JP2001256480A - Automatic picture classifying method and its device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To secure dassification performance for high fault by permitting instruction data for fault classification which is set at every fault attribute to follow the fluctuation of a manufacturing process. SOLUTION: Featured values are calculated from fault detected from a semi- conductor wafer and assigned in a featured values space, categories are assigned from a distribution in the space as shown in a figure 10(a) and instruction data for fault classification is generated. When fault detected from the wafer is classified into the respective fault attributes, the instruction data is used. When fault is classified, the same processing is performed concerning the fault to be classified, which is detected from the wafer, classification dealing data shown in a figure 10(b) is generated and, then, it is compared with instruction data shown in the figure 10(a). A difference occurs in these kinds of data when a fluctuation exists in the semi-conductor manufacturing process and obtained instruction data is corrected in accordance with the difference when the difference exists.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for automatically classifying an image based on the contents included in the image, and more particularly to teaching data used for automatic classification based on the nature of a defect to be classified. The present invention relates to an automatic image classification method and apparatus capable of maintaining high classification performance by updating to an image classification method.

[0002]

2. Description of the Related Art Conventionally, as a method of classifying defect attributes such as a semiconductor wafer pattern based on an image thereof, a method using a neural network has been known as disclosed in Japanese Patent Application Laid-Open No. Hei 8-21803.

In this method, regarding the classification of defects in a semiconductor product in a semiconductor manufacturing process, when information relating to such defects (defect information: defect area, shape, center of gravity position, measurement conditions, etc.) is input to a neural network, The neural network adjusts the defect information by a predetermined weighting factor, and the resulting output information of the neural network is the attribute of the defect (the type of defect such as a flaw or dust, hereinafter referred to as a defect attribute). The defect is classified into the defect attribute corresponding thereto.

[0004] In order to enable defect classification by such a method, an appropriate weighting factor must be set in the neural network. For this purpose, Japanese Patent Application Laid-Open No. Hei 8-21803 discloses a method of setting the weight coefficient. This method prepares, for each defect attribute, image data called “teaching data” (which corresponds to the above-described defect information), which is used as a reference, and executes neural network learning using this. It is. In this learning, the teaching data is input to the neural network, adjusted by the weight coefficient set at that time, the classification result of the teaching data thus made is compared with the defect attribute for the teaching data, and If the error value representing the mismatch state exceeds a preset threshold, the weight coefficient at that time is corrected according to the error value, and the same teaching data is input to the neural network again to adjust with the corrected weight coefficient. So that
This processing is repeated until the error value becomes equal to or less than the threshold value. In this way, when the weighting factor is such that the error value is equal to or less than the above threshold value, the teaching data input to the neural network is classified into the corresponding defect attribute. Such learning is repeated until a weight coefficient is set so that all the teaching data are correctly classified into the corresponding defect attributes.

The weight coefficients of the neural network obtained as a result of such learning are stored as learning data.

At the time of defect classification, the learning data is used for the neural network, and the defect information actually obtained is input to the neural network, so that the defect is classified into a corresponding defect attribute.

[0007]

By the way, when the defect of the semiconductor wafer pattern is classified in the process of manufacturing the semiconductor wafer by using the classification method of the prior art, the learning of the neural network is performed before the actual defect classification is performed. When the defect classification is performed, the learning data of the neural network is determined. Therefore, no matter how much the teaching data is selected during the learning, the defect is classified based on the learning data. Will be.

For this reason, even if a defect occurring in the semiconductor wafer pattern corresponds to a certain predetermined teaching data and should be classified into a predetermined defect attribute corresponding to the predetermined teaching data from the recognition of the user, such a defect is considered. Changes in the semiconductor manufacturing process (eg, changes in the thickness (thickness) of the film formed on the wafer,
Due to changes in exposure time and etching time, minute changes in pattern dimensions and pattern height, changes in surface roughness, etc.) occur, and as a result, defect information such as the shape of the defect is obtained during the above learning. If the teaching data is different from the corresponding teaching data, the defect is not correctly classified into the above-described predetermined defect attribute, but is classified into another defect attribute.

On the other hand, in a classification algorithm such as a neural network, even if defect information having contents other than the teaching data presented at the time of learning is input, the defect attribute is determined based on the similarity to the pattern of the teaching data at the time of learning. Recognized,
Generally, it has generalization ability to perform classification. However, this generalization ability is not enough to follow the above-mentioned semiconductor manufacturing process variation, and conversely, a known defect attribute obtained at the time of learning a new unknown defect caused by the semiconductor manufacturing process variation. There is a problem that there is a risk of missing an unknown defect which is an important clue for extracting a new problem in the semiconductor manufacturing process.

SUMMARY OF THE INVENTION It is an object of the present invention to solve such a problem, to eliminate the influence of fluctuations in the semiconductor manufacturing process between the time of creating teaching data and the time of executing classification, and to make the teaching data follow the process of executing sorting. An object of the present invention is to provide a method and apparatus for automatically classifying images.

