JP2010071826A - Teacher data preparation method, and image sorting method and image sorter - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To speedily and appropriately determine the category of a defective image utilized for teacher data when creating the teacher data used for learning of a sorter for sorting defects on a substrate. <P>SOLUTION: One kind of feature amount U<SB>x</SB>is acquired from a target defective image 7 having an undetermined category, and one kind of feature amount, namely T<SB>x11</SB>, T<SB>x12</SB>, T<SB>x21</SB>, is acquired from typical images 811, 812, 821 indicating typical defects of two categories 81, 82. Then, three feature amount differences between the feature amount U<SB>x</SB>and the feature amount T<SB>x11</SB>, T<SB>x12</SB>, T<SB>x21</SB>is obtained, and one vote is casted for the category 81 to which the typical image 811 having the smallest feature amount difference belongs. The similar processing is performed to the remaining types of feature amount, and the category 81 having the largest number of votes obtained in the categories 81, 82 is determined to be a category to which the target defective image 7 belongs. In this manner, by casting votes by obtaining the feature amount difference for each type of feature amount, the category of the target defective image is determined speedily and appropriately. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a teacher data creation method for creating teacher data used for learning of a classifier, and an image classification method and an image classification apparatus that involve creation of teacher data.

  In manufacturing a semiconductor substrate, a glass substrate, a printed circuit board, and the like, an appearance inspection is performed using an optical microscope, a scanning electron microscope, or the like in order to inspect defects such as foreign matters, scratches, and etching defects. Conventionally, the cause of the defect has been identified by performing further detailed analysis on the defect detected in such an inspection process, and countermeasures against the defect have been taken.

  In recent years, as the pattern on the substrate becomes more complex and finer, the types and quantities of detected defects tend to increase, and automatic classification has been proposed to automatically classify defects detected in the inspection process. Yes. It is possible to quickly and efficiently analyze defects by automatic classification, and it is possible to preferentially take measures by paying attention to the types of defects that occur frequently.

In automatic classification, a classifier using a neural network, a decision tree, discriminant analysis, or the like is used when classifying defects. In order for the classifier to perform automatic classification, it is necessary to prepare the teacher data including a signal indicating the defect image and its category (that is, the type of the defect image) to learn the classifier. In Patent Document 1, the operator observes the teaching defect image displayed on the monitor, and creates the teacher data by adding the corresponding one from the defect category list to the teaching defect image.
JP 2000-57349 A

  By the way, the performance of the automatic classification largely depends on the quality of the teacher data for learning the classifier. In order to prepare high-quality teacher data, a large amount of accurate teaching work by an operator is required, and a great deal of labor is required. In Patent Document 1, it is necessary for an operator to assign a category to a teaching defect image, and teacher data cannot be created quickly. In addition, in the teaching work by the operator, there is a possibility that teacher data including a defect image to which an incorrect category is assigned is created due to the category assignment by the operator's ambiguous judgment and the performance of the automatic classification is lowered.

  The present invention has been made in view of the above problems, and an object of the present invention is to determine a category of a defect image used for teacher data at high speed.

  The invention according to claim 1 is a teacher data creation method for creating teacher data used for learning of a classifier that classifies a defect image indicating a defect on a substrate, wherein the defect image is classified by the classifier. In each of the plurality of categories, at least one typical image indicating a typical defect is designated in advance, and the teacher data creation method is configured to: a) select one type from the target defect image that is an unclassified defect image; Obtaining a feature quantity; b) obtaining a plurality of feature quantity differences that are differences between the one kind of feature quantity of all typical images and the one kind of feature quantity of the target defect image; c) voting to a category to which a typical image corresponding to the smallest value among the plurality of feature quantity differences belongs, and d) processes a) to c) for each type of feature quantity of the target defect image. Process , And a step of determining e) number of votes of the plurality of categories is the largest category as the category to which the target defect image belongs.

