JP2003168114A - Defect sorting device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a defect sorting device for enabling sorting closer to human judgment without requiring registration of defect data regarding defects such as uneven painting, faulty exposure, a flaw, adhesion of dust of a resist mainly in macro inspection of a semiconductor wafer and a liquid crystal panel. <P>SOLUTION: In the defect sorting device for discriminating kinds of detected defects based on an image of a test object, it is provided with a feature extraction means 102 for extracting feature values from an image of a detected defective area, an inference discrimination means 106 for discriminating kinds of the defects based on the feature values extracted by the feature extraction means 102 and an established area exclusion means 107 for excluding an area in which kinds of the defects discriminated by the inference discrimination means 106 are established from the image of the defective area. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a defect classification device, and more particularly to a device for determining the kind of defect type detected on the basis of an image. Regarding classifying device.

[0002]

2. Description of the Related Art In the manufacture of semiconductor wafers, liquid crystal panels and the like, a multilayer wiring pattern is formed by repeating the steps of film formation, resist coating, exposure, development, etching and resist stripping a plurality of times.

In order to improve the yield of products that are finally completed through hundreds of manufacturing processes, early detection of defects and optimization of manufacturing processes are essential items.

Therefore, conventionally, the appearance of the object is inspected before the etching step in which an actual pattern is formed.
In addition to detecting defects and selecting defective products, the type of defects is discriminated and the information is fed back to the manufacturing process.

There are mainly two types of visual inspections here: micro inspection and macro inspection.

The micro inspection has a resolution of about 5 μm.
It detects minute defects and pattern abnormalities.
The macro inspection is an inspection for detecting relatively large defects such as uneven coating of resist, poor exposure, scratches, and dust adhesion with a resolution of about 100 μm.

The macro inspection has advantages that it can detect defects that cannot be seen locally even by micro inspection, and has high throughput because of its low resolution.

Regarding the defect classifying apparatus, Japanese Patent Laid-Open No. 8-21
The "defect type determination device and process management system" described in Japanese Patent No. 803 is constructed with a neural network which is trained to output the defect type for the input pattern of each defect information (= feature amount). The configuration of the device for classifying defects is shown.

Further, the "teaching data creating method and the defect classifying method and the method thereof, which are disclosed in Japanese Patent Laid-Open No. 11-344450, which pointed out and improved the problem concerning the method of giving the teaching data at the time of learning in the prior art, have been proposed. Equipment ”
The configuration of an apparatus is shown in which the distribution of each defect in a multidimensional space based on the image feature amount is preset based on teaching data, and the defect is classified by statistical discriminant analysis.

Furthermore, in the "defect type determination device and process management system" described in Japanese Patent Laid-Open No. 8-94536, measured values of a plurality of evaluation items (feature amounts) for detected defects are divided into stages and coded. , A configuration of a device for classifying defects by referring to a preset defect name based on the code string (a so-called classification tree method) is shown.

[0011]

However, in a defect classifying apparatus using a neural network or a statistical method as disclosed in Japanese Patent Application Laid-Open Nos. 8-21803 and 11-344450, it is necessary to register a large number of defect samples in advance. There is a problem that it takes a lot of time.

Further, usually, since registration is required every time the product type and process are changed, there is also a problem that a sufficient number of defect samples cannot be prepared for a small number of product types.

On the other hand, in the device disclosed in Japanese Patent Laid-Open No. 8-94536, a defect corresponding to each code string can be set based on human experience and knowledge, and the above problem does not occur, but the code Ambiguous processing in the vicinity of switching is not possible, that is, there are objects that are classified as completely different results due to subtle differences in the switching position, or objects that have the same code string but values near switching are clearly separated from switching. You can only give the same result to a target that has a value,
Such problems occur.

Further, many of the feature quantities used in the defect classifying apparatus up to now are such that the detected defect area is regarded as a single unit and processed, and there are a defect and a detection area which are recognized by a human like micro inspection. There is no problem in the case of one-to-one correspondence, but there is no one-to-one correspondence between the defect recognized by humans and the detection area as in the macro inspection (as shown in FIG. If one defect is divided and detected depending on the quantization level of the image), there is a problem that accurate classification is difficult.

Until now, an apparatus for solving the above problems and effectively performing defect classification has not been clarified.

The present invention has been made in view of these problems, and its purpose is mainly to prevent uneven coating of resist, exposure failure, scratches, and dust in macro inspection of semiconductor wafers and liquid crystal panels. It is an object of the present invention to provide a defect classifying apparatus which enables classification closer to human judgment, without requiring registration of defect data with respect to defects such as adherence of defects.

