JP2011158373A - Method for creation of teacher data for use in automatic defect classification, and method and apparatus for automatic defect classification - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a teacher data superior in the repeatability and the objectivity, and alleviate a load of an operator for creating the teacher data, concerning a method for creating the teacher data for automatically classifying defects in a substrate, a defect classification method using it, and an apparatus. <P>SOLUTION: In the method, the operator teaches only a teacher image as a typical example of at least one category in a plurality of classification categories among collected defect image data (step S202). A host computer calculates the feature quantity of each defect image (step S203), and each defect image is tentatively classified in response to a Euclidean distance between the teacher image in a feature quantity space (step S204). The tentatively-classified defect image is reclassified by using an ensemble classifier (step S205-S209) to keep on setting the classification sequentially. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a technique for automatically classifying defects in a substrate such as a semiconductor substrate.

  In the manufacturing technology field of substrates such as semiconductor substrates, glass substrates for liquid crystal display devices, and printed wiring boards, the quality of products is stabilized and improved by analyzing defects that occur in the substrates that are manufactured and feeding them back to the manufacturing process. Is planned. Then, an automatic defect classification apparatus for automatically analyzing such defects has been put into practical use. In this type of automatic defect classification device, several classification categories according to the type of defect such as appearance and cause of occurrence are prepared in advance, and the defect image obtained by imaging the target substrate with an imaging means such as an optical microscope is analyzed. Then, it is determined to which classification category the found defect belongs.

  In order to classify defects with high accuracy by the automatic defect classification apparatus, it is important to create a rule for determining a classification category in which the defect should be classified from the appearance characteristics of the imaged defect. Correspondence between the feature of the defect image and the classification category cannot be automatically performed in principle, and it is necessary for the operator to judge and input to the apparatus. The automatic defect classification apparatus can automatically classify various defect images by learning a classifier using teaching input data as teaching data.

  The present applicant has previously disclosed a technique for supporting such teaching input work by an operator (see Patent Document 1). In the classification support apparatus described in Patent Literature 1, the captured defect image is displayed on the screen, and the result of analyzing the characteristics is displayed on the screen, so that the displayed defect image should be classified into which classification category It is configured to support the operator's judgment as to whether or not. As a result, information contributing to the determination together with the defect image is displayed on the same screen, and the operator's workload is reduced.

Japanese Patent Laying-Open No. 2003-317082 (for example, FIG. 8)

  However, the operation itself of assigning various defect images to several classification categories is still left to the operator's visual judgment. For this reason, there is room for indeterminate elements such as judgment errors, especially in the case of defect images with unclear features, and indeterminate factors such as variations in judgment criteria for each operator. There was room for improvement in terms of sex. In addition, the operator's work of visually judging many defect images is still necessary.

  The present invention has been made in view of the above problems, and in a method of creating teacher data for automatically classifying a defect on a substrate, and a defect classification method and apparatus using the same, it has excellent reproducibility and objectivity. The first object is to provide teacher data. A second object is to reduce an operator's load in creating the teacher data.

  In order to achieve the above object, the teacher data creation method for automatic defect classification according to the present invention collects a plurality of defect images obtained by imaging the defects of the substrate, and for each classification category representing the type of the defect, A setting step of setting at least one of the defect images as a teacher image corresponding to the classification category; a feature amount calculation step of calculating a plurality of feature amounts for each of the defect images; and the unclassified defect image A temporary classification step of calculating a distance between the defect image and the teacher image in the feature amount space formed by the plurality of feature amounts, and temporarily classifying the defect image into the classification category according to the distance; With respect to each of the defect images thus obtained, the classifiers configured by the defect images extracted from the respective classification categories are transferred to the classification categories. And a re-classification step for reclassification is characterized by creating teacher data by repeating said reclassification process and the provisional classification step to classify the results does not change for all of the defective image.

  In the invention configured as described above, if the operator designates at least one teacher image for each classification category, with respect to other defect images, the feature amount of the defect image and the feature amount of the designated teacher image. Therefore, the classification can be performed objectively and automatically, and the operator does not have to make a judgment individually. Therefore, variations and uncertainties due to the subjectivity and misjudgment of the operator are eliminated, and the load on the operator is reduced. In addition, the provisional classification based on the distance between the given teacher image and each defect image and the reclassification using the classifier for the result of the provisional classification are repeated to converge the classification result, and the teacher has high reproducibility and objectivity. Data can be obtained.