[0011]

In order to achieve the above object, the present invention uses a plurality of feature values representing the properties of defect image data. Using this feature,
Each of the defect data has a value of each feature amount as an element.
Are represented as one vector, which is 1 in the feature space.
Can be considered as one point.

Now, a plurality of teaching defect images used for creating teaching data are associated with one point in a feature amount space using the feature amount values, and this point is referred to as a teacher point. Since it is assumed that the attribute (defect attribute) to which the defect belongs is known for the teaching defect, the distribution of the teacher points has the position information of each teaching defect in the feature vector space and at the same time. , And also has position information of each defect attribute. Similarly, a defect image input as a classification target is associated with a feature vector space by its feature. The point associated with the classification target image is referred to as a classification target point.

The distribution of the taught points means that the state of occurrence of the taught defect attribute is represented by the value of the characteristic amount, and the distribution of the classification target points indicates the state of the generation of the defect at the time of the classification. Is represented by the value of Therefore, when the semiconductor manufacturing process fluctuates, these two
A change will occur between the two distributions. In the present invention,
Analysis means for analyzing the properties of the teaching data from the difference in distribution between such teacher points and classification target points, means for notifying the user of the analyzed result and updating the teaching data, or automatically providing the teaching data based on the detected result Is provided, whereby the teaching data can follow the semiconductor manufacturing process state at the time of classification execution.

[0014]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a view for explaining the role of an automatic image classification method and apparatus according to the present invention in a process of manufacturing a semiconductor product such as a semiconductor wafer (semiconductor manufacturing process). Is a defect classification apparatus corresponding to the automatic image classification method and apparatus according to the present invention;
Is an appearance inspection result, 5 is an image at a defective portion, 6 is an occurrence frequency for each failure mode (defect attribute), and 7 is a quality management system.

In the figure, the manufacture of a semiconductor product consists of hundreds of processes (here, such processes are referred to as processes 1, 2,..., N). May require. However, the quality of a semiconductor product is determined only in all processes 1 to n.
By the probe inspection 8 performed after the semiconductor product is completed through the above. Therefore, in order to improve the yield, means for estimating the quality of the process in the middle is indispensable. For this purpose, the appearance inspection is performed by the appearance inspection apparatus 2 and the quality of the process is determined from the abnormal appearance of the wiring pattern in the semiconductor product.

When a process abnormality is confirmed, it is necessary to take countermeasures. As an information collecting means for this, the image automatic classification method and the defect classification apparatus 3 as an apparatus according to the present invention play an important role. I have. That is, the location and the number of defects can be grasped from the result 4 of the visual inspection,
Information such as the nature of the defect is not available. For this reason, the defect classifying device 3 captures the image 5 at the defect site, classifies the type of the defect based on the image, and generates an occurrence frequency 6 for each defect mode.
Is displayed on the quality management system 7. As a result, defect countermeasure candidates can be narrowed down. Here, it is assumed that the frequency of occurrence of foreign matter is higher than that of the pattern defect, and therefore, it is understood that it is only necessary to take measures to prevent the occurrence of foreign matter. That is, for a defect having a high frequency of occurrence, the cause is preferentially estimated and its countermeasures are taken, whereby the yield can be quickly improved. The defect classification device 3 is a device for automating the defect classification work conventionally performed visually, and requires high classification performance.

In FIG. 1, pass / fail estimation is performed for the process 1, but pass / fail judgment may be made for all processes or for a desired process.

FIG. 2 is a block diagram showing one embodiment of the defect classification device 3 in FIG. 1, wherein 9 is an optical system, 10 is a substrate, 11 is a stage, 12 is a TV camera, 13 is an image input device, 14 Is an image recording device, 15 is an image processing device, 1
Reference numeral 6 denotes a stage control unit, 17 denotes a substrate transfer control unit, 18 denotes a monitor, 19 denotes a host computer, 20 denotes a substrate transfer device, 21 denotes input means, and 22 denotes a network. FIG.
3 is an explanatory diagram of a teaching procedure performed by the defect classification device 3 shown in FIG. FIG. 4 is a diagram showing an example of monitor display in creating teaching data. FIG. 5 is a diagram showing an example of the content of defect data recorded as teaching data.

The embodiment shown in FIG. 2 and its classification procedure (automatic image classification method) will be described below with reference to FIGS.

In this embodiment, first, a defect image for teaching data is collected (step 10 in FIG. 3).
0).

For this reason, first, in FIG. 2, the board transfer control unit 17 controls the board transfer device 20 based on a command from the host computer 19, and
0 is mounted on the stage 11. On the other hand, the host computer 19 transmits the board 1 via the network 22.
The defect coordinate information corresponding to 0 is received from the host system. This higher-level system is a quality management system such as a yield management system, a production line monitoring system, a process management system, or an inspection device for inspecting defects. In the semiconductor manufacturing process shown in FIG. Corresponds to this.