[0017]

In order to achieve the above-mentioned object, a first invention is a defect classifying device for discriminating the type of a defect detected based on an image of an object to be detected. Feature amount extraction means for extracting a feature amount from the image, defect type determination means for determining a defect type based on the feature amount extracted by the feature amount extraction means, and defect type determined by the defect type determination means And a definite area excluding means for excluding the area definite as from the image of the defective area.

A second aspect of the present invention is the defect classification apparatus according to the first aspect, wherein the feature quantity extracting means includes means for extracting a feature quantity based on the arrangement shape of at least two or more detection areas.

A third aspect of the present invention is the defect classification apparatus according to the first aspect, wherein the feature amount extraction means is at least two in number by performing expansion processing using structural elements that are adaptive to the shape of each detection region. It includes a connecting unit that connects the above-described detection regions, and a unit that extracts the feature amount of the connected region.

A fourth aspect of the present invention is the defect classification apparatus according to the third aspect, wherein the connecting means calculates a maximum / minimum Feret diameter ratio (≧ 1) and a principal axis angle of each defect area, and A means for selecting the shape of the structural element used for expansion processing from circular or fine linear based on the value of the maximum and minimum Feret diameter ratio, and the fine linear with respect to the principal axis angle of the defect region in which the fine linear structural element is selected. And a means for connecting the overlapping areas by performing expansion processing on the defect area image using the structuring element determined for each defective area.

A fifth aspect of the present invention is the defect classification apparatus according to any one of the first to fourth aspects of the invention, wherein the defect type determination means is a fuzzy inference based on the value of the feature quantity and a preset classification rule. Including means for performing.

A sixth invention is the defect classification apparatus according to any one of the first to fifth inventions, wherein the defect is an exposure defect or coating unevenness caused in a manufacturing process of a semiconductor wafer, a liquid crystal panel or the like. , Scratches, dust, dust, etc.

A seventh invention is the defect classification apparatus according to the sixth invention, wherein the feature quantity extracting means includes means for extracting the feature quantity using the exposure section information in the exposure step.

An eighth aspect of the present invention is the defect classification apparatus according to the seventh aspect, wherein the means for extracting the feature amount using the exposure section information is means for creating a template image showing the shape of the exposure section. Means for scanning the template image in accordance with the exposure division position of the image of the detection defect area, calculating the correlation value of both images at each exposure division position, and the detection defect in the exposure division for which the correlation value is calculated. And a means for adding the correlation value as a feature amount.

A ninth aspect of the invention is the defect classification apparatus according to the seventh aspect of the invention, wherein the means for extracting the feature amount using the exposure section information integrates the pixels in the detected defect area vertically or horizontally. Means, means for calculating the amplitude value at each frequency by Fourier transforming the outline curve of the integrated value, and calculating the ratio of the amplitude value of the frequency corresponding to the exposure section cycle to the average amplitude value of the frequencies corresponding to other cycles. Means, and means for adding the value of the ratio as a feature amount to the detected defects used for the integration.

A tenth aspect of the present invention is the defect classification apparatus according to the seventh aspect, wherein the means for extracting the characteristic amount using the exposure section information is a template image from an arbitrary exposure section in the image of the detected defect area. And a correlation value is calculated by aligning the template image with an exposure section located above or below or on the left or right of the arbitrary exposure section,
It includes means for calculating the average and means for adding the average value to the detected defects in the arbitrary exposure section as a feature amount.

An eleventh aspect of the invention is the defect classification apparatus according to the sixth aspect of the invention, wherein the feature amount extracting means includes means for extracting the feature amount using polar coordinates with respect to the wafer rotation center in the coating step. .

A twelfth aspect of the present invention is the defect classification apparatus according to the eleventh aspect, wherein the feature amount extraction means converts a polar coordinate image obtained by converting the detected defect area into a polar coordinate system with respect to the wafer rotation center in the coating step. It includes a means for creating, a means for detecting a straight line component in the polar coordinate image, and a means for adding the length of the straight line component as a feature amount to a defect area forming the straight line component.

A thirteenth aspect of the present invention is the defect classification apparatus according to any one of the first to twelfth aspects of the present invention, wherein display means for displaying an image of the detected defect area and defect type is determined on this display means. Means for clearly indicating the determined areas in the order in which they have been determined (for example, changing the color), thereby indicating to the user the progress in which the defect type has been determined.

[0030]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described in detail below with reference to the drawings. Prior to describing the embodiments of the present invention, the types of defects to be classified will be described. 1A to 1H are diagrams showing main defect examples (target: semiconductor wafer) in macro inspection.