  The invention according to claim 2 is the teacher data creation method according to claim 1, further comprising the step of f) adding the target defect image to the typical image of the category determined in the step e).

  The invention according to claim 3 is the teacher data creation method according to claim 1 or 2, wherein in the step e), the number of votes of the category having the largest number of votes and the category having the second largest number of votes If the difference is less than or equal to a predetermined value, the target defect image is left unclassified.

  The invention described in claim 4 is an image classification method for classifying a defect image indicating a defect on a substrate, wherein the teacher data created by the teacher data creation method according to any one of claims 1 to 3 is used. Learning the classifier, and classifying the defect image into any of the plurality of categories by the classifier.

  The invention according to claim 5 is an image classification device for classifying a defect image indicating a defect on a substrate, wherein the teacher data generation unit generates teacher data used for learning of a classifier that classifies the defect image; A classifier for classifying defect images, and at least one typical image showing a typical defect is designated in advance in each of a plurality of categories into which the defect images are classified by the classifier, and the teacher data A) a step of a) acquiring one type of feature quantity from a target defect image that is an unclassified defect image; and b) the one type of feature quantity of all typical images and the target defect image. Obtaining a plurality of feature quantity differences that are differences from one type of feature quantity; and c) voting to a category to which a typical image corresponding to the smallest value among the plurality of feature quantity differences belongs. D) the subject Performing steps a) to c) for each type of feature amount of the fallen image; and e) determining a category having the largest number of votes among the plurality of categories as a category to which the target defect image belongs. To do.

  According to the present invention, teacher data is obtained by obtaining a difference related to a plurality of types of feature amounts without correlating the types of feature amounts of the typical image and the target defect image with each other, and further using voting for categories. It is possible to determine the category of the defect image used in the process at high speed and appropriately.

  In the invention of claim 2, it is possible to classify unclassified target defect images into a more appropriate category, and in the invention of claim 3, deterioration of the quality of teacher data is prevented.

  FIG. 1 is a diagram showing a schematic configuration of an image classification apparatus 1 according to an embodiment of the present invention. The image classification apparatus 1 shows defects on a semiconductor substrate 9 (hereinafter simply referred to as “substrate 9”). Classification of defect images is performed. The image classification device 1 picks up a defect into a category to which the defect should belong when the defect is detected based on the image data from the image pickup device 2 and the image pickup device 2 that picks up the inspection target area on the substrate 9. The inspection / classification device 4 for automatic classification and the host computer 5 for controlling the overall operation of the image classification device 1 and generating the classifier 421 used for defect classification in the inspection / classification device 4 are provided. The types of defects present on the substrate 9 are, for example, chipping, protrusion, disconnection, short circuit, and foreign matter, and these are defined as defect categories. In addition, the imaging device 2 is incorporated in the production line of the substrate 9, and the image classification device 1 is a so-called inline system.

  The imaging device 2 captures an inspection target region on the substrate 9 to acquire image data, a stage 22 that holds the substrate 9, and a stage that moves the stage 22 relative to the imaging unit 21. The imaging unit 21 has a drive unit 23 and is imaged by an illumination unit 211 that emits illumination light, an optical system 212 that guides the illumination light to the substrate 9 and receives light from the substrate 9, and an optical system 212. The imaging device 213 converts an image of the substrate 9 into an electrical signal. The stage driving unit 23 includes a ball screw, a guide rail, a motor, and the like, and the inspection area on the substrate 9 is imaged by the host computer 5 controlling the stage driving unit 23 and the imaging unit 21.

  The inspection / classification device 4 includes a defect detection unit 41 that detects a defect while processing image data of an inspection target region, and a defect automatic classification unit 42 that classifies a defect image. The defect detection unit 41 has a dedicated electric circuit for processing image data of the inspection target area at high speed, and performs defect inspection of the inspection target area by comparing the captured image with a reference image having no defect or by image processing. Do. The automatic defect classification unit 42 includes a CPU that performs various arithmetic processes, a memory that stores various types of information, and the like, and classifies defects (that is, defect images) using a classifier 421 that uses a neural network, a decision tree, discriminant analysis, and the like. Classification).

