JP6113024B2 - Classifier acquisition method, defect classification method, defect classification device, and program - Google Patents

Description

  The present invention relates to a technique for classifying defect images indicating defects, foreign matters, and the like included in a fine pattern into defect types.

  In manufacturing a semiconductor substrate, a glass substrate, a printed wiring board, and the like, a fine pattern is formed on the substrate. When such a fine pattern is formed, a defect may occur due to dust or the like being attached to the substrate or being scratched. Such a defect is detected using, for example, an optical microscope or a scanning electron microscope. In order to prevent a decrease in yield due to defects as much as possible, it is preferable to perform detailed analysis on the detected defects to identify the cause of the occurrence of the defects and to take measures against the manufacturing equipment and the manufacturing process.

  The appearance inspection is performed every several series of steps. In the appearance inspection, for example, a reference image and an inspection image are acquired from patterns of two regions originally formed to have the same shape, and defect detection is performed by comparing these images. At the same time, the feature amount of the defective portion is obtained, and the defect is automatically classified based on the feature amount. As the feature amount, for example, the area of the defect, the brightness average, the peripheral length, the flatness, the inclination of the major axis when the defect is approximated to an ellipse, and the like are employed.

  As the automatic defect classification method, there are a rule-based type and a teaching type. In the rule-based type, defects are classified according to a predetermined rule. A rule is often expressed as a threshold for a feature amount. In the teaching type, teaching is performed in which a feature quantity vector that is a set of feature quantities is associated with a defect type. Then, the classifier is automatically acquired using the teacher data that is a pair of the feature vector and the teaching information. In teaching-type defect classification, the classifier itself is a black box, so even if the classification accuracy is unsatisfactory, the user can reselect the classifier generation algorithm from several types of algorithms and re-acquire the classifier It is.

  As a method of improving the accuracy of teaching type defect classification, conventionally, a plurality of classifiers are arranged in parallel, and a comprehensive judgment is made using the classification results of these classifiers (for example, FIG. 6), and a technique for providing a classifier in multiple stages and sequentially performing rough classification and detailed classification (for example, see FIG. 7 of Patent Document 1) has been proposed.

JP 2012-181209 A

  By the way, in the conventional classification method, as the accuracy is increased, the processing becomes more complicated. Further, in some cases, the limit of accuracy obtained even if it is complicated does not satisfy the required accuracy.

  The present invention has been made in view of the above problems, and an object thereof is to obtain a highly accurate classifier by a simple process.

  The invention according to claim 1 is a classifier acquisition method for obtaining a classifier for classifying pattern defects indicated by a defect image into N defect types, wherein a) one primary classifier and the N number of classifiers Preparing N classifiers, each of which is associated with each of the defect types, wherein each of the N secondary classifiers is connected under the primary classifier; b A) preparing a plurality of defect images; c) associating a teaching defect type with each of the plurality of defect images; and d) all or part of the plurality of defect images and associated with them. Learning the primary classifier using a taught defect type, e) classifying all or some of the plurality of defect images by the primary classifier to obtain a primary classification type, and f) All or one of the defect images classified in step e) A step of obtaining the classifier by associating a combination teaching defect type obtained by combining the teaching defect type and the primary classification type, and learning the primary classifier and the N secondary classifiers. Prepare.

  Invention of Claim 2 is the classifier acquisition method of Claim 1, Comprising: After the said f) process, all or one part of the defect image used for learning is used for the said classifier. To obtain a secondary classification type and further obtain a combination classification type obtained by combining the secondary classification type and the primary classification type, and h) a primary classification type included in each combination teaching defect type, Using the step of changing to the primary classification type included in the combination classification type, i) all or a part of the defect images classified in the step g) and the changed combination teaching defect type associated therewith. And re-learning the classifier.

  The invention according to claim 3 is the classifier acquisition method according to claim 2, wherein j) the defect image and the combination after the step i) or after the steps g) to i) are repeated. A step of displaying a primary classification type included in the classification type.