In the resist coating step, a coating material (liquid) is dropped on the central portion of the wafer by using, for example, a coater device,
The resist is applied to the entire surface of the wafer by rotating the wafer around it. The defects generated in the same process are as follows.

Coating unevenness 1 (Fig. 1 (A)): When the dropping is not successful and the coating is scattered, or when the viscosity of the coating material is not appropriate and it does not spread smoothly during rotation, it is radial to the center of rotation. Unevenness occurs.

Coating unevenness 2 (FIG. 1 (B)): When foreign matter adheres to the wafer, the coating material does not flow smoothly from that portion to the outside, resulting in uneven comet shape.

Coating unevenness 3 (FIG. 1C): As in FIG. 1B, when foreign matter adheres, unevenness in a spiral shape may occur depending on the viscosity of the coating material and the rotation speed during coating. Occurs.

In the exposure process, in order to form a pattern on the wafer after resist coating, transfer is performed in units of predetermined exposure sections. A typical device is a stepper. The defects generated in the same process are as follows.

Exposure failure 1 (FIG. 1 (D)): The focus shifts in units of exposure sections, which causes the section to have a different surface state from other sections (the pattern is sagging).

Exposure failure 2 (FIG. 1 (E)): The exposure apparatus is tilted, and the step is continued while the focus is partially deviated in the exposure section, so that a surface state different from the normal state occurs in a cycle of the step unit.

Exposure failure 3 (FIG. 1 (F)): A foreign film is mixed into the back side of the resist film to be exposed, the resist film is partially raised, and the focus is deviated at that part, so that the surface is different from the others. It becomes a state.

In addition to the defects represented by the above-mentioned steps, there are the following defects.

Scratches (FIG. 1 (G)) Dust, dust adhesion (FIG. 1 (H)) In an appearance inspection using an image, as shown in Japanese Patent Laid-Open No. 8-21803, the appearance of the subject is interfered with, By imaging with a diffraction image or the like and comparing it with a reference non-defective wafer image, or by comparing in units of sections having the same pattern, a portion (defect portion) different from usual is detected.

The embodiments of the present invention will be described in detail below. FIG. 2 is a diagram showing the configuration of the defect classification device according to the present embodiment. The defect classification device according to the present embodiment is provided with a condition setting unit 101 for setting various conditions necessary for classification, a region single feature extraction unit 103 for obtaining a feature relating to the shape and brightness of a detected region alone, and a detected region. Area connecting means 104 for connecting the areas according to a predetermined method, and area arrangement feature extracting means 105 for obtaining characteristics according to the arrangement shape of the detected areas.
And the possibility of each defect type is deduced based on the feature amount obtained by each feature extracting means of the region single feature extracting means 103 and the area arrangement feature extracting means 105 and the classification rule set by the condition setting means 101. The inference determining means 106 for determining a defect, the definite area excluding means 107 for excluding the area in which the defect type is definite by the determination, and the memory 108 used for each of the above processes. The region single feature extraction unit 103, the region arrangement feature extraction unit 105, and the region connection unit 104 constitute the feature extraction unit 102.

The defect classification apparatus of the present invention uses a binary image in which defective pixels detected by comparison with the above-described non-defective wafer image or comparison in the same pattern section are bright and the other are dark (hereinafter, Defect detection image) and the appearance (multivalued) image of the subject used for this defect detection (hereinafter referred to as inspection image)
Is obtained, and the result of determining what the defect type is for each detection region is output.

The operation of the above-described structure will be described below with reference to FIGS.
This will be described with reference to the flowchart in FIG. After acquiring the defect detection image and the inspection image (step S1), the defect classification apparatus extracts the feature amount of the detection region alone in the region alone feature extraction means 103 (step S2). Next, the area arrangement feature extraction unit 105 extracts the characteristic amount based on the arrangement shape of the plurality of areas, and adds the characteristic amount to each detection area forming the arrangement shape (step S3). afterwards,
Based on the information obtained by the region single feature extraction unit 103 and the defect detection image, the region connection unit 104 performs expansion processing using a structural element that is adaptive to the shape of each detection region, and an image in which the regions are connected (hereinafter referred to as an image). , Connected area image) is created (step S4).

The connected area image is sent again to the area-only feature extraction means 103, where the characteristic amount of the connected area is extracted and this characteristic amount is added to each detection area forming the connected area (step S5). Based on the feature amount obtained in steps S2, S3, and S5 and the classification rule set by the condition setting unit 101, the inference determination unit 106 infers the possibility value of each defect type and determines the defect type ( Step S6). Based on the result of step S6, the definite area excluding unit 107 excludes the area in which the defect type is definite from the defect detection image (step S7). If all the areas are excluded as a result of step S7, the defect types of all the detection areas have been determined, so the processing ends.