  In the present invention, for example, in the reclassification step, when the defect image is classified into the same classification category in the reclassification and the temporary classification, the classification result for the defect image is determined, while the reclassification and the When the classification results in the temporary classification are different, the defect image may be set in an unclassified state. By doing so, even if an inappropriate classification result is obtained in the temporary classification, it can be corrected to increase the objectivity of the data.

  Further, for example, in the reclassification step, the defect image for which the classification result is confirmed may be additionally set as the teacher image. Thus, the accuracy of classification can be further increased by gradually increasing the number of teacher images.

  Further, in this teacher data creation method, for example, based on the feature amount of the defect image whose classification result is determined by the reclassification step, the classification result is greatly influenced from the plurality of feature amounts constituting the feature amount space. A feature amount selecting step for selecting effective feature amounts to be given, and the defect image is classified based on the selected effective feature amounts in the temporary classification step and the reclassification step after the feature amount selection step. It may be. As the feature amount representing the feature of the defect image, for example, various things such as luminance, texture information, and the size of an area satisfying a predetermined condition can be considered. However, it is not always necessary to classify defects, and all of the possible features are not necessarily required, and classification with a high degree of accuracy can be performed even with classification that mainly uses features that characterize defects. There are cases where it is possible. In such a case, if the classification is performed using only the feature amount that greatly affects the feature of the defect, the processing time can be significantly shortened.

  According to a first aspect of the automatic defect classification method of the present invention, in order to achieve the above object, learning is performed based on teacher data created by any one of the teacher data creation methods described above, A learning step for obtaining classification data indicating a correlation with the classification category, a defect candidate image is obtained from the inspection target substrate, a feature amount thereof is calculated, and the defect candidate image is calculated based on the calculation result and the classification data. And a defect classification step of classifying into any of the classification categories.

  Further, in order to achieve the above object, the automatic defect classification device according to the present invention performs learning based on teacher data created by any one of the teacher data creation methods described above, and calculates the feature amount and the classification category. Learning means for obtaining classification data indicating correlation, imaging means for imaging a substrate to be inspected to obtain a defect candidate image, calculating a feature amount of the defect candidate image, and based on the calculation result and the classification data And defect classification means for classifying the defect candidate image into any of the classification categories.

  In these inventions, defect candidate images acquired on the inspection target substrate are classified based on classification data obtained by learning based on the teacher data created by the teacher data creation method described above. Highly accurate defect classification can be performed automatically.

  According to a second aspect of the automatic defect classification method of the present invention, in order to achieve the above object, learning is performed based on teacher data created by the teacher data creation method including the feature quantity selection step described above, and the feature A learning step for obtaining classification data indicating a correlation between a quantity and the classification category, a defect candidate image is obtained from the inspection target substrate, the effective feature quantity is calculated for the image, and the calculation result and the classification data are And a defect classification step of classifying the defect candidate image into any one of the classification categories.

  In the invention configured as described above, the teacher data created by the above-described teacher data creation method is applied, and the defects are classified based on the effective feature amount among the feature amounts of the defect candidate image acquired on the inspection target substrate. Therefore, defect classification with high reproducibility and objectivity can be performed automatically, and classification can be performed in a short time.

  According to the teacher data creation method for automatic defect classification according to the present invention, it is possible to obtain teacher data having high reproducibility and high objectivity without variations due to the subjectivity of the operator. In addition, according to the automatic defect classification method and the automatic defect classification apparatus according to the present invention, defect classification is performed based on the classification data learned based on the teacher data having high reproducibility and objectivity obtained in this manner. It is possible to automatically perform highly objective defect classification.

It is a figure showing a schematic structure of an inspection system as one embodiment of an automatic defect classification device concerning this invention. It is a flowchart which shows the automatic defect classification | category process in this inspection system. It is a flowchart which shows a teacher data creation process. It is a figure which shows the mode of provisional classification typically. It is a figure which shows an ensemble classifier typically. It is a figure which shows typically the classification state after temporary classification | category and a reclassification operation | work. It is a flowchart which shows the optimization process of feature-value space.