  Invention of Claim 4 is the classifier acquisition method of Claim 2 or 3, Comprising: By performing said g) process, each defect image contains the teaching defect contained in a combination teaching defect classification Teach-class information in which a type, a primary classification type included in the combination teaching defect type, a secondary classification type included in the combination classification type, and a primary classification type included in the combination classification type are sequentially associated is associated with each other. In the step h), the combination teaching is performed by deleting the primary classification type included in the combination teaching defect type and the secondary classification type included in the combination classification type from the teaching-classification information. The teaching defect type included in the defect type and the primary classification type included in the combination classification type are obtained as the combined teaching defect type after the change.

  The invention according to claim 5 is a defect classification method for classifying the pattern defect indicated by the defect image into a plurality of defect types, the step of preparing the defect image, and the classification according to any one of claims 1 to 4 And classifying the defect image using a classifier acquired by a container acquisition method.

  The invention according to claim 6 is a defect classification device for classifying pattern defects indicated by a defect image into a plurality of defect types, an image storage unit for storing defect image data, and any one of claims 1 to 4. And a classification control unit that classifies the defect image using the classifier.

  The invention according to claim 7 is a program for obtaining, using a computer, a classifier that classifies pattern defects indicated by a defect image into N defect types, and the execution of the program by the computer is performed on the computer, a) including one primary classifier and N secondary classifiers respectively associated with the N defect types, wherein each of the N secondary classifiers is the primary classifier Preparing a classifier connected below; b) preparing a plurality of defect images; c) associating a teaching defect type with each of the plurality of defect images; and d) the plurality of the defect images. E) learning the primary classifier using all or part of the defect images and teaching defect types associated therewith, and e) using the primary classifier to obtain all or part of the plurality of defect images. Classified and primary A step of obtaining a classification type, and f) a combination teaching defect type combining a teaching defect type and a primary classification type is associated with all or a part of the defect images classified in the step e), The step of obtaining the classifier is performed by learning the classifier and the N secondary classifiers.

  According to the present invention, it is possible to acquire a highly accurate classifier by simple processing.

It is a figure which shows schematic structure of a defect classification apparatus. 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 structure of a classifier. It is a figure which shows the flow of a classifier acquisition. It is a figure which shows the flow of a classifier acquisition. It is a figure which illustrates the screen at the time of teaching. It is a figure which illustrates a confusion matrix. It is a figure which illustrates the screen at the time of performing re-teaching. It is a figure which illustrates a confusion matrix. It is a figure which illustrates the reduced confusion matrix. It is a figure which illustrates the confusion matrix after repeating learning.

  FIG. 1 is a diagram showing a schematic configuration of a defect classification apparatus 1 according to an embodiment of the present invention. In the defect classification apparatus 1, a defect image indicating a pattern defect on a semiconductor substrate 9 (hereinafter simply referred to as “substrate 9”) is acquired, and the defect image is classified. The defect classification apparatus 1 includes an imaging apparatus 2 that images an inspection target area on the substrate 9, an inspection / classification apparatus 4 that automatically classifies defects, and a host computer 5. The inspection / classification device 4 performs defect inspection based on the image data from the imaging device 2 and, when a defect is detected, the defect type to which the defect should belong (also called “class”, “category”, etc.). Classify defects. The host computer 5 controls the overall operation of the defect classification apparatus 1 and generates a classifier 422 used for defect classification in the inspection / classification apparatus 4. The types of defect of the pattern existing on the substrate 9 are, for example, a defect, a protrusion, a disconnection, a short, and a foreign substance. Further, the imaging device 2 is incorporated in the production line of the substrate 9, and the defect classification device 1 is a so-called inline system. The defect classification apparatus 1 can also be regarded as an apparatus in which a function of automatic defect classification is added to a defect inspection apparatus.