If a region remains in the defect detection image, steps S3 to S7 are repeated based on this image.

Since the images are different, the feature amount of the arrangement shape and the feature amount of the connected area are different. That is, a more accurate feature amount can be extracted by excluding the confirmed region, and inference and discrimination can be performed based on this.

Next, the details of the processing by each of the above means will be described. The condition setting means 101 inputs various conditions required for classification. The input contents will be described along with the processing by each means described later.

FIG. 4 is a flow chart for explaining the details of the processing in the region single feature extraction means 103. In the region single feature extraction means 103, the following procedure is performed.
The feature amount of each detection area is obtained.

1. Labeling the plurality of detected areas in the defect detection image (step S10 in FIG. 4, FIG. 5).
(B)).

2. The position characteristics (the center of gravity position, the circumscribed rectangle position) of the detection area are calculated (step S11 in FIG. 4, FIG. 6).
(A)).

3. Shape characteristics (area, circularity, maximum / minimum Feret diameter ratio (≧ 1), main axis angle) of the detection region are calculated (step S12 in FIG. 4, FIG. 6A).

4. Luminance characteristics of detection area (average, variance)
Is calculated (step S13 in FIG. 4, FIG. 6A).

5. A feature amount list is created for each detection region (step S14 in FIG. 4, FIG. 6B).

The series of processes 1, 2, 3 and 4 described above is a general image processing method called particle analysis (or blob analysis).

FIG. 7 is a flow chart for explaining the details of the processing in the area connecting means 104. 8 and 9 are explanatory diagrams of region connection.

The area connecting means 104 uses the features obtained by the area single-feature extracting means 103 to create a connected area image in which the respective detection areas are connected according to the following procedure.

1. Based on the maximum and minimum Feret's diameter ratio of each area, the structural element for the expansion processing for that area is selected. It is determined whether or not the maximum / minimum Feret diameter ratio is equal to or greater than a certain threshold value (either a fixed value of the trial result or a value set by the condition setting means 101) (step S20 in FIG. 7).
If YES, it is regarded as a region having directionality, and a fine linear structuring element (FIG. 8A) is selected (step S2 in FIG. 7).
1).

Further, if the judgment in step S20 is NO, that is, if it is smaller than the threshold value, the structural element of the circle diameter (see FIG. 8).
(B)) is selected (step S23 in FIG. 7). In addition,
The above selection may be performed using the circularity.

2. When the fine linear structuring element is selected, the fine linear direction is further aligned with the main axis direction of the region (step S22 in FIG. 7). FIG. 8C shows an example of directions of the fine linear structuring element. The direction is limited when the size of the structural element is small, but the direction can be subdivided by increasing the size. It should be noted that the direction of the area may be estimated from the histogram distribution in the contour direction, and the same direction may be obtained.

3. Expansion processing (step S24 in FIG. 7) is performed using the selected structuring element to create a connected area image. The expansion processing here is one of the general image processing methods called morphology processing.
The overlapping areas will be connected. The number of times of expansion processing is set by the condition setting means 101 based on the imaging system, various conditions of the subject, and the size of each structural element.

If it is determined in step S25 in FIG. 7 that all the detection areas have been determined, the process ends.

FIG. 8D shows the result of performing the expansion process of this example on the detection region of FIG. 5A.

By performing the expansion process with the structural element adaptive to the shape of the detection area by the above procedure, it is possible to expand and connect the directional area in the longitudinal direction and the other areas in the radial direction. Further, if the structural element is re-determined based on the area shape after the expansion and the contraction processing is performed, even better area connection is possible.

Identical structural element (circular) in all detection areas
In the general expansion or expansion / contraction process using the above, the directionality of the detection region is not considered, and therefore, FIG.
As shown in (A) and (B), the flaw and the dust in the vicinity thereof cannot be connected in a good state, and the subsequent feature extraction cannot be performed accurately.

The connected area image created by the area connecting means 104 is sent again to the area single feature extracting means 103,
It is processed by the following procedure. The basic processing flow is the same as in FIG.

1. Each area after connection (referred to as connection area)
Are labeled (FIG. 10 (A)).

2. Positional characteristics (barycentric position, circumscribing rectangle position) of the connected region are calculated.

3. Shape characteristics (area, circularity, maximum / minimum Feret diameter ratio (≧ 1), principal axis angle) of the connection region are calculated.

4. The calculated features are added to the detection areas that are the respective constituent parts to create a feature amount list for each detection area (FIG. 10 (B)).

The series of treatments 1, 2, and 3 described above is based on particle analysis.