  FIG. 1 is a diagram showing a schematic configuration of an inspection system as an embodiment of an automatic defect classification apparatus according to the present invention. The inspection system 1 is an inspection system that performs defect inspection of an inspection target and automatically classifies detected defects. The inspection system 1 performs an inspection of defects based on image data from the imaging device 2 and the imaging device 2 that images an inspection target region on a semiconductor substrate (hereinafter referred to as “substrate”) W, and a defect when a defect is detected. And an inspection / classification device 4 that automatically classifies defects into categories to which the image belongs, and a host computer 5 that controls the overall operation of the inspection system 1. The imaging device 2 is incorporated in the production line of the substrate W, and the inspection system 1 is a so-called inline system.

  The imaging device 2 moves the stage 22 relative to the imaging unit 21 that acquires image data by imaging the inspection target area on the substrate W, the stage 22 that holds the substrate W, and the imaging unit 21. The imaging unit 21 has a stage driving unit 23, and an imaging unit 21 forms an image by an illumination unit 211 that emits illumination light, an optical system 212 that guides the illumination light to the substrate W and that receives light from the substrate W, and an optical system 212. The image pickup device 213 converts the image of the substrate W thus formed into an electrical signal. The stage driving unit 23 includes a ball screw, a guide rail, and a motor. The host computer 5 controls the stage driving unit 23 and the imaging unit 21 so that the inspection target area on the substrate W is imaged.

  The inspection / classification apparatus 4 includes a defect inspection processing unit 41 that performs defect inspection while processing image data of an inspection target region, and a defect classification unit 42 that classifies defects. The defect inspection processing unit 41 has a dedicated electrical circuit that processes 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 or by image processing. That is, the imaging device 2 and the defect inspection processing unit 41 play a role as an inspection device in the inspection system 1. When the defect inspection processing unit 41 detects a defect from the inspection target area, image data of the defect and various data used for the inspection are temporarily stored in the memory of the defect inspection processing unit 41.

  On the other hand, the defect classification unit 42 includes a CPU that performs various arithmetic processes, a memory that stores various types of information, and the like. Software that performs a process for classifying detected defects using a neural network, a decision tree, discriminant analysis, or the like. Run it. In addition to the role as the control of the inspection system 1, the host computer 5 also generates various parameters (that is, classification data) used for classification performed by the defect classification unit 42.

  The host computer 5 has a general computer system configuration having a processing unit 51 in which a CPU for performing various arithmetic processes, a ROM for storing basic programs, a RAM for storing various information, and the like are connected to a bus line. As shown in FIG. 1, the processing unit 51 configured as described above further includes a display unit 52 such as a display for displaying various information such as images, an input such as a keyboard and a mouse for receiving input from the operator. The receiving unit 53, although not shown, transmits / receives a signal to / from another configuration of the inspection system 1 and a reading device that reads information from a computer-readable recording medium such as an optical disk, a magnetic disk, or a magneto-optical disk. The communication units to be connected are appropriately connected through an interface (I / F) or the like.

  The classification data is generated by learning, and the host computer 5 performs the classification work (teaching work) in which the operator inputs the defect category (classification) to the image displayed on the host computer 5, whereby the host computer 5 performs the teacher data. Further, learning is performed, and the learning result is output to the defect classification unit 42 as classification data used for automatic classification. The host computer 5 has a large number of functions for assisting an operator in classification work (ie, category input), and also serves as a classification support apparatus.

  FIG. 2 is a flowchart showing an automatic defect classification process in this inspection system. In this process, first, the host computer 5 acquires teacher data stored in advance in a recording medium (step S101), and based on this, a learning operation is executed (step S102). As this learning process, for example, a known method such as a support vector machine (SVM), a discriminant analysis method, or a method using a neural network can be applied. The classification data obtained by learning is stored in the RAM of the host computer 5 or an appropriate recording medium, and is also transmitted to the defect classification unit 42.