  FIG. 2 is a diagram showing a flow of defect image classification by the image classification apparatus 1. First, when the imaging device 2 shown in FIG. 1 images the substrate 9, the defect detection unit 41 of the inspection / classification device 4 acquires image data (step S11). Next, when the defect detection unit 41 performs the defect inspection of the inspection target region and a defect is detected (step S12), the image of the defective portion (that is, the defect image) is transmitted to the automatic defect classification unit 42. The The automatic defect classification unit 42 calculates a plurality of types of feature values of the defect image (step S13), and the feature values of the defect image are input to the classifier 421 of the automatic defect classification unit 42 and the classification result is output. That is, the defect image is classified into one of a plurality of categories by the classifier 421 (step S14). In the image classification device 1, the feature amount is calculated in real time every time a defect is detected by the defect detection unit 41, and automatic classification of a large number of defect images is performed at high speed.

  Next, classifier learning by the host computer 5 will be described. FIG. 3 is a diagram showing the configuration of the host computer 5. The host computer 5 has a general computer system configuration in which a CPU 51 that performs various arithmetic processes, a ROM 52 that stores basic programs, and a RAM 53 that stores various information are connected to a bus line. The bus line further includes a fixed disk 54 for storing information, a display 55 for displaying various information such as images, a keyboard 56a and a mouse 56b (hereinafter referred to as “input unit 56”) for receiving input from the operator. ), A reading device 57 that reads information from a computer-readable recording medium 8 such as an optical disk, a magnetic disk, or a magneto-optical disk, and a communication unit 58 that transmits and receives signals to and from other components of the image classification device 1. Are appropriately connected through an interface (I / F) or the like.

  In the host computer 5, the program 80 is read in advance from the recording medium 8 via the reading device 57 and stored in the fixed disk 54. Further, the program 80 is copied to the RAM 53 and the CPU 51 follows the program in the RAM 53. Arithmetic processing is executed.

  FIG. 4 is a block diagram showing a functional configuration for learning a classifier realized by the CPU 51, ROM 52, RAM 53, fixed disk 54 and the like of the host computer 5, and also shows the inspection / classification device 4. The host computer 5 includes a teacher data creation unit 61 that creates teacher data used for learning the classifier and a learning unit 62 that learns the classifier using the teacher data. In addition, these functions may be constructed by a dedicated electric circuit, or a dedicated electric circuit may be partially used.

  The teacher data includes defect image data, a feature signal of the defect image, and a teaching signal which is information indicating the category of the defect. As the feature amount of the defect image, for example, the defect area, brightness average, perimeter length, flatness, defect The inclination of the major axis when approximating to an ellipse is adopted. In the learning unit 62, the feature amount of the defect image read from the teacher data is input to a classifier (not shown) in the host computer 5, and the output of the classifier is the same as the teaching signal indicating the defect category. Learning is performed, and the learning result, that is, the learned classifier 421 (more precisely, information indicating the structure of the classifier 421 and the value of the variable) is transferred to the automatic defect classification unit 42.

  FIG. 5 is a block diagram showing a functional configuration of the teacher data creation unit 61 of the host computer 5, and also shows a display 55, an input unit 56, and a learning unit 62. The teacher data creation unit 61 includes a storage unit 611 that stores defect image data and various types of information related to the defect image, a feature amount calculation unit 612 that calculates a feature amount of the defect image, and a category that determines a category to which the defect image should belong. A determination unit 613 is included.