The area arrangement feature extraction means 105 obtains area arrangement features based on the defect detection images before connection. The typical layout features are listed below and the calculation method is shown.

1) A feature amount indicating the degree of construction of the exposure section shape by the detection area (used for determining exposure failure 1).

<Calculation Method> 1. A template is created in which the portion of the exposure section shape is bright and the periphery thereof is dark (FIG. 11A).

2. Arrange them so that the exposure partition position in the defect detection image (binary image in which the defective pixels are bright and the others are dark) and the exposure partition shape in the template completely overlap, and the sum of the absolute differences between the two images within the template range. Is calculated (FIG. 11 (B)).

The sum R of the absolute differences between the two images is

[Equation 1] Can be obtained by Here, I (a, b) is the pixel value (1 or 0) at the coordinates (a, b) of the defect detection image.
Is. Further, T (a, b) is a pixel value (1 or 0) at the template coordinates (a, b).

3. The value calculated in 2 is added as a feature amount to the detection area in the target exposure section.

4. Scan the template (Figure 12
(A)), processes 2 and 3 are performed at all exposure division positions in the defect detection image.

Incidentally, in order to reduce the calculation amount, FIG.
You may use the template as shown in FIG.

2) A feature amount indicating the degree of periodicity of the exposure division unit in the horizontal and vertical directions (used for determining the exposure failure 2).

<Calculation Method 1> 1. The detection pixels (bright pixels) in the defect detection image are integrated in the vertical direction (FIG. 13A).

2. The curve obtained from the integration result is Fourier transformed (FIG. 13 (B)).

3. Power spectrum value X corresponding to the exposure section cycle and power spectrum value a corresponding to another cycle
The ratio (averagely calculated in portions other than the exposure division cycle) is calculated (FIG. 13B).

4. The value calculated in 3 is added to each detection area in the image as a feature amount (horizontal condition).

In order to obtain the local feature quantity,
It is also possible to set a plurality of strip areas as shown in FIG. 4 and integrate the detection pixels in each strip area for processing. In that case, the calculated feature amount is added to each detection area in the strip area.

In the present embodiment, the vertical integration is shown, but the horizontal integration can be used as a feature quantity indicating the vertical cycle condition.

<Calculation Method 2> 1. A template 52 is created from the detection area image in the target exposure section (FIG. 15 (A)).

2. The template of No. 1 is arranged so as to completely overlap the exposure section (FIG. 15B) on the left of the target exposure section, and the sum of the absolute difference values of both images is calculated within the template range.

3. The template of No. 1 is arranged so as to completely overlap the exposure section (FIG. 15B) on the right of the target exposure section, and the sum of the absolute difference values of both images is calculated within the template range.

4. The average value of the values obtained in 2 and 3 is added to each detection region in the target exposure section as a feature amount (horizontal condition in the horizontal direction).

5. Scan the template (Figure 12
(A)), 1 to 1 at all exposure division positions in the defect detection image
Process 4 is performed.

When the target exposure section is located at the edge and there is no left or right adjacent exposure section, the value on only one side is used.

Further, in the present embodiment, the difference between the horizontal direction and the left and right neighbors is shown, but by performing the difference between the vertical upper and lower neighbors, it is possible to obtain the feature amount indicating the vertical cycle condition.

3) A characteristic amount (a coating unevenness 1) indicating the total amount of radial components (relative to the rotation center at the time of coating) in the entire image.
Used to judge).

<Calculation method> 1. The edges of the detection area are extracted (FIG. 16 (A)).

2. Pixels of the extracted edge part are displayed in polar coordinate system image (horizontal axis: angle θ,
Converted to vertical axis: distance r from center (Fig. 16)
(B)). The vertical line segment in the polar coordinate system image corresponds to the radial line segment in the original image. The vertical line segment is detected using an operator (template) for detecting the vertical line segment (FIG. 17A).

3. The total length (total area) of the detected vertical line segments is added as a feature amount to each detection region in the image.

Performing the connection processing in the vertical coordinate thin line structuring element in the polar coordinate system image corresponds to performing the radial direction connection in the original image. Further, the vertical line segment as shown in FIG. 17 (B) can be detected by using contour tracing from the outer peripheral side together.

4) A characteristic amount (a coating unevenness 2) indicating the amount of the radial component (relative to the rotation center at the time of coating) in each detection area.
Used to judge).

<Calculation method> 1. The processing 1 to 3 of the feature amount calculation is performed.

2. The length of the detected vertical line segment is added as a feature amount to each corresponding detection area.

In the polar coordinate system image, the feature amount obtained by performing the connection process with the thin line structuring elements in the vertical direction (corresponding to the radial direction connection in the original image) and then measuring the length of the vertical line segment. May be added to each area that is a constituent part of the connection area.