  When learning is completed, defect detection / classification work is started. First, the imaging device 2 images the surface of the substrate W to be inspected (step S103). Then, the defect inspection processing unit 41 performs defect detection on the captured image (step S104). Hereinafter, an image that may contain a defect is referred to as a “defect candidate image”.

  When a defect candidate image is detected, the defect classification unit 42 receives an image (defect candidate image) of the inspection target area from the defect inspection processing unit 41 and data (difference image, reference image, etc.) obtained during the inspection process. ), And the feature amount of the defect candidate image is calculated based on the data (step S105). The feature amount may be calculated in the defect inspection processing unit 41, and the region in which the feature amount is calculated is appropriately determined by the defect inspection processing unit 41 or the defect classification unit 42.

  Subsequently, the detected defect is automatically classified (step S106). That is, the feature amount and various data are input to the classifier of the defect classification unit 42 in which the learning result is input as classification data from the host computer 5 in advance, and the defect classification unit 42 outputs the classification result. Specifically, which of the plurality of classification categories set in advance according to the feature of the defect corresponds to the defect candidate image is output as a result. Note that the feature amount refers to a value obtained by processing the pixel value according to a predetermined rule, and most of the feature amount is obtained by applying some kind of filter to the image. For example, an index value indicating some feature of the image such as the average luminance of the image, texture information, the size of an area satisfying a predetermined condition, the amount of an edge to be extracted, and the like is used as the feature amount. A feature amount based on higher-order local autocorrelation using a plurality of autocorrelation filters may be used.

  In the inspection system 1, every time a defect is detected by the defect inspection processing unit 41, the operations shown in steps S104 to S106 in FIG. 2 are performed in real time, and a large number of images indicating defects are classified at high speed. The defect candidate images thus detected are classified into corresponding classification categories.

  Next, a description will be given of how the host computer 5 supports the defect image classification work by the operator in the teaching work that is the preliminary preparation of the classification by the defect classification unit 42. In the inspection system 1, the imaged defect candidate images are classified into several classification categories (for example, chipping, pinholes, foreign matter, etc.) according to features such as the shape and the cause of occurrence. For this purpose, it is necessary to obtain various defect image data in advance and classify them into the above-mentioned classification category and give it to the host computer 5 as teacher data.

  In this case, it is not necessary for the operator to perform classification work for all defect images. That is, in this embodiment, if an operator teaches at least one representative defect image to be classified into each classification category as a typical example of a teacher image from among a large number of defect images, Then, the host computer 5 automatically performs classification based on the given teacher image data, thereby creating teacher data.

  FIG. 3 is a flowchart showing teacher data creation processing. More specifically, this processing is realized by the host computer 5 executing a predetermined program, and is processing for receiving teaching input from an operator and creating teacher data.

  In this process, first, various types of defect image data are collected as information for creating teacher data (step S201). As the defect image data, data that has been captured in advance and stored in a recording medium of the host computer 5 may be used, and each part of the sample substrate placed on the stage 22 of the inspection system 1 in FIG. A newly obtained image may be used. In order to increase the accuracy of classification, it is desirable to prepare a large number of defect image data.

  These defect images are displayed side by side or sequentially on the display unit 52, and a teaching input by the operator is accepted. More specifically, the operator selects at least one image including typical defects to be classified into the classification category for each classification category prepared in advance from the displayed various defect images. Teaching input is performed from the input receiving unit 53 as a representative teacher image corresponding to the classification category. Thereby, one or more representative teacher images are set for each classification category (step S202).

  Thus, after receiving the teaching input from the operator, the host computer 5 basically classifies other defect images. That is, a plurality of feature amounts are calculated for each defect image including the representative teacher image (step S203), and each defect image other than the representative teacher image is compared with the representative teacher image in the feature amount space including a plurality of feature amounts. Based on the Euclidean distance, they are provisionally classified into each classification category (step S204). Specifically:

  FIG. 4 is a diagram schematically showing the provisional classification. In FIG. 4, “x” marks indicate the feature vector of each defect image distributed in the feature space FS. Here, it is assumed that three categories A, B, and C are set as classification categories. When the teaching input by the operator is finished, as shown in FIG. 4A, in the feature quantity space FS, a teacher image (representative teacher image) A0 as a typical example of the classification category A and a typical example of the classification category B Only the teacher image B0 and the teacher image C0 as a typical example of the classification category C are in a state where the classification is defined. In the figure, black circles indicate defect images whose classification has been determined.