The host computer 19 refers to the defect coordinate information, sends an instruction to the stage control unit 16, moves the stage 11, and moves the substrate 10 to a position where a defect can be observed. Defects on the substrate 10 are transmitted to the TV via the optical system 9.
An image is taken by the camera 12, and the obtained defect image is recorded in the image recording device 14 via the image input device 13.

The above processing is repeated for all the defects specified by the defect coordinate information, and the respective defect images are stored in the image recording device 14. In this embodiment, the TV camera 12 is used as a means for obtaining a defect image. However, the present invention is not limited to this, and another imaging device such as a line sensor or an electron microscope may be used.

Next, creation of teaching data (assignment of a category (defect attribute)) is executed (step 10 in FIG. 3).
1).

That is, when defect images of all the defects on the substrate 10 are recorded in the image recording device 14, the host computer 19 displays these defect images on the monitor 18. The user observes these displayed defect images and assigns a category using the input unit 21. This process is performed on all the defect images required as teaching data for the collected defect images.

As shown in FIG. 4, the accumulated defect images are displayed in the unclassified window 24 on the monitor 18. On the other hand, the user sets the user category window 23 according to the user-defined category (defect attribute).
A, 23B, and 23C are created, and the defect images in the unclassified window 24 are moved to the corresponding user category windows 23A, 23B, and 23C.
A teaching category is assigned to the defective image. The category specified by the user is recorded as teaching data in the host computer 19. Here, the user category is displayed in the user category windows 23A, 23B, and 23C.
However, it is clear that the present invention is not limited to this.

Next, the process proceeds to a feature extraction process (image processing) (step 102 in FIG. 3).

That is, each of the defect images stored in the image recording device 14 is transferred to the image processing device 15, where the image processing is performed, and as shown in FIG. 5, a characteristic amount is extracted for each defect image. Here, the feature amount of the defect image includes color information, shape, size, and the like of the defect. For example, the foreign matter is dark and almost circular, but the pattern defect is the same color as the peripheral pattern and has a complicated shape. Predetermined types of feature amounts (five types in the example shown in FIG. 5) are calculated, and these feature amounts are recorded in the host computer 19 as teacher data corresponding to the defect numbers (numbers assigned to the defects). You.

Next, teaching is performed by the host computer 19 based on the teacher data to create teaching data (step 103 in FIG. 3), and the teaching data is recorded (step 104 in FIG. 3).

Here, the teaching (step 103) will be described.

First, an image feature amount of a defect or its background is extracted from a defect image (teacher image) targeted for teacher data. Here, the feature amount is, for example, a geometric quantitative numerical value such as a circularity or an area of a defective portion, an image average luminance value,
It is a quantitative value related to image information such as a variance value, a value related to color (such as a distribution of RGB, hue, and color value) or a quantitative value related to image information including a background image of a defect.

Now, suppose that the feature quantity is c, the number of preset feature quantities is n, and a combination (c0, c1,..., Cn) of n feature quantities extracted from one image is a feature quantity vector. And This feature vector corresponds to one point in the feature space. As shown in FIG. 5, the feature vector extracted from each teacher image is recorded together with a defect category corresponding to a defect number which is an index of each teacher image. Hereinafter, a point in the feature amount space represented by the feature amount vector extracted from each teacher image will be referred to as a teacher point.

FIG. 6 is a diagram schematically showing a specific example of the distribution of teacher points belonging to two categories A and B, with two feature amounts.

In the figure, in order to classify these teacher points into two categories of categories A and B, generally,
A straight line L or an ellipse E (EA, EB) is determined and determined. Here, when the two feature amounts are f1 and f2, for example, when using a straight line L, [Expression 1] L (f1, f2) <0 → belonging to category A L (f1, f2)> 0 → category It belongs to B. However, [Equation 2] L (f1, f2) = a * f1 + b * f2 + c = 0 represents a straight line L.

When the ellipses EA and EB are used, [Equation 3] EA (f1, f2) <0 → belongs to category A EB (f1, f2) <0 → belongs to category B . However, [Equation 4] EA (f1, f2) = d * f1 * f1 + e * f2 * f2
+ F * f1 * f2 + g * f1 + h * f2 + i = 0 EB (f1, f2) = j * f1 * f1 + k * f2 * f2
+ L * f1 * f2 + m * f1 + n * f2 + o = 0 represents ellipses EA and EB.

As described above, the straight line L and the ellipse E can be used as a discriminant function.

Although the discriminant function determined by a neural network or the like takes a more complicated form, it can be considered similarly. The above-described discriminant function is determined from the given feature amount vector of the teaching image and its category, and this is stored as teaching data. The above is the teaching process (Step 104 in FIG. 3).

In the classification process, the feature amount is calculated from the classification target defect whose classification is unknown, as in the teaching process (step 103 in FIG. 3), and the calculated value is generated in the teaching process. It maps to the feature space and determines which category it belongs to. More specifically, the category is determined based on a function value obtained as a result of applying the feature value of the defect to be classified to the discriminant function obtained in the teaching process.