5) A feature amount (used for determining the coating unevenness 3) indicating the ratio of the spiral component (with respect to the rotation center at the time of coating) in each detection area.

<Calculation Method> 1. The processing 1 and 2 for calculating the characteristic amount is performed. The spiral component in the original image corresponds to the diagonal line segment in the polar coordinate system image.

2. The diagonal line segment is detected using an operator for detecting the diagonal line segment.

3. The length of the detected diagonal line segment is added to each corresponding detection region as a feature amount of.

Note that the Hough transform may be used for line segment detection in the calculation of the feature amount of.

In the inference discrimination means 106, 103, 105
And the condition setting means 10 obtained by each of the feature extracting means
The possibility of each defect type is inferred based on the classification rule set by 1.

Before showing the inference process, the classification rules will be described. If you can say that "small and elongated areas are more likely to be scratches" or "if they are small, there is a possibility of scratches even if they are not long and slender" from the experience and knowledge that we have gained The maximum and minimum ferret diameter ratio is used as an index to express the area and the slenderness (the smaller the slender, the larger the value)
Can be used as a rule as follows.

: IF area = small AND Maximum and minimum ferret diameter ratio = large THEN Possibility of scratches = high: IF area = small AND Maximum and minimum ferret diameter ratio = small THEN Possibility of scratches = medium: IF area = large AND Maximum / minimum ferret diameter ratio = Large THEN Possibility of scratches = Medium: IF Area = Large AND Maximum / minimum ferret diameter ratio = Small THEN Possibility of flaws = Low Next, “small”, “large” used in the rule,
For the expressions "high" and "low", set a change line (membership function) that indicates the degree to which the numerical values on each characteristic axis match with a value of 0 to 1 (fitness) (FIG. 18 (A) to FIG. 18 (C)).

A process for inferring the possibility of a flaw in a certain detection region (area = x, maximum / minimum Feret diameter ratio = y) based on the above settings is shown in FIGS. This is a general reasoning method as a direct method in fuzzy reasoning ((reference) Russell: Introduction to fuzzy theory for those who aim for application: Kazuo Tanaka).

In FIG. 20, the sets obtained by the respective rules ~ in FIG. 19 are combined (logical sum) to obtain an intermediate solution (center of gravity Z). As a result, it is equivalent to obtaining the possibility of a flaw in the case of considering the four rules from the viewpoint of each rule and obtaining a complementary determination result.

In this example, only four rules are shown for scratches, but the accuracy can be improved by setting more rules.

Similarly, by setting a plurality of rules (& membership function) based on the knowledge of other defects and processing these rules, the probability values for all defect types can be calculated.

After the possibility value of each defect type is inferred, the defect type is determined by the following processing.

1. The probability values of each defect type in the detection area of the determination target are arranged in descending order.

2. In the case of “highest probability of defect type−probability of second highest defect type> threshold”, it is determined that the defect type of the target detection region is fixed.

3. 1. And 2. This process is performed for all the detection areas in the defect detection image.

Incidentally, 2. The threshold value in 1 may be a fixed value of the trial result or a value set by the condition setting means 101.
In addition, the above-mentioned repeated processing (FIG. 3, steps S3 to S
It may be set so as to decrease as the number of times of 8) increases.

The confirmed area excluding means 107 excludes the area where the defect type is confirmed by changing the pixels in the detection area where the defect type is confirmed in the defect detection image from bright to dark.

[0120]

According to the present invention, a more accurate feature can be obtained and the classification accuracy can be improved by extracting the feature based on the arrangement shape based on the image excluding the area where the defect type is determined.

Further, by extracting the feature of the region obtained by performing the expansion (concatenation) process on the basis of the image excluding the region where the defect type is determined, a more accurate feature is obtained and the classification accuracy is improved. .

Further, since the defect classification is performed by the knowledge-based classification rule, a teaching sample is not required.

Further, it is possible to show a region having a remarkable feature indicating a defect type and a region not having such a feature as a difference in the possibility of the defect type.

Further, not only the feature amount of the detection region alone,
The classification accuracy can be improved by using the feature amount based on the arrangement.

[0125] Further, by connecting the detection areas with adaptive structuring elements, accurate connected area characteristics can be obtained and the classification accuracy is improved.

[Brief description of drawings]

FIG. 1 is a diagram showing an example of main defects in a macro inspection.

FIG. 2 is a diagram showing a configuration of a defect classification device according to the present embodiment.

FIG. 3 is a flowchart for explaining details of defect classification processing.