  In the temporary classification, as shown in FIG. 4B, for each of the teacher images A0, B0, and C0, only defect images that are within a predetermined distance from the teacher image at the Euclidean distance are classified into the same classification category as the teacher image. Temporary classification. That is, the defect images A1, A2 and A3 within the range of the distance Ra from the teacher image A0 are provisionally classified into the category A. Similarly, the defect images B1 and B2 within the range of the distance Rb from the teacher image B0 are provisionally classified into the category B. Further, the defect images C1 and C2 within the range of the distance Rc from the teacher image C0 are provisionally classified into the category C. These distances Ra, Rb, and Rc are predetermined according to the nature of the defect that can occur, and may be the same value or different values.

  At this point, it is desirable to apply relatively strict criteria, and there may be many defect images that are not classified into any classification category (unclassified). As a determination criterion for this, in addition to the method of temporarily classifying only those whose Euclidean distance from the teacher image is within a predetermined range, for example, the Euclidean distance from the teacher image belonging to the classification category A is A method of classifying an image that is less than half of the Euclidean distance from the teacher images belonging to the classification categories B and C into the classification category A can be used.

  Subsequently, reclassification work is performed based on the temporary classification result thus obtained (steps S205 to S209). In the reclassification operation, whether or not the result of the temporary classification is appropriate is evaluated by comparing the Euclidean distance in the feature amount space FS between the temporarily classified defect images (step S205), and according to the evaluation result. (Step S206), whether to confirm the classification result (step S207) or discard the classification result and return to the unclassified state (step S208). As a specific example, as will be described below, an ensemble classifier can evaluate a provisionally classified defect image.

  FIG. 5 is a diagram schematically showing an ensemble classifier. As shown in FIG. 5 (a), the ensemble classifier 10 extracts N weak classifiers 11k (k = 1, 2, and 2) configured by extracting the temporarily classified defect images one by one from each classification category. .., N) and a majority decision judgment unit 12 for taking a majority decision of judgment results by these N weak classifiers 11k. These configurations are generally realized by software, but are not limited thereto.

  Each weak classifier 11k includes one defect image Al extracted from the defect image provisionally classified into category A, one defect image Bm extracted from the defect image provisionally classified into category B, and provisional category C. One defect image Cn extracted from the classified defect image (l, m, n are parameters for specifying the defect image). A symbol X represents a defect image to be evaluated. The weak classifier 11k has a high probability that the defect image X of the classification categories A, B, and C should be classified with respect to the defect image X as an input. Returns one as output.

  As an evaluation rule for the weak classifier 11k, for example, the one shown in FIG. 5B can be adopted. In this example, the Euclidean distances Da, Db, and Dc in the feature amount space FS between the defect image X to be evaluated and the respective defect images Al, Bm, and Cn constituting the weak classifier 11k are obtained, respectively. The classification category to which the defect image with the smallest Euclidean distance belongs is set as the classification category to which the defect image X should belong. In the example of FIG. 5B, since Da is the smallest with respect to Db and Dc, the output is A, that is, an evaluation result that the probability that the defect image X belongs to the classification category A is high.

  When the defect image X is evaluated by each of the N weak classifiers 11k (111 to 11N) obtained by extracting and combining the temporarily classified defect images one by one from each classification category, the defect to be extracted It is considered that the evaluation results vary depending on the combination of images. Therefore, by taking a majority vote for the evaluation results of these weak classifiers 11k by the majority decision determiner 12, it is obtained as an output to which classification category the defect image X should be classified. As for the accuracy, it is experimentally known that the determination result almost converges if N = 200 or so.