  The storage unit 611 stores a large number of defect images prepared in advance, and some of these defect images are defect images (hereinafter referred to as “typical images”) indicating typical defects belonging to each category by an operator as will be described later. ")"). The defect image may be acquired using the imaging device 2 and the defect detection unit 41 illustrated in FIG. 1 or may be separately prepared. The category determination unit 613 automatically determines the category of an unclassified defect image (hereinafter referred to as “untaught image”) based on the typical image, and the feature amount calculation unit 612 determines the typical image and the untaught image. Calculate all types of features.

  FIG. 6 is a diagram showing a part of the structure of information stored in the storage unit 611. In the storage unit 611, the feature amount of the defect image calculated by the feature amount calculation unit 612 shown in FIG. 5 is stored in association with the defect image data together with its type. When a category is determined for a defect image, a category number indicating the category is stored in a categorical variable associated with the defect image data, and when a defect image is designated as a typical image, the category number is associated with the defect image data. 1 is stored in the set typical image identification variable. In the initial state, 0 is stored in the categorical variable and the typical image identification variable.

  7 and 8 are diagrams showing a flow in which teacher data is created by the host computer 5 and the classifier is learned. First, a large number of defect image data are prepared in the storage unit 611 of the teacher data creation unit 61 shown in FIG. 5, and the feature amount calculation unit 612 calculates all types of feature amounts of all defect images (step S21). . The calculated feature value is stored in the storage unit 611. Next, the defect image is displayed on the display 55 according to the defect image data, the operator designates at least one typical image in each category, and a teaching signal indicating the category of the typical image is transmitted to the teacher data via the input unit 56. The data is input to the creation unit 61 (step S22). As a result, the category number is stored in the category variable shown in FIG. 6 of the defect image data designated as the typical image, and 1 is stored in the typical image identification variable.

  When the typical image is designated, the category determining unit 613 shown in FIG. 5 determines one untaught image as the target defect image (step S23), and all types of feature amounts of the target defect image and the target defect image are stored in the storage unit 611. All types of feature values of all typical images are acquired. Further, one type of feature amount of the target defect image and all typical images is selected (hereinafter, the selected type of feature amount is referred to as “selected feature amount”) (step S24).

  Then, a plurality of feature quantity differences, which are the differences between the selected feature quantities of all the typical images and the selected feature quantities of the target defect image, are obtained (each feature quantity difference is the selected feature quantity of one typical image and the target defect image. (Step S25), the smallest feature value among the plurality of feature value differences is identified, and one vote is voted for the category to which the corresponding typical image belongs (Step S26). ). That is, in the category determination unit 613, a voting variable is assigned in advance for each category, and 1 is added to the variable. The variable is set to 0 at initialization.

  FIG. 9 is a diagram for explaining a specific example of voting for a category, and illustrates the positions of a plurality of typical images and the positions of target defect images in a feature amount space having feature amounts as coordinate values. In FIG. 9, for convenience of explanation, the number of types of feature quantities and the number of categories are two, but in practice, the feature quantity space is determined by a number of types of feature quantities, and a large number of categories are set. Further, it is assumed that two typical images 811 and 812 are designated in the left category 81 in FIG. 9 and one typical image 821 is designated in the right category 82.

In FIG. 9, when the feature amount on the horizontal axis is selected as the selected feature amount (step S24), the feature amount difference between the feature amount U _x of the target defect image 7 and the feature amount T _x11 of the typical image 811 of the category 81 is _calculated. A certain absolute value | U _x −T _x11 |, an absolute value | U _x −T _x12 | that is a feature amount difference between the feature amount U _x and the feature amount T _x12 of the typical image 812 in the category 81, and the feature amount U _x And an absolute value | U _x −T _x21 | which is a feature amount difference between the feature amount T _x21 of the typical image 821 and the category 82 is obtained (step S25). Among these feature quantity differences, | U _x −T _x11 | is the smallest, so one vote is voted for the category 81 to which the typical image 811 belongs (step S26).