Incidentally, in a semiconductor product, the life cycle of one product may be several months to several years or more. In such a case, there is a difference in the manufacturing process for manufacturing the product between the initial stage of the life cycle, the middle stage, and the final stage. Defects that occur in the manufacturing process depend on the manufacturing process, which means that different defects appear depending on the life cycle of the product. Here, the “differences in defects that occur” means “differences in defect attributes” such that a defect attribute that did not exist initially is newly generated as a result of fluctuations in the semiconductor manufacturing process, and a defect attribute that has existed in the past. , The meaning of “difference in defect characteristics” such that the appearance (shape, color, etc.) of the defect is different from that observed conventionally. Considering that a product having such a relatively long life cycle is subjected to production control by applying automatic defect classification, the above-described fluctuation in the semiconductor manufacturing process becomes a serious problem.

In the automatic classification system, defect images actually generated are collected, and using this as a teaching sample, teaching data for each category is created. Then, referring to the teaching data, it is determined to which category the defect to be classified belongs. In other words, this means that only the information on the defects that can occur in the semiconductor manufacturing process at the time of creating the teaching data is reflected in the teaching data, and the semiconductor manufacturing process at the time of classification is the same as that at the time of teaching. This means that the processing is performed on the assumption that they match. Therefore, when a change occurs in the semiconductor manufacturing process, it is necessary to update the teaching data in accordance with the change. Otherwise, the correlation between the feature amounts of the teaching data and the classification target decreases, and the expected classification performance cannot be obtained.

However, the existing automatic classification apparatus generally does not allow the user to easily detect a change in the semiconductor manufacturing process. All that the user can understand is the fact that the classification accuracy is reduced. Moreover, this can be understood only when the user visually confirms whether or not all the results of the automatic classification match the expected classification results. Such a visual review of all defects is inefficient, and the automatic classifier that would otherwise be required to reduce these efforts would have the opposite effect. Further, it cannot be read from the classification result alone whether the cause of the decrease in the classification performance is due to a change in the semiconductor manufacturing process or another cause. Given these facts, the automatic classification process can
It needs to be robust.

This embodiment makes it possible to determine whether or not the teaching data at the time of classification is suitable for the classification process in order to provide a classification performance that is robust against variations in the semiconductor manufacturing process. It is.

FIGS. 7A and 7B are diagrams showing a specific example of a method of judging a variation in a semiconductor manufacturing process. FIG. 7A shows the distribution of feature points of teaching data, and FIG. The distribution of the quantities is shown respectively.

Here, as shown in FIG. 7A, it is assumed that there are two defect attributes of categories A and B at the time of teaching. The feature amount distribution of the classification target illustrated in FIG. 7B is obtained by plotting the feature amounts of a plurality of defects acquired as the classification target data on a feature amount space. The defects to be classified may be all or some defects of one semiconductor wafer, or may be all or some defects of each of a plurality of semiconductor wafers. Further, for example, defect data obtained from a semiconductor wafer manufactured within a predetermined period, such as one day or one week, may be used. In FIG. 7B, there is no information on which category each feature point belongs to. The case of including information on which category each feature point belongs to will be described later with reference to FIG.

By comparing the distributions of FIGS. 7 (a) and 7 (b), the distribution of the entire feature point of the defect at the time of teaching and the distribution of the feature amount of the entire defect at the time of defect classification are compared. Can be. For example, in the distribution shown in FIG. 7A, there are two places where the feature points are densely arranged, whereas in FIG.
It can be seen that there are three locations in the distribution shown in FIG.
If an area where the feature points are densely gathered in the feature amount space is referred to as a cluster, a portion forming the cluster means that there are many defects having the same kind of feature amount. Therefore, as shown in FIG. 7B, when the number of clusters at the time of performing the classification is increased as compared with the time of teaching shown in FIG. 7A, a defect of a new attribute that did not exist at the time of teaching occurred. It means that it is likely.

FIG. 8 is a flowchart showing a specific example of a processing algorithm for actually detecting a cluster from a feature space.

In the following, a set of N feature points in a feature space is P (P (1), P (2),..., P (N)), and an existing cluster is C (C (1 (1)). ), C (2), ...
...), the variable NP represents the number of processed patterns, and the variable NC represents the number of clusters created. For a certain cluster, a geometric center point of the cluster in the feature space is referred to as a standard pattern of the cluster.

In FIG. 8, first, variables NP and NC are initialized to NP = 1 and NC = 1, respectively (step 200), and the feature point P (NP) = P
Cluster C (NC) = C with (1) as a standard pattern
(1) is created (step 201). That is, the feature point P (1) becomes the center point of the cluster C (1).