For example, the evaluation can be performed by a calculation formula as follows. Each weak classifier 11k includes the distances Da (k), Db (k), and the like in the feature amount space FS between the input defect image X and the respective defect images Al, Bm, and Cn constituting the weak classifier 11k. Dc (k) is obtained individually. And the value obtained by cumulatively adding the above distances for each classification category, that is,
Σ _k Da (k), Σ _k Db (k), Σ _k Dc (k)
And the classification category in which these values are the smallest can be set as the classification category in which the defect image X is to be classified.

  Returning to FIG. 3, the description of the teacher data creation process will be continued. The result of the evaluation for the temporarily classified defect image X as described above is compared with the result of the temporary classification (step S206). If the evaluation does not change and the two results match, the defect image X is It can be said that the probability of belonging to the classification category is extremely high. Therefore, the classification result for the defect image X is determined to be appropriate (step S207). Specifically, in the subsequent processing, the defect image X is regarded as one of the teacher images representing the classification category.

  On the other hand, if the result of the temporary classification and the evaluation result do not match, the result of the temporary classification for the defect image is invalidated and the defect image is returned to the unclassified state (step S208). In this way, once the defect images with uncertain evaluation are returned to the unclassified state, the defect images having ambiguous intermediate characteristics of the category to be classified are short-term classified into one of the classification categories. The classification is suspended.

  Note that when confirming and invalidating the classification, for example, a confirmation work by an operator may be performed before the execution. That is, a defect image for which the classification is to be confirmed or invalidated may be displayed on the display unit 52 so that the operator can cancel or change the decision or invalidation. In addition, confirmation work may be performed when the temporary classification is completed. In these cases, for example, the technique described in Patent Document 1 can be applied to assist the operator's work.

  When the evaluation for one defect image is completed in this way, all the results of the temporary classification are evaluated by repeating the processes of steps S205 to S208 described above until there is no unevaluated defect image after the temporary classification (step S209). Will be. At this time, each collected defect image is divided into one in which classification is confirmed (including one taught by the operator) in one of the classification categories and one that is not classified in any classification category.

  FIG. 6 is a diagram schematically showing the classification state after the temporary classification and reclassification operations. When temporary classification and reclassification are performed as described above, as shown in FIG. 6A, a defect image relatively close in the feature amount space FS from the teacher image (A0, B0, C0) initially taught by the operator. (A1 to A3, B1 to B2, C1 to C2) are each definitely classified into the same classification category as the teacher image. Based on the new teacher images (A0 to A3, B0 to B2, and C0 to C2) to which the defect images whose classification has been determined in this way are added, as shown in FIG. Then, the defect image A4 is newly classified into the classification category A, the defect images B3, B4 and B5 are classified into the classification category B, and the defect image C3 is temporarily classified into the classification category C. In this case, the Euclidean distances Ra ′, Rb ′, and Rc ′ in the determination of the provisional classification are not necessarily the same as the values Ra, Rb, and Rc in the previous classification, and may be changed as appropriate.

  As described above, when the above-described provisional classification and reclassification based on the evaluation are repeated for the unclassified defect images, the number of defect images in which the classification category is determined gradually increases. Finally, all the defect images are classified, or most of the defect images are classified and some of them are left unclassified. Is repeated, the classification result for each defect image is in a state of being completely coincident with the previous time. After this, even if the process is repeated, the result does not change and it can be said that the classification result has converged.

  Accordingly, when the reclassification is completed, it is determined whether or not the result has converged (step S210), and if it has converged, the process is terminated, and the classification result and each defect image data at this time are determined as teacher data. To do. On the other hand, if the result has not converged, the process returns to step S204, and the above-described provisional classification and reclassification are performed again. Prior to this, the following feature amount space optimization processing is performed (step S211). This optimization process is not indispensable for creating the teacher data, but this process classifies the feature quantities that make up the feature quantity space that contribute significantly to the classification and those that do not have much influence, and affects the classification. The feature space is optimized by omitting features that are not used. As a result, the dimension of the feature amount space can be reduced, and the number of feature amounts to be calculated is reduced in the subsequent processing, and the time required for the processing can be shortened. As described below, for example, a known genetic algorithm can be applied to this processing.