When one vote is completed, the type of the selected feature amount is then changed (steps S27 and S24), and a plurality of feature amount differences between the selected feature amount of all the typical images and the selected feature amount of the target defect image are obtained. One vote is voted for the category of the typical image that is found and corresponds to the smallest feature amount difference (steps S25 and S26). In the case of FIG. 9, the feature quantity indicated by the vertical axis is selected as the selected feature quantity (step S24), the feature quantity U _y of the target defect image 7 on the vertical axis and the feature quantity T _{y11 of the} category 81 typical images 811 and 812. , T _y12 , and absolute values | U _y −T _y11 |, | U _y −T _y12 |, | U _y −T _y21 | which are feature amount differences from the feature amount T _y21 of the typical image 821 of the category 82. One vote is obtained for the category 81 to which the typical image 812 corresponding to the smallest | U _y −T _y12 | As a result of the above voting, the number of votes for category 81 is 2 and the number of votes for category 82 is 0, as shown in FIG.

When q categories are set in the feature amount space where the number of types of feature amounts is p, the feature amount vector U = (U ₁ , U ₂ ,. , U _p ) of one kind of feature quantity U _k (k is an arbitrary integer from 1 to _p ) and a feature quantity vector T _ij of a typical image j of category C _i = (T _ij1 , T _ij2 ,..., T _ijp ) Is calculated as an absolute value | U _k −T _ijk | which is a feature amount difference from one type of feature amount T _ijk . However, i is an arbitrary integer of 1 to q, and j is an arbitrary integer of 1 to r, where r is the total number of typical images included in the category C _i . Then, one vote is voted for the category of the typical image corresponding to the smallest feature amount difference among the plurality of feature amount differences (steps S24 to S26).

  Further, it is confirmed whether or not voting has been performed for all types of feature amounts of the target defect image (step S27). If there are types of feature amounts for which voting has not been performed, steps S24 to S24 are performed for the types. S26 is performed. When voting is performed for all types of feature quantities, a voting result indicating the number of votes obtained for each category (one category obtains the maximum p votes) is obtained.

  When the voting result is acquired, next, the category determining unit 613 (see FIG. 5) determines whether or not to determine the category of the target defect image based on the voting result (FIG. 8: Step S31). If the difference in the number of votes between the category with the largest number of votes and the category with the second largest number is greater than a predetermined value, for example, 10% of the total number of votes, the category with the largest number of votes among a plurality of categories is the target defect. The category to which the image belongs is determined (step S32). As a result, as shown in FIG. 6, the category number indicating the category with the largest number of votes is stored in the category variable assigned to the target defect image in the storage unit 611. Hereinafter, an untaught image whose category has been determined by the above processing is referred to as a “taught image”. It is preferable that the determined category is finally confirmed after being confirmed by the operator.

  On the other hand, if the difference in the number of votes between the category with the largest number of votes and the category with the second largest number of votes is equal to or less than a predetermined value, the category of the target defect image is not determined (that is, the target defect image is not yet determined). It is left as a classification.) Even when there are two or more categories having the largest number of votes, the target defect image is left unclassified.

  Thereafter, the image classification device 1 confirms whether or not there is an untaught image that has not been voted (step S33), and if present, the teacher data creation unit 61 shown in FIG. The target defect image is determined (step S23) and the above-described steps S24 to S27 and S31 are executed, and the category is appropriately determined for the next target defect image (step S32).

  When the voting process is performed for all the untaught images, the data of the typical image and the data of the taught image stored in the storage unit 611 of the teacher data creation unit 61, and the information on the feature amount and the category are stored in the teacher. The data is transferred to the learning unit 62, and the learning unit 62 learns the classifier using the teacher data (step S34). That is, the values of variables constituting the classifier (see FIG. 4) are determined, and the structure is determined to generate a classifier.