Next, a distance d defined by the following equation is obtained (step 202). [Equation 5] d (min) = d (P (NP + 1), C (i)) Here, if there are m clusters C, i = 1,
2,..., M, and d (P (NP + 1), C
(I)) shows a cluster C (i) and a feature point P (NP + 1)
Represents the distance between There are several methods for defining the distance between the cluster and the feature point. Here, the distance between the standard pattern of the cluster and the feature point (ie, 2
Distance between points). feature point P for each of m clusters
Find the distance from (NP + 1). Initially, NP = NC
= 1 and there is only cluster C (1), so only the distance between feature point P (2) and cluster C (1) is determined.

Then, for a predetermined threshold T defining the range of the cluster C (i), the distance d (min) to one of the m clusters C (i) is expressed by the following equation (6). If d (min) ≦ T (step 203), the characteristic point P (N
P + 1) belongs to the cluster C (i) (step 204). Therefore, when NP = NC = 1 at the beginning, it is assumed that the feature point P (2) belongs to the cluster C (1). If d (min)> T for any of the m clusters (step 203), the feature point P (N
P + 1) does not belong to any cluster, a new cluster C (NC + 1) using this feature point P (NP + 1) as a standard pattern is created, and a variable N
The value of C is incremented by 1 to indicate that the number of clusters has increased by one to m + 1 (step 205).
Therefore, when the first NP = NC = 1, the feature point P
A new cluster C (2) having (2) as a standard pattern (center point) is added, and clusters C (1) and C (1) are added.
There will be two of (2).

When the processing of step 204 or 205 is completed, the value of the variable NP is increased by one (step 2).
06), when NP> P, the processing has been completed for all feature points, but when NP ≦ N, the process returns to step 202 (step 20).
7) The operation from step 202 is repeated for new variables NP and NC. That is, the processing up to this point is the feature point P
In the case of (2), the next feature point P (3) is designated in step 206, and this feature point P (3)
, The operation from step 202 is repeated so that this feature point P (3)
(I) (step 204) or the feature point P
A new cluster with (3) as a standard pattern is newly created (step 204), and thereafter, this process is repeated for all feature points.

In this way, all the feature points P (NP)
Are classified into one of the clusters C (NC).

By applying the above processing to each of the teacher point and the classification target point, it becomes possible to calculate the number of clusters in the feature quantity space and to compare the numbers. It is possible to detect a change in the semiconductor manufacturing process between the time and the execution of the classification. When there is a difference between them, it means that the semiconductor manufacturing process has fluctuated. The calculation of the cluster is not limited to the above example, and another method may be used.

Further, it is also possible to calculate statistical quantities which are easy to quantitatively calculate from the distribution of each feature point at the time of teaching and at the time of classification, respectively, and to compare these. FIG. 9 illustrates the concept of the average point and variance of the feature point group, the principal component axis, and the variance with respect to the principal component axis as a specific example of the statistics. It is also possible to approximate each distribution with a sphere or the like, and calculate the area of the overlap between the spheres as a statistic. By comparing the feature amount distributions at the time of teaching and at the time of classification using such statistics, the relative change can be calculated quantitatively. Then, a threshold value is set in advance for each statistical value, and when the actual change exceeds the threshold value, a message indicating to the user that there is a change in the semiconductor manufacturing process is displayed. Thus, the user can know it. Although the user can see the image of the defect, it is difficult to determine a subtle difference in the shape, color, and the like. However, according to this embodiment, the characteristics (shape, color, etc.) of each defect are quantitatively evaluated, and the deviation of the overall tendency is notified to the user. Fluctuations can be detected. This allows the user to determine whether or not to update the teaching data as needed.

FIG. 10 is a diagram showing another specific example of the method of judging the fluctuation of the semiconductor manufacturing process according to the present invention. FIG. 10A shows the distribution of the defect feature amount at the time of teaching. b)
Indicates the distribution of defect feature amounts at the time of classification.

FIG. 10A shows a distribution similar to that of FIG. 7A. On the other hand, in this specific example, FIG.
The distribution shown in FIG. 7B is different from the distribution shown in FIG. This means that in the classification target distribution shown in FIG. 10B, information indicating to which category each defect belongs is added. If the distribution of each category is known for the defect to be classified, as described with reference to FIG. 7, in addition to evaluating the change in the statistics of the entire defect, it is possible to detect a change in the distribution for each category. Will be possible. In the evaluation for each category, a category that formed only one cluster at the time of teaching was detected at the time of classification, when it was formed into two clusters, when the cluster was moving, or when it was expanded or reduced. The teaching data can be created for each category.

For example, in FIG. 10B, FIG.
, A new cluster that did not exist was generated, which means that there was a new defect not included in any of the teaching categories. In addition, it can be confirmed that the distribution of the category A is moving in the feature amount space.
It can be confirmed that the property of the feature amount for expressing the expression has changed from that at the time of teaching.

In FIG. 10, information indicating to which category each defect to be classified belongs is added.
How to add this will be described.