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

  The inspection / classification apparatus 4 includes a defect detection unit 41 and a classification control unit 42 that classifies defect images. The defect detection unit 41 detects defects while processing the image data of the inspection target area. 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 classification control unit 42 includes a CPU that performs various arithmetic processes, a memory that stores various types of information, and the like, and includes a feature amount calculation unit 421 and a classifier 422. The classifier 422 executes defect classification, that is, defect image classification, using a neural network, a decision tree, discriminant analysis, or the like.

  FIG. 2 is a diagram showing a flow of defect image classification by the defect 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 defect inspection of the inspection target region and a defect is detected, data of a defect image that is an image of the defective portion is generated and prepared (step S12). The defect image data is transmitted to the classification control unit 42. The feature amount calculation unit 421 of the classification control unit 42 calculates a feature amount vector that is an array of a plurality of types of feature amounts of the defect image (step S13), and the feature amount vector is input to the classifier 422 of the classification control unit 42. The classification result is output. That is, the defect image is classified into one of a plurality of defect types using the classifier 422 (step S14). In the defect classification device 1, every time a defect is detected by the defect detection unit 41, the feature amount vector is calculated in real time, and a large number of defect images are automatically classified at high speed.

  Next, acquisition of a classifier 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 defect classification device 1. Are appropriately connected through an interface (I / F) or the like.

  The host computer 5 reads in advance the program 80 from the recording medium 8 via the reading device 57 and stores it in the fixed disk 54. Further, the CPU 51 calculates in accordance with the program 80 using the RAM 53, the fixed disk 54, and the like. Processing is executed.

  FIG. 4 is a block diagram showing a functional configuration for learning the classifier, which is realized by the CPU 51, ROM 52, RAM 53, fixed disk 54, etc. of the host computer 5. FIG. 4 also shows an inspection / classification device 4. The host computer 5 includes a classification control unit 61 including a classifier 611 and a learning unit 62 that causes the classifier 611 to learn. The classifier 611 indicates a structure that stores information necessary for performing classification in a predetermined storage area. The same applies to the classifier 422 of the inspection / classification apparatus 4.

  The host computer 5 further includes an image storage unit 64 and an information storage unit 65. The image storage unit 64 stores defect image data 801 that is data of each defect image and a feature quantity vector 802. The defect image data 801 corresponding to each defect image and the feature quantity vector 802 are associated with each other. The information storage unit 65 stores a teaching defect type 811 associated with each defect image, a combined teaching defect type 812, a primary classification type 821, a secondary classification type 822, and a combination classification type 823.

  As described above, the feature quantity vector 802 is an array of a plurality of types of feature quantities obtained from each defect image. The teaching defect type 811 is a defect type assigned to each defect image by the user. That is, the teaching defect type 811 is information indicating the result of teaching work that associates the type of foreign matter, the type of scratch, the type of pattern defect, and the like with each defect image. The combination teaching defect type 812 will be described later. The primary classification type 821, the secondary classification type 822, and the combination classification type 823 are defect types indicated by the classification result of each defect image obtained during or after generation by the classifier 611.

  When the classifier 611 is acquired by the host computer 5, the classifier 611 is transferred to the classifier 422 of the inspection / classification apparatus 4. Of course, the function of the host computer 5 can be included in the inspection / classification apparatus 4.

  FIG. 5 is a diagram illustrating the structure of the classifier 611. The classifier 611 includes one primary classifier 631 and a plurality of secondary classifiers 632. Each secondary classifier 632 is associated with one defect type output from the primary classifier 631. That is, when the number of defect types to be classified by the classifier 422 of the inspection / classification device 4 is N (where N is an integer equal to or greater than 2), N secondarys respectively associated with the N defect types. A classifier 632 is prepared. More precisely, a storage area for storing N secondary classifiers 632 is prepared. Each of the N secondary classifiers 632 is connected under the primary classifier 631. “Connected below” means that information indicating the classification result output from the primary classifier 631 (that is, the classified defect type) is input to the secondary classifier 632. Thus, the classifier 611 has a tree structure.

