JP4050273B2 - Classification apparatus and classification method - Google Patents

Description

  The present invention relates to a classification device and a classification method.

  Conventionally, various classification devices are known. For example, in JP-A-8-21803, a defect region to be classified is extracted from an image, and feature quantities such as area, perimeter, ferret diameter, circularity, and barycentric coordinates of the region are obtained in advance. A configuration of an apparatus that classifies defects using a neural network that is trained to output a predetermined defect type with respect to an input pattern of these feature amounts is disclosed.

Japanese Patent Application Laid-Open No. 11-344450 discloses a configuration of a classification device that performs statistical classification based on color information, shape, size, and the like of a defect image, a configuration that supports creation of teaching data (teacher data), and the like. It is shown.
JP-A-8-21803 JP 11-344450 A

  In the above classification device, when classification is performed using feature quantities related to the size (area, ferret diameter, circumference length, etc.) and shape (circularity degree, ferret diameter ratio, etc.) of the classification target area, the target area is extracted. If is not appropriate, the calculated feature value is not appropriate, and an accurate classification result cannot be obtained. The same applies to the teacher data. If the feature amount is not calculated and created after appropriate extraction, it will cause a decrease in performance in the subsequent classification process.

  For example, when appropriate extraction is performed on the classification target region shown in FIG. 20A (FIG. 20B) and when inappropriate extraction is performed (FIG. 20C). The calculated feature values vary greatly, and even if classification is performed based on a teacher data group including such a difference, accurate classification is not performed.

  On the other hand, there is a case where even a human who should serve as an example is confused about the correct answer category input at the time of teacher data creation. In this case, it is not appropriate to treat teacher data with a low confidence level in the same way as other teacher data. Naturally, a classification considering these low confidence levels is necessary. A classification device that shows an effective solution to such a problem has not been proposed.

  The present invention has been made paying attention to such a problem, and the object of the present invention is to provide a classification apparatus and a classification capable of performing high-accuracy classification by considering a validity scale for teacher data. It is to provide a method.

To achieve the above object, the classification apparatus according to the first aspect of the present invention, from the image, a region extracting means for extracting a region for creating teacher data, and classifying the target area, the teacher An operation means for inputting a primary candidate category that is a primary candidate of a correct answer category and a secondary candidate category that is a secondary candidate as correct answer categories for each of the areas for creating data ; It includes a teaching data creation means for creating teacher data including information and information of the secondary candidate categories of the primary candidate category, and a classification means for classifying the region to be the classification target by using the teacher data , the classification means, the teacher data having said primary candidate category of information and the secondary candidate categories of information, is regarded as two teacher data existing in the same position in the feature space, Carry out the classification of the region to be a serial classified.

The classification device according to the second aspect of the present invention relates to the classification device according to the first aspect , and different weights are set for the two teacher data .

The classification device according to the third aspect of the present invention relates to the classification device according to the first or second aspect, and the classification device includes a plurality of data located in the vicinity of the data of the classification target region in the feature amount space. Classification based on teacher data .

A classification device according to a fourth aspect of the present invention relates to a classification device according to the first to third aspects, and the classification device includes data representing a teacher data distribution of the same category in a feature amount space, and Classification is performed based on the distance from the data of the classification target area .

The fifth classification device according to an aspect of the present invention, in the third one of the aspects of the first, the classifier, classifying the data representing the training data distribution of the same category in the feature value space Classification is performed based on the distance to the target area data.

  According to the present invention, since the confidence level of the teacher data is taken into consideration, it is possible to perform classification with high accuracy.

  In addition, since the secondary candidate category is taken into consideration, it is possible to classify with high accuracy.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Here, a defect classification apparatus for a semiconductor wafer will be described as an embodiment of the present invention. However, the present invention is also applicable to other classification devices that extract classification objects (for example, cells) in an image and perform classification based on the characteristics of the region.

  FIG. 1 is a diagram showing a configuration of a defect classification apparatus according to an embodiment of the present invention, in which a defect area extraction unit 101 that extracts a defect area to be classified from an inspection image and a display unit 102 that displays an extraction result. A feature amount calculation unit 103 that calculates a feature amount for the extracted region, an operation unit 104 that selects an extraction region and inputs a correct category or a correct category and a confidence level for the correct category, and a correct category And teacher data creating means 105 for creating teacher data including information on confidence level, teacher data accumulating means 106 for accumulating teacher data, and classification means 107 for classifying the extraction area.