In the first method, the operator visually checks all defects to be classified and determines which category the defect belongs to. For example, as shown in FIG. 11, a window 25 for displaying defects to be classified and windows 26A and 26B for displaying defects of each category (here, two terminals are provided) on a terminal of the defect classification device. However, the present invention is not limited thereto, and each defect image is displayed as an icon 26 on the windows 25, 26A, and 26B. Therefore, the user sets the defect icon 28 in the classification target defect window 25 to 1
Each of them is moved to the window 26A, 26B of one of the categories to which each belongs while visually confirming each one.

In this case, inside the defect classification device, the icon 2 in the feature amount space is synchronized with the movement of the icon 28.
Information on which category the point corresponding to 8 belongs to is given. Also, considering that there is a defect belonging to a category that does not exist at the time of teaching, a window for a new category can be created as an unknown defect window 27 at an arbitrary timing.

As a second method, there is a method of automatically classifying defects to be classified using teaching data. As a method of automatic classification, a method using a discriminant function as shown in the above-described equations 1 to 4 is generally used. Also in the classification process, if the discriminant function value for any category is less than a certain threshold value, it can be determined that the defect is not included in any of the categories existing at the time of teaching. Automatic classification is calculated inside the defect classification device, and a category (defect attribute) can be assigned to each feature point in the feature amount space according to the calculation result. In this case, from the viewpoint of the user,
It is necessary to confirm which category is assigned to which defect. This is performed by automatically moving the icon 28 shown in FIG. 11 according to the classification defect.

The third method is a combination of the above first and second methods. The first method is inefficient when the number of defects to be classified is very large. In addition, in the second method, when the semiconductor manufacturing process varies greatly between the teaching data and the classification target data when they are acquired, it is difficult to obtain a correct automatic classification result, and errors may be mixed in the automatic classification result. There is a problem that there is. In order to avoid this problem, first, the defect classifying apparatus performs automatic classification, confirms the result on the screen shown in FIG. 11, and moves only the icon 28 of the defect having a classification error to the correct category window. That should be done.

In the above method, after a category (defect attribute) is assigned to a defect to be classified, the teaching data can be automatically updated using this information.

This is based on statistics (such as the number of clusters, average values, variances, principal components, and the like described above) representing the distribution of feature amounts, and information on the results of automatic classification of defects to be classified using teaching data. Can be realized. For example, a threshold value is set in advance for a statistic obtained from the above-described feature amount, and when the value of the statistic exceeds this threshold value, it is determined that re-teaching is necessary, and the defect classification apparatus automatically Can be used to update the teaching data.

Other examples include re-teaching when the number of clusters does not match between the teaching data and the classification target data, or the overlapping volume when each category is approximated by a sphere or ellipse exceeds the threshold. Can be re-taught. In addition, as a method of re-teaching using information of a result of automatically classifying the classification target data using the teaching data,
Calculate the classification accuracy rate and perform it when the value exceeds a certain threshold, or when the number of defects (unknown defects) not included in the teaching category exceeds a certain threshold. Can also. Further, it is also possible to perform re-teaching by judging information as a result of visually classifying the defects to be classified by the operator, for example, the number of defects (unknown defects) not included in the teaching category.

Next, a description will be given of a method of selecting a newly required teaching sample in order to update the teaching data.

In the first method, a defect of the data to be classified is added to the teaching sample used to create the original teaching data. The defect to be added may be all or a part of the classification target data. When adding all the data to be classified, all data on the screen shown in FIG. 11 is added. In addition, for example, only a defect misclassified by the automatic classification process may be added to the teaching sample. In this case, the fact that the operator moves the defect icon 28 on the screen shown in FIG. 11 and corrects the error of the automatic classification result and stores the fact that this defect should be added as teaching data is linked. It is possible to As described above, adding the defect of the classification target data to the teaching sample is suitable when the distribution of each category is little changed due to a change in the semiconductor manufacturing process.

In the second method, it is also possible to use only the data to be classified as a new teaching sample without using the original teaching sample as the teaching sample at the time of re-teaching. All of the classification target data may be used as teaching samples,
Only some of the data may be used as the teaching sample. This is suitable when the variation in the semiconductor manufacturing process is so great that the teaching sample originally used has no meaning at all.

The third method is a method of re-teaching an original teaching sample and a defect selected from the classification target data as a teaching sample. This is suitable when the distribution of each category is moving in the feature amount space due to fluctuations in the semiconductor manufacturing process. In this case, the teaching sample that is no longer included in the distribution of the defect to be classified due to the movement of the distribution is excluded from the teaching sample. For example, for each of the teacher point and the classification target point, the distribution of each category is approximated by a sphere or an ellipsoidal sphere, and a defect to be removed from the teaching sample and a defect to be added to the teaching sample can be selected based on the degree of their overlap. . FIG.
2, the part a was used as a teaching sample in the past.
The part b represents a defect added at the time of re-teaching as a teaching sample. For the defect included in the region c, both the original teaching sample and the defect to be classified may be used for re-teaching.