  Various methods or mechanisms for generating each of the primary classifier 631 and the secondary classifier 632 may be used as long as they are teaching types. For example, nearest neighbor method, minimum distance method, subspace method, decision tree method, linear discriminant method, fuzzy voting method, flexible naive Bayes method, neural network, support vector machine (SVM), etc. can be adopted. Each of the primary classifier 631 and the plurality of secondary classifiers 632 may be a different type of classifier.

  6 and 7 are diagrams showing the flow of classifier acquisition by the host computer 5. Acquisition of a classifier means that a classifier is generated by assigning a value to a variable included in the classifier or determining a structure. First, a classifier 611 in which variable values and structures are not set is prepared (step S21). Specifically, a predetermined storage structure is prepared in the storage space of the host computer 5 for storing the variable values, structure, etc. of the classifier 611.

  On the other hand, separately from step S 21, data of a large number of defect images detected by the inspection / classification device 4 is input to the host computer 5 and stored as defect image data 801 in the image storage unit 64. Thereby, a plurality of defect images are prepared (step S22). The defect detection unit 41 of the inspection / classification apparatus 4 obtains the feature vector of the defect image when detecting the defect. Therefore, the feature vector is also input to the host computer 5 together with the defect image data and stored in the image storage unit 64. Is done. Note that the stored feature quantity vector 802 may be generated again by the host computer 5 based on the defect image data 801.

  Next, teaching is performed by the user (step S23). That is, when the input unit 56 of the host computer 5 receives a teaching input from the user, the defect type is associated with each of the plurality of defect images. The associated defect type is stored in the information storage unit 65 as a taught defect type 811. Since each defect image is associated with the feature vector 802, step S23 is substantially an operation for associating the feature vector 802 with the defect type. FIG. 8 is a diagram illustrating a screen of the display 55 of the host computer 5 when teaching is performed. Five defect types “A”, “B”, “C”, “D”, and “E” are prepared as folders, and the defect image 803 is dragged and dropped into one of the folders, thereby teaching the defect image 803. Done.

  When the teaching work is completed, the primary classifier 631 learns using all or a part of the plurality of defect images for which teaching has been performed and the teaching defect type associated therewith (step S24). That is, a defect image data 801, a feature vector 802, and a teaching defect type 811 are prepared for each defect image, and the primary classifier 631 learns using a plurality of combinations as teacher data. As a result, the values of variables in the primary classifier 631, undecided structures, and the like are determined.

  Next, the defect image is classified by the classification control unit 61 using the primary classifier 631 (step S25). The classification result by the primary classifier 631 is temporarily stored in the information storage unit 65 as the primary classification type 821. The defect image to be classified is normally used at the time of learning by the primary classifier 631, but other defect images may be used as long as they are taught. That is, all or part of the defect images prepared and taught in advance are classified.

  FIG. 9 is a diagram illustrating a confusion matrix (confusion matrix) that summarizes the classification results. “A” to “E” in the leftmost “MAN” column indicate teaching defect types designated by the operator. “ADC” indicates an automatic defect classifier, and “A” to “E” in the top row indicate primary classification types. The number of times that the defect image that should belong to the teaching defect type “A” is determined to belong to the primary classification type “B” is indicated at the position (A row, B column).

  The value in the column enclosed by the bold frame is the number of correctly classified defect images. The value in the other column is the number of defect images that are misclassified. In the example of FIG. 9, 71.5% of defect images are correctly classified. As the performance of the classifier 611, for example, when a correct answer rate of 85% is required, the primary classifier 631 alone cannot satisfy the request, and the process proceeds to the next step.