  FIG. 7 is a flowchart showing the optimization process of the feature amount space. In this processing, a reference image is selected from among defect images whose classification has already been determined by an appropriate sampling method, for example, a bootstrap sampling method (step S301), and for example, M is extracted from an M-dimensional (M is a natural number) feature amount space. Several smaller feature quantities are selected (step S302). Then, a feature quantity subset space described only by the selected feature quantity selected in this way is set (step S303), and the distance from the reference image of each defect image in the subset space is calculated to classify the classification category. This is performed (step S304). The classification by the ensemble classifier illustrated in FIG. 5 can also be applied to the classification work in this case.

  Thus, the correct answer rate is calculated by comparing the classification result classified based on the lower level (that is, the number of feature quantities is smaller) than the previous classification (temporary classification and reclassification) work with the previous classification result ( Step S305). As the feature amount used for classification is reduced, the correct answer rate may be less than 100%. In particular, if the feature amount that greatly contributes to characterizing the defect is excluded from the selected feature amount, the correct answer rate is expected to be greatly reduced.

  Until the correct answer rate converges (step S306), the feature amount selected based on the genetic algorithm is appropriately replaced (step S307). By repeating the above processing, the optimum feature value that ultimately yields the highest correct answer rate is obtained. Can be selected (step S308). If the necessary and sufficient correct answer rate is obtained by this combination, classification with sufficient accuracy can be performed by working in a lower-dimensional feature amount space consisting of only these feature amounts in the subsequent classification work. It becomes possible. In this way, the feature amount space is optimized. The feature amounts constituting the optimized feature amount space function as “effective feature amounts” that are substantially significant for classification.

  In the classification operation to be executed next, since the Euclidean distance in the optimized feature amount space has only to be obtained, the time required for processing is shortened. Further, since the optimized feature amount space can be applied to the automatic classification of the defect image in the defect classification unit 42, the defect detection and classification for the inspection target substrate W performed in real time by the inspection / classification device 4 is performed. It also leads to faster work.

  As described above, in this embodiment, when creating a teacher data by classifying a number of defect images into several categories, the host is based on a representative teacher image for each classification category taught by an operator. The computer 5 automatically classifies other defect images and creates teacher data. For this reason, the operator's work is greatly reduced. In addition, since the operator is basically required only to input a typical defect image for each classification category, there is no problem caused by variations in judgment or misjudgment for each operator. Teacher data with high reproducibility and objectivity can be created.

  In particular, in the teacher data creation process of this embodiment, classification is performed by repeatedly classifying a defect image based on the Euclidean distance from the teacher image in the feature amount space, evaluating the result by an ensemble classifier, and reclassifying the result. Are sequentially determined, so that highly accurate teacher data can be created without misclassification.

  In addition, since the defect images are automatically classified based on the teacher data having high reproducibility and objectivity created in this way, the inspection system 1 according to this embodiment performs automatic classification of defects detected on the inspection target substrate with accuracy. Can be done well.

  Furthermore, in this embodiment, since the feature amount space is optimized based on the classification result, only a large number (usually 100 or more) of feature amounts necessary for the classification are extracted, and this is extracted from the teacher data. By being used for creation and automatic defect classification, it becomes possible to perform processing at high speed while obtaining sufficient classification accuracy.

  As described above, in this embodiment, the host computer 5 functions as the “learning means” of the present invention, while the imaging device 2 and the inspection / classification device 4 are the “imaging means” and “defect classification means of the present invention, respectively. Is functioning.

  In the teacher data creation process (FIG. 3) of the above embodiment, steps S201, S202, S203, and S204 are respectively “collection process”, “setting process”, “feature amount calculation process”, and “temporary classification process” of the present invention. Steps S205 to S209 correspond to the “reclassification step” of the present invention. Step S211 corresponds to the “feature amount selection step” of the present invention.

  In the automatic defect classification process (FIG. 2) of the above embodiment, step S101 corresponds to the “learning process” of the present invention, and steps S103 to S106 correspond to the “defect classification process” of the present invention. Yes.

  The present invention is not limited to the above-described embodiment, and various modifications other than those described above can be made without departing from the spirit of the present invention. For example, as described above, the optimization of the feature amount space in the teacher data creation process is not essential and can be omitted. Also, the effective frequency is not optimized every time in the provisional classification and reclassification processing loop, but is optimized only a limited number of times, for example, when a predetermined number of defect images have been determined. May be performed.

