated from the joint histogram feature space between the corresponding local regions formed by the forming means. And using the calculated maximum value (Diff) of the difference and brightness (Bs) at that time, it is obtained from the image signal cut out from the local region of interest set in the local region setting step. 10. The defect inspection apparatus according to claim 9, wherein the threshold value is determined based on the brightness (Bf) to be obtained. 9. The defect inspection apparatus according to claim 8, further comprising image quality improvement means for restoring deterioration of the image signal acquired from the inspection target substrate using a point spread function representing deterioration of the optical system. . A defect inspection apparatus that acquires an image signal from a substrate to be inspected on which a repeated circuit pattern is formed, and inspects a defect present on the substrate to be inspected based on the acquired image signal,
For the acquired image signal, a local region of interest to be cut out around the pixel including the pixel of interest,
Local region setting means for setting each of a plurality of corresponding local regions to be cut out around each of the plurality of corresponding pixels corresponding to the target pixel;
Search means for searching for the similarity between the image signal of the local area of interest set by the local area setting means and the image signals of a plurality of corresponding local areas set by the local area setting means;
Determination means for determining a plurality of image signals of the corresponding local area similar to the image signal of the local area of interest using similarity information searched by the search means;
A defect inspection apparatus comprising defect determination means for comprehensively determining defects by comparing the image signals of the plurality of corresponding local areas determined by the determination means and the image signals of the local area of interest . Further, forming means for forming a joint histogram feature space between the corresponding local regions based on the image signal cut out from each corresponding local region set by the local region setting means,
Threshold determination means for determining a threshold based on the joint histogram feature space formed by the forming means;
Defect detection means for detecting the defect existing in the local region of interest based on a difference image between the local region of interest and the corresponding local region, using the threshold value determined by the threshold value determining unit. 14. The defect inspection apparatus according to claim 13, further comprising:

Images