  In the host computer 5, a combination of the teaching defect type and the primary classification type is prepared as a combination defect type. For example, a combination defect type whose teaching defect type is “A” and whose primary classification type is “A” is expressed as “aA”. A combination defect type whose teaching defect type is “A” and whose primary classification type is “B” is expressed as “aB”. This is a combination defect type that is taught as “A” but is misclassified as “B” by the primary classifier 631. From the taught defect type “A”, the combination defect types “aA”, “aB”, “aC”, “aD”, and “aE” are derived as lower classes.

Similarly, a combination defect type whose teaching defect type is “B” and whose primary classification type is “A” is expressed as “bA”. From the taught defect type “B”, combination defect types of “bA”, “bB”, “bC”, “bD”, and “bE” are derived. As described above, 25 combination defect types “aA” to “eE” are prepared. Generally speaking, when the number of defect types is N, N ² combination defect types are prepared.

  Then, teaching for associating the combination defect type corresponding to each defect image is performed (step S26). For example, “aA” is taught to a defect image classified as “A” by teaching “A”. A defect image taught as “A” but classified as “B” is taught “aB”. That is, information indicating that the defect image is classified as “B” by the primary classifier 631 but should be classified as “A” is given to the defect image. A defect image taught as “B” but classified as “A” is taught as “bA”. The combination defect type associated with the defect image is stored as the combination teaching defect type 812 in the information storage unit 65. Since this teaching is a simple process automatically performed by the host computer 5, it need not be regarded as teaching. In other words, the classification itself by the primary classifier 631 can be regarded as teaching using the combination defect type.

  When the re-teaching is performed manually, for example, as shown in FIG. 10, a folder indicating each upper defect type is expanded and a folder indicating a lower combination defect type is displayed. The defect image 803 is dragged and dropped into a folder indicating the combination defect type.

  Next, using the defect image data 801, the feature quantity vector 802, and the combination teaching defect type 812 as teacher data, learning of the classifier 611, that is, learning of the primary classifier 631 and the N secondary classifiers 632 is performed. (Step S31). The defect images used for learning are all or part of the images classified in step S25.

  For example, when a feature vector of a defect image whose combination teaching defect type is “aA” is input to the classifier 611, the primary classifier 631 outputs “A” as the primary classification type. Learning by the classifier 631 is performed. Next, learning of the secondary classifier 632 is performed so that the classification result of the feature vector by the secondary classifier 632 associated with the primary classification type “A” becomes “A”. However, at this stage, the primary classifier 631 outputs “A” by learning in step S24, and therefore the learning of the primary classifier 631 is not substantially performed.

  When a feature vector of a defect image whose combination teaching defect type is “aB” is input to the classifier 611, the primary classifier so that “B” is output from the primary classifier 631 as the primary classification type. 631 is learned. Next, learning of the secondary classifier 632 is performed so that the classification result of the feature vector by the secondary classifier 632 associated with the primary classification type “B” becomes “A”. Also in this case, since the primary classifier 631 outputs “B” by learning in step S24, learning of the primary classifier 631 is not substantially performed. Thus, in the first learning of the entire classifier 611, the learning of the primary classifier 631 is not performed as a result.

  Generally speaking, the defect type “X” (teaching defect type) is taught, and the defect class classified as the defect type “Y” (primary classification type) by the primary classifier 631 is combined teaching. “XY” is associated as the defect type. In the first learning of the entire classifier 611, the primary classifier 631 is not substantially learned, and the secondary classifier 632 associated with “Y” receives the feature vector of the defect image. Learning is performed so as to output “X”. Thereby, even if an erroneous classification result is obtained by the primary classifier 631, the secondary classifier 632 functions to correct this classification result. In other words, learning that “positively allows” erroneous classification in the primary classifier 631 is performed. In the first learning of the classifier 611 as a whole, whether or not the learning of the primary classifier 631 is substantially performed depends on the learning algorithm.