The method of adding and exchanging teaching data samples in the first to third methods has been described. However, which of these three methods is adopted, the teaching point of the feature amount space and the classification target It can be determined automatically from the change in distribution between points. For this purpose, the cluster may be approximated by a sphere or the like for each category, and the degree of overlap may be compared.

The re-teaching process by exchanging the teaching data samples described above can be performed for each category. In other words, re-teaching is not performed for a category in which there is no significant difference in the distribution between the teaching point and the classification target point, and re-teaching is performed only for a category having such a difference. Thus, the teaching process can be performed efficiently.

The method described above is a method of automatically triggering the re-teaching process using the difference in the feature amount distribution. However, the method is not automatically performed, but information obtained at each stage of the process is used. May be interactively notified to the user, and the process may proceed while obtaining the confirmation of the user.

FIG. 13 is a view showing a specific example of a display screen for notifying the user of the result of the analysis of the teaching data when the user re-teaching interactively.

In this figure, an image display unit 30 for displaying a list of icons 31 of defects to be classified is displayed on this display screen.
And analysis result display unit 2 for displaying the analysis result of the teaching data
There are nine. The analysis result display unit 29 displays how much the statistic amount (average, variance, etc.) of the feature amount data of the defect to be classified fluctuates with respect to the teaching data. In the defect classification apparatus, a result 32 of determining whether or not re-teaching is necessary for each category based on these fluctuation amounts is displayed.
The user can execute re-teaching by pressing the teach button 33.

FIG. 14 is a flowchart showing a specific example of a process for updating the teaching data by using the method of sensing the variation of the feature amount distribution shown in FIGS. 7 and 10.

The evaluation of teaching data that does not require the category (defect attribute) of each defect to be classified shown in FIG. 7 can be performed fully automatically in the defect classification device.
The method of evaluating the feature amount of a defect using the classification category information of each defect to be classified shown in FIG. 10 requires a process for giving correct classification category information to all the defects. May be mixed.

The maintenance of the teaching data obtained by mixing the processes shown in FIGS. 7 and 10 is performed as follows.

That is, in the normal state, the distribution change of the entire defect distribution shown in FIG. 7 is stored. This information can be browsed by the user at any time. Then, when an explicit instruction from the user or a statistic representing the distribution observed on the monitor exceeds a certain threshold value, the situation of the distribution change for each category shown in FIG. 10 is analyzed. Then, as a result of the analysis, if the user determines that re-teaching is necessary, or if the statistic obtained as a result of analyzing the change in the distribution exceeds a certain threshold, re-teaching is performed. This re-teach may be performed for all categories or only for a selected category. Further, the teacher sample required for re-teaching can be determined from the distribution of defect points in the feature space as described above.

As described above, by examining the difference in the distribution of the defect feature amount data between the teaching time and the classification execution time,
A change in the semiconductor manufacturing process can be detected, and when the change in the semiconductor manufacturing process is detected, the teaching data can be updated so that the teaching data can follow the change in the semiconductor manufacturing process.

Although the embodiments described above relate to inspection for defects in a semiconductor product manufacturing process, the present invention is not limited to this, and may be applied to a glass substrate to which a liquid crystal panel resist is applied. It goes without saying that the present invention can be applied to the classification of images obtained from any other products.

[0082]

As described above, according to the present invention, even when the feature amount between the teaching data and the classification target data changes due to a change in the semiconductor manufacturing process, it is possible to follow the teaching data suitable for the classification target. This makes it possible to secure stable classification performance.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a role of the present invention in a semiconductor manufacturing process.

FIG. 2 is a configuration diagram illustrating an embodiment of the defect classification apparatus in FIG. 1 corresponding to the automatic image classification method and apparatus according to the present invention.

FIG. 3 is a flowchart illustrating a specific example of a classification learning processing procedure performed by the defect classification device in FIG. 1;

FIG. 4 is a diagram showing a specific example of a teacher data creation screen in the automatic image classification method and apparatus according to the present invention.

FIG. 5 is a diagram showing a specific example of teacher data in the automatic image classification method and apparatus according to the present invention.

FIG. 6 is a diagram for explaining the principle of teaching in the automatic image classification method and apparatus according to the present invention.

FIG. 7 is a diagram for explaining a comparison between teaching data and classification target data in the automatic image classification method and apparatus according to the present invention.

FIG. 8 is a flowchart showing processing for calculating a cluster from a feature amount space in the automatic image classification method and apparatus according to the present invention.

FIG. 9 is an explanatory diagram of statistics calculated from a set of feature points in a feature amount space in the automatic image classification method and apparatus according to the present invention.

FIG. 10 is a diagram illustrating a comparison between teaching data and classification target data in the automatic image classification method and apparatus according to the present invention.

FIG. 11 is a diagram of a display screen for associating a defect to be classified with a category in the automatic image classification method and apparatus according to the present invention.