  Although the configuration and operation of the image classification device 1 have been described above, the image classification device 1 determines the feature amount for each type of the typical image and the target defect image when determining the category of the target defect image that is an untaught image. By calculating differences between multiple types of feature quantities without correlating them with each other, it is assumed that the difference is the Euclidean distance (the difference between all types of feature quantities) connecting the position of the typical image and the position of the target defect image in the feature quantity space. The calculation is simplified as compared with the case where it is obtained as the square root of the sum of the squares. As a result, classification of unteached images (that is, determination of categories to which the images belong) is performed at high speed. In addition, by voting to the category for each type of feature value, the category that has not been voted is not affected by the feature value of that type at all, and calculation is performed independently for each type of feature value. Compared to the case where the category is determined by obtaining the Euclidean distance, all types of feature quantities are used to prevent the variation of the numerical range for each feature quantity type and the influence of a particular type of feature quantity. A process for normalizing each difference and a process for eliminating unnecessary types of feature amounts by analysis of variance (so-called ANOVA (analysis of variance)) or the like become unnecessary. As a result, it is possible to classify untaught images at high speed and appropriately.

  In the image classification device 1, since the operator automatically determines the category of the untaught image (that is, classifies the untaught image) based on the typical image previously designated in each of the plurality of categories, The burden of teacher data creation work by the person is reduced.

  Furthermore, when determining whether or not to determine the category of the target defect image that is an untaught image, the difference in the number of votes between the category with the largest number of votes and the category with the second largest number is less than or equal to a predetermined value. Since the target defect image is left unclassified and this target defect image is not used as teacher data, deterioration of the quality of the teacher data is prevented. Further, in the image classification device 1, since the presence of a defect image belonging to the wrong category is reduced as compared with the case where the operator assigns a category to the defect image, a plurality of taught images in each category are representative images. Is distributed in a state with little variation, so that the learning of the classifier can be finished quickly.

  FIG. 11 is a diagram showing a part of the flow of another operation for determining the category of the untaught image by the image classification apparatus 1. In the operation shown in FIG. 11, after the category is determined for the target defect image that is an untaught image, the operator confirms whether or not to add the target defect image to the typical image. Other processes in determining the category are the same as those shown in FIGS.

  When the category of the target defect image is determined in step S32 shown in FIG. 11 (that is, when the target defect image is a taught image), the data of the taught image and its category are stored from the storage unit 611 shown in FIG. The category number shown is read and displayed on the display 55. Then, the operator determines whether or not to add the taught image to the typical image of the determined category (step S41), and if the instruction to add is given by the operator via the input unit 56, the taught 1 is stored in the typical image identification variable (see FIG. 6) assigned to the image (step S42). Thereafter, the process proceeds to a process for determining the category of another untaught image (step S33).

  When the taught image is not added to the typical image, the typical image identification variable is left as 0. Further, as described above, if the difference between the category with the largest number of votes and the number of votes with the category with the second largest number of votes is equal to or less than a predetermined value in step S31, the target defect image remains unclassified. Steps S41 and S42 are not performed.

  In the image classification device 1, the taught images (appropriate ones) are added to the typical image, so that other untaught images can be classified into a more appropriate category. Further, a category determination process may be performed again on an untaught image for which a category has not been determined once, after another taught image is added to the typical image. The determination may be repeated until there is no determinable defect image, that is, all the images are taught images, or there are finally untaught images left unclassified. After determining the category of all unteached images (or determining to keep unclassified) without adding the taught image to the typical image, add the taught image to the typical image and remain as an unteached image The category may be determined again.

  Although the embodiment of the present invention has been described above, the present invention is not limited to the above embodiment, and various modifications can be made. For example, in the above embodiment, the feature amount of a large number of defect images used for teacher data may be calculated by the inspection / classification device 4, and the host computer 5 may acquire the feature amount from the inspection / classification device 4. .