  When the classifier 611 is acquired, all or part of the defect images used for learning are classified into combination defect types by the classification control unit 61 using the classifier 611 (step S32). For example, the defect image being classified as “aB” means that it was classified as “B” by the primary classifier 631 but was classified as “A” as taught by the secondary classifier 632. It shall be determined. Hereinafter, the classification result by the secondary classifier 632 is referred to as “secondary classification type”. On the other hand, if the combination teaching defect type is “aA”... “AD” or “aE”, “A” is taught but “A”. It means that it was classified into “E”.

  Therefore, when a defect image whose combination teaching defect type is “aA”... “AD” or “aE” is classified as “aB”, regardless of whether the primary classifier 631 correctly determines, or not. As a result, the classifier 611 performs correct classification. Similarly, the classification results of defect images whose teaching defect types are “aA”... “AD” or “aE” (hereinafter referred to as “combination classification types”) are “aA”, “aC”, and “aD”. In the case of “aE”, the correct classification is performed as a result. The combination classification type is a combination of the secondary classification type and the primary classification type. That is, the combination classification type 823 is generated from the primary classification type 821 and the secondary classification type 822 stored in the information storage unit 65.

  When the combination teaching defect type is “aA”... “AD” or “aE” and the combination classification type is “bA”... “BD” or “bE”, “A” is taught. Indicates that it is classified as “B”. Generally speaking, if the defect type on the front side of the combination teaching defect type (teaching defect type) matches the defect type on the front side of the combination classification type (secondary classification type), the correct classification has been performed. If the defect type on the front side of the combination teaching defect type is different from the defect type on the front side of the combination classification type, it means that an incorrect classification has been performed.

  FIG. 11 is a diagram illustrating a confusion matrix showing the relationship between 25 combination teaching defect types and 25 combination classification types. In this matrix, the statistical value column is omitted. In FIG. 11, five blocks of 5 rows and 5 columns surrounded by a thick line indicate the number of defect images for which a correct classification result is obtained. FIG. 12 shows a summary of 25 cells of 5 rows and 5 columns. Comparing FIG. 9 and FIG. 12, it can be seen that the classification of the entire classifier 611 is more accurate than the classification of the primary classifier 631 alone. However, the classification accuracy of the defect image taught as “C” is still not sufficient. Therefore, learning is performed again by changing the combination teaching defect type (steps S33 and S34).

  In changing the combination teaching defect type, first, “xA”, “xB”, “xC”, “xD”, and “xE” are taught to one of “xA”, “xB”, “xC”, “xD”, and “xE”. In other words, for the defect image of 5 blocks surrounded by a bold line in FIG. 11, the combination teaching defect type is replaced with the combination classification type. “X” is any one of “a” to “e”. For example, for a defect image whose combination teaching defect type is “aA” and whose combination classification type is “aC”, the combination teaching defect type is replaced with “aC”. This respects the fact that correct classification has been performed regardless of the primary classification type, and positively recognizes the primary classification type.

  In the remaining 20 blocks, for example, when the combination teaching defect type is “aA” and the combination classification type is “bA”, the primary classifier 631 correctly classifies, so the combination teaching defect type is changed. I will not. In other words, when the defect type subsequent to the combination teaching defect type matches the defect type subsequent to the combination classification type, the combination teaching defect type is not changed.

  On the other hand, if the combination teaching defect type is “aA” and the combination classification type is “bB”, the classification by the primary classifier 631 is incorrect, so the combination teaching defect type is changed to “aB”, and the primary classification defect type is “aB”. While the classification result by the classifier 631 is recognized, it is set as teaching information indicating that it should be classified as “A”.

  The above-described change of the combination teaching defect type is generally expressed as an operation of “changing the primary classification type included in each combination teaching defect type to the primary classification type included in the combination classification type”. As the arithmetic processing, for example, if the teaching of “aA” is classified as “aC” and expressed as “aA-aC” with five characters, “aC” is obtained by deleting the middle three characters. can get. Similarly, “aA-bA” is converted to “aA”, and as a result, the combination teaching defect type is not changed. From “aA-bB”, “aB” is obtained as a new combination teaching defect type. Thereby, the arithmetic processing is simplified. In addition, the change of the combination teaching defect type can be easily realized.