  In the above embodiment, the teacher image taught by the operator is not subject to classification evaluation and the classification is not changed, but the present invention is not limited to this, and the teacher image taught is also evaluated. Reclassification may be performed.

  In the above embodiment, the defect image is classified by an ensemble classifier combining a plurality of weak classifiers, the reference image is selected by bootstrap sampling, and the feature space is optimized by a genetic algorithm. However, these are merely examples of processing methods, and methods other than those described above may be employed.

  Further, in the above embodiment, automatic classification (in learning of the inline type defect image) is performed from the viewpoint that the classification work in generating the teacher data can be performed by using the data obtained in the inspection. However, the present invention can also be applied to a so-called off-line type classification apparatus.

  The present invention is suitably applied to a classification device and a classification method for detecting and classifying defects in various substrates such as a semiconductor substrate, a glass substrate for a liquid crystal display device, and a printed wiring board, and a technical field for creating teacher data therefor. be able to.

2 Imaging device (imaging means)
4 inspection / classification equipment (defect classification means)
5 Host computer 5 (learning means)
S201 Collection step S202 Setting step S203 Feature amount calculation step S204 Temporary classification step S205 to S209 Reclassification step S211 Feature amount selection step S101 Learning step S103 to S106 Defect classification step W substrate

Claims (7)

A collection step of collecting a plurality of defect images obtained by imaging defects on the substrate;
For each classification category representing the type of defect, a setting step of setting at least one of the defect images as a teacher image corresponding to the classification category;
A feature amount calculating step for calculating a plurality of feature amounts for each of the defect images;
For the unclassified defect image, a distance between the defect image and the teacher image in the feature amount space formed by the plurality of feature amounts is calculated, and the defect image is provisionally classified into the classification category according to the distance. A classification process;
A reclassification step for reclassifying the classification image by the classifier constituted by the defect image extracted from each of the classification categories for each of the temporarily classified defect images,
A teacher data creation method for automatic defect classification, wherein teacher data is created by repeating the provisional classification step and the reclassification step until classification results for all of the defect images no longer change.   In the reclassification step, when the defect images are classified into the same classification category in the reclassification and the temporary classification, a classification result for the defect image is determined, while a classification result in the reclassification and the temporary classification The teacher data creation method according to claim 1, wherein the defect images are placed in an unclassified state when the two differ.   The teacher data creation method according to claim 2, wherein in the reclassification step, the defect image whose classification result is determined is additionally set as the teacher image. A feature amount selection step of selecting an effective feature amount that greatly affects the classification result from a plurality of feature amounts constituting the feature amount space, based on the feature amount of the defect image whose classification result is determined by the reclassification step. With
4. The teacher data generation method according to claim 1, wherein in the temporary classification step and the reclassification step after the feature amount selection step, the defect image is classified based on the selected effective feature amount. . A learning step of performing learning based on the teacher data created by the teacher data creation method according to any one of claims 1 to 4 to obtain classification data indicating a correlation between the feature quantity and the classification category;
A defect classification step of acquiring a defect candidate image from an inspection target substrate, calculating a feature amount thereof, and classifying the defect candidate image into one of the classification categories based on the calculation result and the classification data. Feature automatic defect classification method. A learning step of performing learning based on teacher data created by the teacher data creating method according to claim 4 to obtain classification data indicating a correlation between the feature quantity and the classification category;
A defect classification step of obtaining a defect candidate image from an inspection target substrate, calculating the effective feature amount for the image, and classifying the defect candidate image into one of the classification categories based on the calculation result and the classification data; An automatic defect classification method comprising: Learning means for performing learning based on the teacher data created by the teacher data creation method according to any one of claims 1 to 4 and obtaining classification data indicating a correlation between the feature quantity and the classification category;
Imaging means for capturing a defect candidate image by imaging a substrate to be inspected;
An automatic defect classification device comprising: defect classification means for calculating a feature amount of the defect candidate image and classifying the defect candidate image into any of the classification categories based on the calculation result and the classification data .

Images