FIG. 12 is a diagram illustrating sampling of a defect for re-teaching in the automatic image classification method and apparatus according to the present invention.

FIG. 13 is a diagram showing a specific example of a display screen for notifying a user of a result of analysis of teaching data when the user interactively re-teaches in the automatic image classification method and apparatus according to the present invention.

FIG. 14 is a flowchart illustrating a procedure for updating teaching data in the automatic image classification method and apparatus according to the present invention.

[Explanation of symbols]

 Reference Signs List 2 appearance inspection device 3 defect classification device 4 appearance inspection result 5 defect image 6 screen showing occurrence frequency for each failure mode 7 quality control system 8 probe inspection 9 optical system 10 substrate 11 stage 12 TV camera 13 image input device 14 image recording device Reference Signs List 15 Image processing device 16 Stage control unit 17 Substrate transfer control unit 18 Monitor 19 Host computer 20 Substrate transfer device 21 Input means 22 Network 23A to 23C User category window 24 Unclassified window 25 Classification defect window 26A, 26B Category window 27 Unknown defect Window 28 Icon 29 Analysis result display part 30 Image display part 31 Icon 32 Re-teach result 33 Teach button

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G06T 7/00 300 G06T 7/00 300F 350 350B (72) Inventor Kenji Ohara Yoshida-cho, Totsuka-ku, Yokohama-shi, Kanagawa 292 Hitachi Manufacturing Co., Ltd.Production Technology Research Laboratories (72) Inventor Yasuhiko Ozawa 882 Omo, Hitachinaka-shi, Ibaraki Pref.Hitachi Co., Ltd. 882, Omo, Ichigo Co., Ltd. within Hitachi Measuring Instruments Group (72) Inventor Shizushi Isogai 882, Ooichi, Oita, Hitachinaka-shi, Ibaraki Pref. Hitachi Measuring Instruments Group F-term (reference) FF01 FF04 JJ02 JJ03 JJ19 JJ25 JJ26 PP12 PP24 QQ24 QQ31 QQ36 QQ41 RR06 SS02 SS13 2G051 A A51 AB01 AB07 CA04 EB01 EC01 ED21 5B057 AA03 DA03 DA12 DB02 DC03 DC25 DC36 DC40 5B075 ND06 NK46 NR12 PP02 PP03 PQ02 PQ46 QP01 QT04 5L096 BA03 CA02 FA32 FA33 FA59 FA64 GA38 JA11 JA22 KA04 MA07

Claims (10)

[Claims] 1. A category to which an image obtained by imaging an object to be inspected belongs is determined by referring to teaching data created based on an image obtained by imaging the object to be inspected. In the automatic image classification method, a statistic of a feature calculated from classification target data which is data of one or more images as a classification target, and a statistic of a feature calculated from the teaching data, An automatic image classification method characterized by analyzing properties of teaching data or classification target data and automatically or manually updating the teaching data according to the analysis result. 2. The automatic image classification method according to claim 1, wherein at least one of the analysis result and the update result of the teaching data is notified to an operator. 3. The automatic image classification method according to claim 1, wherein the property of the teaching data or the classification target data is
An automatic image classification method, wherein the classification target data is analyzed based on a result of automatic classification using the teaching data. 4. The automatic image classification method according to claim 1, wherein, as a result of automatically classifying the classification target data using the teaching data, it is determined that the classification target data does not belong to any of the taught categories. An automatic image classification method, characterized by analyzing the properties of the teaching data or the classification target data using the number of classification target data. 5. The automatic image classification method according to claim 1, wherein the updating of the teaching data is selectively performed for each of the taught categories. Method. 6. The automatic image classification method according to claim 1, wherein a property of the teaching data or the classification target data is analyzed for each of the taught categories. Automatic image classification method. 7. The automatic image classification method according to claim 1, wherein at least one of the analysis of the property of the teaching data or the classification target data and the updating of the teaching data is determined in advance. An automatic image classification method, wherein the method is performed at regular intervals. 8. A category to which an image obtained by imaging the object to be inspected belongs is determined by referring to teaching data created based on an image obtained by imaging the object to be inspected. In the automatic image classification device, a statistic of a feature calculated from classification target data, which is data of one or a plurality of images as a classification target, and a statistic of a feature calculated from the teaching data, An automatic image classification apparatus, comprising: an analysis unit that analyzes the properties of teaching data or the classification target data; and a unit that updates the teaching data in accordance with a processing result of the analysis unit. 9. The automatic image classification apparatus according to claim 8, further comprising a display unit that displays a processing result by the analysis unit or a result of the update processing of the teaching data. 10. The automatic image classification device according to claim 8, wherein at least one of the teaching data or the analysis result of the classification target data calculated for a plurality of classification target sets is stored. An automatic image classification apparatus, comprising: a storage unit; and a display unit that displays at least one of information stored in the storage unit.