  When the above arithmetic processing is further expressed using names such as the teaching defect type and the classification type, “By performing classification, each defect image includes the teaching defect type included in the combined teaching defect type and the combined teaching defect type. Are associated with teaching-classification information in which the primary classification types included in the combination classification classification and the primary classification classification included in the combination classification classification are arranged in order. By deleting the primary classification type included in the teaching defect type and the secondary classification type included in the combination classification type, the teaching defect type included in the combination teaching defect type and 1 included in the combination classification type The next classification type is obtained as the changed combination teaching defect type. "

  As described above, in the combination teaching defect type, only the primary classification type that becomes the midway of classification is changed and relearning is performed (step S31). For example, when a feature vector of a defect image whose combination teaching defect type is “aB” is input to the classifier 611, learning is performed so that the primary classifier 631 outputs “B”, and “B”. The secondary classifier 632 associated with the primary classification type is learned so as to output “A”. Generally speaking, the primary classifier 631 is learned to output the primary classification type of the combination teaching defect type, and the secondary classifier 632 corresponding to the primary classification type is the teaching defect of the combination teaching defect type. Learning to output the type. The defect image used for relearning is all or part of the defect image used in the immediately preceding classification process (step S32). The performance of the classifier 611 is further improved by the relearning.

  FIG. 13 is a diagram showing a confusion matrix showing the classification result after repeating steps S31 to S34 five times. In FIG. 13, it can be seen that learning has been performed for each defect type until sufficient classification accuracy is obtained. When sufficient classification accuracy is obtained (step S33), the primary classification types included in each defect image and combination classification type are displayed on the display 55 as necessary (step S35). Of course, when sufficient accuracy can be obtained by performing step S31 only once, the primary classification type may be displayed together with the defect image without performing repeated learning.

  In the above-described processing for acquiring the classifier 611, since the error of the primary classification type is positively recognized, the primary classification type may be greatly different from the teaching defect type by repeatedly changing the combination teaching defect type. is there. However, the primary classification type after iteratively changing the combination teaching defect type strongly reflects the ease of identification or error in the primary classifier 631, and the classification of the teaching policy and defect type by the user is strongly reflected. This is important information for examining the question of whether or not the method was appropriate. Therefore, the host computer 5 which is a classifier acquisition apparatus provides meaningful information to the user by displaying the primary classification type.

  As described above, the host computer 5 can obtain a classifier with high performance by simple processing by obtaining the classifier 611 while allowing misclassification by the primary classifier 631.

  In the classifier 611, when the primary classifier 631 always outputs a correct result for a specific primary classification type, the secondary classifier 632 corresponding to the defect type is a classifier that does not substantially function. Become. The host computer 5 can perform all operations automatically except for the first teaching in step S23. Of course, the second and subsequent teachings may include teaching by the user.

  The defect classification device 1 and the host computer 5 which is a classifier acquisition device, and their operations can be variously changed.

  In the above embodiment, the feature quantity vector and the defect image may be identified. The input of the feature vector is only one form of input to the defect image data. The defect image may be input to the classifier 611 in a form other than the feature vector or used for learning. The primary classification type displayed together with the defect image may be displayed together with the element value of the feature vector.

  The defect image may indicate a pattern defect of a substrate other than the semiconductor substrate or a defect such as a foreign substance. For example, the defect image may indicate a defect of a glass substrate, a photomask substrate, a film substrate, or a printed wiring substrate used in a thin film device such as a hard disk substrate or a thin display such as a plasma display or a liquid crystal display.

  The light applied to the substrate when acquiring the defect image may be lamp light, laser light, or the like, and the substrate may be irradiated with an electron beam. A defect image is acquired using radiation in a broad sense.