  The configuration related to the acquisition of the inspection image is not particularly limited in the present invention and is not shown. Also, the memory used for holding the acquired image data and each process is not particularly shown.

  The processing of this defect classification apparatus has two processing flows: a teacher data registration process and a classification process. First, the processing flow of the teacher data registration process will be described with reference to FIG. First, an image of a defect (teacher sample) for creating teacher data is acquired using an imaging system (not shown) (step S11). Next, the defect area extraction unit 101 extracts a defect area from the acquired inspection image (step S12).

Here, two methods are shown as the method for extracting the defect area. In the first method, a threshold value that is a non-defective level luminance range is set for the inspection image 133 ((A) in FIG. 3), and a pixel region having a luminance that deviates from this threshold value is extracted as the defect extraction image 140. (B in FIG. 3). Here, the threshold value indicating the non-defective level luminance range may be set in advance, or may be adaptively determined based on the luminance distribution in the image. Further, it may be changed according to the position in the image, or may be common in the image. (See The University of Tokyo Press: Image Analysis Handbook: Mikio Takagi, Yoshihisa Shimoda, Supervision: 502P, Binarization)
In the second method, a non-defective wafer image 134 (or an image of a fixed section that is non-defective) as shown in FIG. 4B is held, and this image and the image shown in FIG. The inspection image 133 (or the corresponding section in the processing target image) is aligned, a luminance difference between overlapping pixels is obtained, and a difference image 135 ((C) in FIG. 4) is created. A defective area is extracted by threshold processing similar to the first method. Note that the region extraction method is changed according to the classification target, and does not limit the contents of the present invention.

  In addition, in the case of a semiconductor wafer or the like, the same defect may be divided and extracted at the time of region extraction due to the influence of the base pattern and the dicing line. Therefore, if necessary, a process for connecting regions may be performed using a morphological process ((Reference): Corona: Morphology: Hidefumi Obata).

  (A) and (B) of FIG. 5 show an example of region connection processing using morphological processing (closing processing). The continuous resolution failure 150 and the unevenness 151 shown in FIG. 5A become connection defect regions 150-1 and 151-1 (FIG. 5B) by the region connection processing, respectively.

  After the defective area is extracted, it is displayed using the display means 102 so that the extracted shape of the area can be visually identified quickly (step S13). As a display method, it is conceivable to display the extracted region so as to be superimposed on the inspection image, or to display the outline of the extracted region so as to be superimposed on the inspection image.

  FIG. 7 is a diagram illustrating a state in which the extraction result for the inspection image 160 illustrated in FIG. 6 is displayed by overlapping the extracted region shapes. The extraction results of the defects 1 to 6 existing in the inspection image 160 of FIG. 6 are shown as the shapes of the regions 1 to 6, respectively. FIG. 8 is a diagram illustrating a state in which the extraction result for the inspection image 160 illustrated in FIG. 6 is displayed by superimposing the outlines of the extraction regions. The extraction result is shown by the shape of the outline. It should be noted that the display can be switched to the original inspection image by a simple operation so that details of the overlapped portion can be easily confirmed. The display scale is also variable so that the details of the shape can be easily confirmed.

  Here, the user operates the operation unit 104 to select a region with a good extraction result by, for example, pointing with a pointer, and inputs a correct answer category of the region and a confidence level for the correct answer category (step S14). ).

  FIG. 9 shows a state of selecting a region and inputting a correct answer category. That is, a state in which the region 1 is selected by the pointer 170 and the defect type A is input as the correct category, and a region 5 is selected by the pointer 170 and the defect type B is input as the correct category is shown. ing.

  Next, FIG. 10 shows a state in which a determination level for the correct answer category is input in addition to the correct answer category. When it can be determined that the defect area (area 5 in the figure) of a certain teacher sample is clearly the defect type B, the confidence level is set to “high” by the pointer 170. On the other hand, if the defect type B is considered to be a defect type B, the confidence level is set to “medium”. Furthermore, here, the confidence level is set to two levels “high” and “medium”, but it is of course possible to divide the level into more levels.