FIG. 4 is a flowchart for explaining the details of the processing in the region single feature extraction means 103.

FIG. 5 is an explanatory diagram (part 1) of region single feature extraction.

FIG. 6 is an explanatory diagram (part 2) of region-only feature extraction.

FIG. 7 is a flow chart for explaining the details of the processing in the area connecting unit 104.

FIG. 8 is an explanatory diagram (1) of region connection.

FIG. 9 is an explanatory diagram (2) of region connection.

FIG. 10 is an explanatory diagram of connected region single feature extraction.

FIG. 11 is an explanatory diagram of a feature amount indicating a configuration of an exposure section shape, (A) shows a template in which an exposure section shape is bright and its periphery is dark, and (B) is a defect detection image. The exposure section is shown.

FIG. 12 is an explanatory diagram of a feature amount indicating a configuration of an exposure section shape, (A) shows an example of scanning a template,
(B) shows an example of a template for reducing the calculation amount.

13A and 13B are explanatory diagrams (No. 1) of the feature amount indicating the periodicity of the exposure section size, in which (A) shows integration of the detection pixels in the vertical direction, and (B) shows Fourier transform of the integration curve. The characteristic amount showing the periodicity of the exposure section size is shown.

FIG. 14 is an explanatory diagram (No. 1) of the characteristic amount indicating the periodicity of the exposure section size, showing integration in the vertical direction in the strip area.

15A and 15B are explanatory diagrams (No. 2) of the feature amount indicating the periodicity of the exposure section size, in which (A) shows the creation of a template based on the detection result image of the target exposure section, and (B) shows the target exposure. The exposure sections on the left and right in the horizontal direction of the section are shown.

16A and 16B are explanatory diagrams of a feature amount indicating the amount of a radial component in the entire image, FIG. 16A shows edge extraction of a detection region, and FIG. 16B shows polar coordinate conversion with respect to a rotation center at the time of application. There is.

FIG. 17 is an explanatory diagram of a feature amount indicating the amount of a radial component in the entire image, (A) showing a vertical line segment 1 in a polar coordinate image.
(B) shows the vertical line segment detection 2 in the polar coordinate image.

FIG. 18 is an explanatory diagram of calculation of a possibility of a defect by fuzzy inference, showing a membership function with respect to an expression in a flaw determination rule.

FIG. 19 is an explanatory diagram of calculation of a possibility of a defect by fuzzy inference, showing an inference process (1).

FIG. 20 is an explanatory diagram of calculation of a possibility of a defect by fuzzy inference, showing an inference process (No. 2).

[Explanation of symbols]

101 Condition setting means 102 feature extraction means 103 area single feature extraction means 104 area connecting means 105 area arrangement feature extraction means 106 Inference discrimination means 107 Definite area excluding means 108 memory

[Procedure amendment]

[Submission date] June 21, 2002 (2002.6.2)
1)

[Procedure Amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0076

[Correction method] Change

[Correction content]

The value calculated in 3.2 is added to the detection area in the target exposure section as the characteristic amount of 1) .

[Procedure Amendment 2]

[Document name to be amended] Statement

[Name of item to be corrected] 0083

[Correction method] Change

[Correction content]

The value calculated in 4.3 is added to each detection region in the image as the feature amount (horizontal condition) in 2) .

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0089

[Correction method] Change

[Correction content]

The average value of the values obtained in 4.2 and 3 is added to each detection region in the target exposure section as the feature amount (horizontal condition in the horizontal direction ) of 2) .

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0096

[Correction method] Change

[Correction content]

3. The total length (total area) of the detected vertical line segments is added to each detection region in the image as a characteristic amount of 3) .

[Procedure Amendment 5]

[Document name to be amended] Statement

[Correction target item name] 0099

[Correction method] Change

[Correction content]

<Calculation Method> 1. 3) Perform 1 to 3 processing of feature amount calculation.

[Procedure correction 6]

[Document name to be amended] Statement

[Correction target item name] 0100

[Correction method] Change

[Correction content]

2. The length of the detected vertical line segment is added to each corresponding detection area as a feature amount of 4) .

[Procedure Amendment 7]

[Document name to be amended] Statement

[Correction target item name] 0103

[Correction method] Change

[Correction content]

<Calculation Method> 1. 3) Perform the processing 1 and 2 of the feature amount calculation. The spiral component in the original image corresponds to the diagonal line segment in the polar coordinate system image.

[Procedure Amendment 8]

[Document name to be amended] Statement

[Correction target item name] 0105

[Correction method] Change

[Correction content]

3. The length of the detected diagonal line segment is added to each corresponding detection area as the feature amount of 5) .