  In the above embodiment, the category determining unit 613 acquires all types of feature amounts from the storage unit 611 in advance, and the selected feature amounts are selected in step S24, but only the selected feature amounts are stored from the storage unit 611 for each vote. May be acquired. Furthermore, the selected feature value of the typical image or the target defect image may be obtained by calculation at the stage where the type of feature value to be calculated is selected. As described above, in the category determination unit 613, how to calculate and prepare the feature amount is the same as the operation of acquiring the selected feature amount from the typical image and the target defect image for each vote in one type of feature amount. It may be done.

  Further, the defect image data may be omitted from the teacher data. In this case, a variable for specifying the defect image is provided in the teacher data. The feature amount may be omitted from the teacher data, and the learning unit 62 may obtain the feature amount based on the defect image data of the teacher data.

  In the above embodiment, a glass substrate, a printed wiring board, or a mask substrate used for exposure of the substrate may be inspected instead of the semiconductor substrate.

It is a figure which shows the structure of an image classification device. It is a figure which shows the flow of a classification | category of a defect image. It is a figure which shows the structure of a host computer. It is a figure which shows the function structure of a host computer. It is a figure which shows the function structure of a teacher data creation part. It is a figure which shows the structure of the information memorize | stored in the memory | storage part of a teacher data creation part. It is a figure which shows the flow which produces teacher data and learns a classifier. It is a figure which shows the flow which produces teacher data and learns a classifier. It is a figure which illustrates the position of the typical image in the feature-value space, and the position of a target defect image. It is a figure which shows the result of the vote to a category. It is a figure which shows a part of flow of the other operation | movement which determines the category of an object defect image.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Image classification device 7 Target defect image 9 Board | substrate 61 Teacher data preparation part 81,82 (defect image) category 421 Classifier 811,812,821 Typical image S11-S14, S21-S27, S31-S34, S41, S42 Step

Claims (5)

A teacher data creation method for creating teacher data used for learning a classifier that classifies defect images indicating defects on a substrate,
In each of a plurality of categories into which defect images are classified by the classifier, at least one typical image indicating a typical defect is specified in advance.
The teacher data creation method includes:
a) acquiring one type of feature amount from a target defect image which is an unclassified defect image;
b) obtaining a plurality of feature amount differences that are differences between the one type of feature amount of all the typical images and the one type of feature amount of the target defect image;
c) voting on a category to which a typical image corresponding to the smallest value among the plurality of feature amount differences belongs;
d) performing the steps a) to c) for each type of feature quantity of the target defect image;
e) determining a category having the largest number of votes among the plurality of categories as a category to which the target defect image belongs;
A method for creating teacher data, comprising: The teacher data creation method according to claim 1,
f) A teacher data creation method further comprising the step of adding the target defect image to the typical image of the category determined in the step e). The teacher data creation method according to claim 1 or 2,
In the step e), when the difference in the number of votes between the category with the largest number of votes and the category with the second largest number of votes is equal to or less than a predetermined value, the target defect image is left unclassified. Characteristic teacher data creation method. An image classification method for classifying defect images indicating defects on a substrate,
Learning a classifier using the teacher data created by the teacher data creation method according to any one of claims 1 to 3,
Classifying the defect image into any of the plurality of categories by the classifier;
An image classification method comprising: An image classification device for classifying defect images indicating defects on a substrate,
A teacher data creation unit for creating teacher data used for learning a classifier for classifying defect images;
A classifier for classifying defect images;
With
In each of a plurality of categories into which defect images are classified by the classifier, at least one typical image indicating a typical defect is specified in advance.
The teacher data creation unit
a) acquiring one type of feature amount from a target defect image which is an unclassified defect image;
b) obtaining a plurality of feature amount differences that are differences between the one type of feature amount of all the typical images and the one type of feature amount of the target defect image;
c) voting on a category to which a typical image corresponding to the smallest value among the plurality of feature amount differences belongs;
d) performing the steps a) to c) for each type of feature quantity of the target defect image;
e) determining a category having the largest number of votes among the plurality of categories as a category to which the target defect image belongs;
An image classification device characterized by executing

Images