  Next, the feature amount calculation means 103 calculates the feature amount for the region where the correct answer category is input (step S15). Regarding the feature amount, in general, in addition to the size, shape, position, and density information of a single region, the feature amount relates to an arrangement configuration of several regions. The feature amount in the macro inspection is disclosed in Japanese Patent Laid-Open No. 2003-168114 by the present inventor. However, the feature amount calculation method is changed according to the classification target, and the contents of the present invention are It is not limited.

  After the feature amount is calculated, teacher data is created by the teacher data creation means 105 (step S16). As shown in FIG. 11, the teacher data here includes the feature amount information calculated by the feature amount calculation unit 103 and the correct category information (including the confidence level) of the target region input using the operation unit 104. ) Is a set.

  Thereafter, the created teacher data is stored in the teacher data storage means 106 (step S17). The teacher data registration process is completed by repeating the above processing as necessary.

  In this description, the feature amount is calculated after the correct category is input. However, the feature amount is calculated for all the regions first, and the teacher uses the feature amount information only for the region to which the correct category is input later. Data may be created.

  Next, the processing flow of the classification process will be described with reference to FIG. Among the processes of the classification process, the image acquisition (step S21), the defect area extraction (step S22), and the feature amount calculation for the defect area (step S23) are performed only when the imaging target is an inspection object. The processing is the same as steps S11, S12, and S15 of the teacher data registration process described above. However, the feature amount is calculated for all the extraction regions.

  After the feature amount of each defective area is calculated in step S23, the classification unit 107 classifies the calculated feature quantity (classification target region data) based on the distance in the feature amount space between the pre-registered teacher data. Is performed (step S24).

  At this time, the classification means 107 performs classification while treating teacher data having a high certainty level as data more significant than teacher data having a low certainty level. Before explaining the method, the basic classification method will be described.

  FIG. 15 is a diagram for explaining the principle of classification by the k-nearest neighbor method, which is one of methods for performing classification using teacher data. In FIG. 15, ◯, Δ, and □ indicate positions in the feature amount space of the teacher data of the defect type A, the defect type B, and the defect type C, respectively. On the other hand, P is the position in the feature amount space of the area to be classified.

The k-nearest neighbor method is a method in which the largest defect type among the k (in the example, five (pre-set)) teacher data closest to the target region P is set as the defect type of the target region. > 1 resolution failure> 1 unevenness and most scratches, so it is determined that the target area = scratches.In this method, two points (x _i : teacher data, x _j ) in the feature space (N dimension) : Distance target) is required, but there are the following distance calculation methods.

As a method other than the k-nearest neighbor method, as shown in FIG. 16, the distance from the representative point of the teacher data distribution for each defect type (for example, the center of gravity of the entire distribution or the cluster center of gravity when the distribution is clustered) There is a method to classify based on. In this case, the value μ ^l (superscript L) of the feature value l (el) (1 ≦ l (el) ≦ N) at the representative point is calculated by the following formula, and the above distance calculation is performed. Classify as the closest defect type.

Here, n is the number of teacher data (of the same defect type) collected at the representative points, a _i is a weighting factor according to the confidence level of the teacher data x _i or the category candidate order, and is usually 1.

  Note that both the k-nearest neighbor method and the representative point distance comparison method increase the calculation load as the number of features (the number of dimensions) increases. It is also possible to calculate the distance after determining the calculation method of the above and performing the feature amount reduction processing based on the calculation method.

Next, a method for considering the confidence level in the classification method will be described. For example, when classifying by the k-nearest neighbor method, the number of teacher data having a low confidence level is counted as 0.6 instead of 1 when comparing the number of k nearby teacher data. When classification is performed using the representative point distance comparison method, the weight of the teacher data a _{i having} low confidence is set to 1 or less (for example, 0.6), and the representative point value μ ^l (superscript L) is calculated. To do.

  As described above, it is possible to classify with high accuracy by considering the confidence level of the teacher data.

  In addition, as a modification of the above-described embodiment, an example in which a secondary candidate category is input to an object with a low confidence level will be described. In this case, teacher data as shown in FIG. 12 is created, and classification is performed based on this. For example, when classifying by the k-nearest neighbor method, when the number comparison is performed on k neighboring teacher data, only the primary candidate is input as the primary candidate defect type. The teacher data input up to the secondary candidate is weighted in the order of candidates, such as 0.6 for the defect type of the category of the primary candidate and 0.4 for the defect type of the secondary candidate. Set to count. FIG. 13 is a diagram for explaining defect type determination by the k-nearest neighbor method using teacher data including secondary candidates. When counted as described above, in FIG. 13, defect type B: 2.2> defect type C: 1.8> defect type A: 1 and target region = defect type B.