  The configurations in the above-described embodiments and modifications may be combined as appropriate as long as they do not contradict each other.

DESCRIPTION OF SYMBOLS 1 Defect classification apparatus 5 Host computer 61 Classification control part 64 Image storage part 80 Program 611 Classifier 631 Primary classifier 632 Secondary classifier 803 Defect image 811 Teaching defect classification 812 Combination teaching defect classification 821 Primary classification classification 822 Secondary Classification type 823 Combination classification type

Claims (7)

A classifier acquisition method for obtaining a classifier for classifying a pattern defect indicated by a defect image into N defect types,
a) including one primary classifier and N secondary classifiers respectively associated with the N defect types, wherein each of the N secondary classifiers is the primary classifier Preparing a classifier connected below;
b) preparing a plurality of defect images;
c) associating a teaching defect type with each of the plurality of defect images;
d) learning the primary classifier using all or a part of the plurality of defect images and the teaching defect type associated therewith;
e) classifying all or part of the plurality of defect images by the primary classifier to obtain a primary classification type;
f) All or part of the defect images classified in the step e) are associated with a combination teaching defect type in which a teaching defect type and a primary classification type are combined, and the primary classifier and the N pieces of defect images are associated with each other. Obtaining the classifier by learning a secondary classifier;
A classifier acquisition method comprising: The classifier acquisition method according to claim 1,
g) After the step f), all or part of the defect images used for learning are classified using the classifier to obtain secondary classification types, and further, secondary classification types and primary classification types Obtaining a combination classification type that combines
h) changing a primary classification type included in each combination teaching defect type to a primary classification type included in the combination classification type;
i) re-learning the classifier using all or part of the defect images classified in the step g) and the combined teaching defect type after modification associated with the defect images;
A classifier acquisition method, further comprising: The classifier acquisition method according to claim 2,
j) after the step i) or after repeating the steps g) to i), further comprising a step of displaying the primary classification type included in the defect image and the combination classification type. Container acquisition method. The classifier acquisition method according to claim 2 or 3,
When the step g) is executed, each defect image includes a teaching defect type included in the combined teaching defect type, a primary classification type included in the combined teaching defect type, and a secondary classification type included in the combination classification type. And teaching-classification information in which primary classification types included in the combination classification type are arranged in order,
In the step h), the combination teaching defect is obtained by deleting the primary classification type included in the combination teaching defect type and the secondary classification type included in the combination classification type from the teaching-classification information. A classifier acquisition method, wherein a teaching defect type included in a type and a primary classification type included in the combination classification type are obtained as the changed combination teaching defect type. A defect classification method for classifying a pattern defect indicated by a defect image into a plurality of defect types,
Preparing a defect image;
Classifying the defect image using the classifier acquired by the classifier acquisition method according to claim 1,
A defect classification method comprising: A defect classification device that classifies pattern defects indicated by a defect image into a plurality of defect types,
An image storage unit for storing defect image data;
A classifier acquired by the classifier acquisition method according to claim 1,
A classification control unit for classifying the defect image using the classifier;
A defect classification apparatus comprising: A program for obtaining, using a computer, a classifier that classifies pattern defects indicated by a defect image into N defect types, the computer executing the program,
a) including one primary classifier and N secondary classifiers respectively associated with the N defect types, wherein each of the N secondary classifiers is the primary classifier Preparing a classifier connected below;
b) preparing a plurality of defect images;
c) associating a teaching defect type with each of the plurality of defect images;
d) learning the primary classifier using all or a part of the plurality of defect images and the teaching defect type associated therewith;
e) classifying all or part of the plurality of defect images by the primary classifier to obtain a primary classification type;
f) All or part of the defect images classified in the step e) are associated with a combination teaching defect type in which a teaching defect type and a primary classification type are combined, and the primary classifier and the N pieces of defect images are associated with each other. Obtaining the classifier by learning a secondary classifier;
A program characterized by having executed.

Images