[Procedure Amendment 9]

[Document name to be amended] Statement

[Correction target item name] 0106

[Correction method] Change

[Correction content]

The Hough transform may be used for line segment detection in the feature amount calculation of 3), 4) and 5) .

   ─────────────────────────────────────────────────── ─── Continued front page    F term (reference) 2G051 AA51 AB01 AB02 EA11 EA12                       EA14 EC01 EC02 EC04 ED01                       ED23                 2H088 FA11 FA30 MA20                 4M106 AA01 AA20 CA38 CA41 DB21                       DJ20 DJ27 DJ40                 5B057 AA03 BA02 CA12 CA16 DA06                       DB02 DC36                 5L096 BA03 CA02 GA34 JA11

Claims (13)

[Claims] 1. A defect classification device for determining the type of a defect detected based on an image of a subject, a feature amount extraction means for extracting a feature amount from an image of a detected defect area, and this feature amount extraction means. Based on the feature amount extracted by
A defect classification comprising: a defect type discriminating unit for discriminating a defect type; and a definite region excluding unit for excluding a region in which the defect type discriminated by the defect type discriminating unit is confirmed from the image of the defect region. apparatus. 2. The defect classification device according to claim 1, wherein the feature quantity extraction means includes means for extracting a feature quantity based on an arrangement shape of at least two or more detection areas. 3. The feature quantity extracting means connects the at least two or more detection areas by expansion processing using a structuring element that is adaptive to the shape of each detection area, and the feature quantity of the connection area. The defect classification device according to claim 1, further comprising: 4. The connecting means is means for calculating a maximum / minimum Feret diameter ratio (≧ 1) and a principal axis angle of each defect region, and a structure used for expansion processing based on the calculated value of the maximum / minimum Feret diameter ratio. A means for selecting the shape of the element from a circular shape or a fine linear shape, a means for aligning the fine linear direction of the fine linear structural element with the principal axis angle of the selected defect area, and a structural element determined for each defect area 4. The defect classification apparatus according to claim 3, further comprising means for performing expansion processing on the defect area image by using, and connecting the overlapping areas. 5. The defect according to claim 1, wherein the defect type determining means includes means for performing fuzzy inference based on a value of a feature amount and a preset classification rule. Classifier. 6. The defects include exposure defects, coating unevenness, scratches, dust, which may occur in manufacturing processes of semiconductor wafers and liquid crystal panels.
The defect classification device according to claim 1, wherein the defect classification device is dust or the like. 7. The defect classification device according to claim 6, wherein the feature quantity extraction means includes means for extracting the feature quantity using the exposure section information in the exposure step. 8. The means for extracting a feature amount using the exposure division information includes means for creating a template image showing an exposure division shape, and scanning the template image in accordance with the exposure division position of the image of the detection defect area. And a means for calculating a correlation value of both images at each exposure section position, and a means for adding the correlation value as a feature amount to the detected defect in the exposure section for which the correlation value is calculated. The defect classification device according to claim 7. 9. A means for extracting a feature amount using the exposure section information is a means for vertically or horizontally integrating pixels in a detected defect area, and a Fourier transform of an outline curve of the integrated value to obtain an amplitude value at each frequency. Means for calculating the ratio of the amplitude value of the frequency corresponding to the exposure section cycle to the average amplitude value of the frequencies corresponding to other cycles, and the value of the ratio for the detection defect used in the integration 8. The defect classification device according to claim 7, further comprising means for adding as a feature amount. 10. The means for extracting a feature amount using the exposure section information includes a means for creating a template image from an arbitrary exposure section in an image of a detected defect area, and the template image being located above and below the arbitrary exposure section. Alternatively, it includes means for calculating a correlation value in accordance with the exposure sections located on the left and right sides and calculating an average thereof, and means for adding the average value as a feature amount to the detected defects in the arbitrary exposure section. The defect classification device according to claim 7, which is characterized in that. 11. The defect classification device according to claim 6, wherein the feature quantity extraction means includes means for extracting the feature quantity using polar coordinates with respect to a wafer rotation center in the coating step. 12. The feature quantity extracting means, means for creating a polar coordinate image in which a detected defect area is converted into a polar coordinate system with respect to a wafer rotation center in a coating step, and means for detecting a linear component in the polar coordinate image. 12. The defect classification device according to claim 11, further comprising: a unit that adds the length of the straight line component as a feature amount to a defect area forming the straight line component. 13. A display means for displaying an image of a detected defect area, and means for clearly indicating the areas in which the defect type has been determined on the display means in the order in which they have been determined. The defect classification device according to claim 1, wherein the defect classification device is shown to a user.