Further, when classification is performed using the representative point distance comparison method, the weight a _i corresponding to the candidate order is set, and the representative point value μ ^l (superscript L) is calculated.

  According to the above-described modification, it is possible to perform highly accurate classification by considering the category of the secondary candidate.

  This completes the description of the classification process. After classifying each defective area, information on the classification result is output (step S25).

  FIG. 17 is a diagram illustrating a modification of the defect classification apparatus described in FIG. This modification is basically the same as the configuration of FIG. 1, but the user interactively changes the degree of extraction of the defect area for an arbitrary range in the image displayed on the display unit 102 of FIG. It is possible.

  FIG. 18 is a flowchart for explaining the processing flow of the defect classification device according to the present modification. FIG. 18 illustrates only the process corresponding to the process added to the process in FIG. 2 and the processes before and after the process. After the extraction area shape is displayed (step S13), the user checks the degree of extraction and determines whether re-extraction is necessary (step S31).

  If re-extraction is necessary, the operation means 104 is used to set a range that requires re-extraction and extraction parameters (such as a threshold for binarization) (step S32). After the range and parameters are set, the defect area extraction unit 101 re-extracts the defect area based on the set content (step S33). After the re-extraction of the defective area, the process returns to step S13, the extracted area shape is displayed, and the determination in step S31 is performed again.

  The extent of partial range extraction is interactively optimized through the above flow. FIG. 19 shows an example of a screen for adjusting the degree of extraction. After defect 4 is selected from defects 1 to 6 existing in inspection image 160, slide bar 170 is slid left and right to extract the defect. It is shown that the degree of is adjusted and optimized.

  As a result of the adjustment, if there is no need for re-extraction, the correct answer category is input to the region where extraction is appropriate (step S14), and thereafter, the teacher data registration process is terminated according to the flow described above.

It is a figure which shows the structure of the defect classification device of one Embodiment of this invention. It is a flowchart for demonstrating the flow of a process of a teacher data registration process. It is a figure for demonstrating the 1st method of defect area extraction. It is a figure for demonstrating the 2nd method of defect area extraction. It is a figure for demonstrating an example of the area | region connection process using a morphological process (closing process). It is a figure which shows an example of a process target image. It is a figure which shows a mode that the extraction result with respect to the process target image shown in FIG. 6 was displayed by area | region superimposition. It is a figure which shows a mode that the extraction result with respect to the process target image shown in FIG. 6 was displayed by outline superimposition. It is a figure which shows the mode of selection of an area | region, and the input of a correct category. It is a figure which shows the state of the setting of the certainty level based on determination of a defect kind. It is a figure which shows an example of the teacher data (plurality) containing the information of confidence level. It is a figure which shows an example of the teacher data (plurality) created by inputting the category of a secondary candidate with respect to the object with a low confidence level. It is a figure for demonstrating the determination of the defect kind by k neighborhood method using the teacher data containing a secondary candidate. It is a figure for demonstrating the flow of a process of a classification | category process. It is a figure for demonstrating the principle of the classification | category by the k-nearest neighbor method which is one of the methods of classifying using teacher data. It is a figure for demonstrating the classification method based on the distance with the representative point of the teacher data distribution for every defect type. It is a figure which shows the modification of this defect classification device demonstrated in FIG. It is a flowchart for demonstrating the flow of a process of the defect classification device which concerns on this modification. It is a figure which shows the example of the screen which adjusts the degree of extraction. It is a figure for demonstrating the fault of the conventional classification method.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 101 ... Defect area extraction means, 102 ... Display means, 103 ... Feature-value calculation means, 104 ... Operation means, 105 ... Teacher data creation means, 106 ... Teacher data storage means, 107 ... Classification means.

Claims (4)

From the image, a region extracting means for extracting a region for creating teacher data, and classifying the target area,
Operation means for inputting a primary candidate category that is a primary candidate of a correct answer category and a secondary candidate category that is a secondary candidate as correct answer categories for each of the regions for creating the teacher data. When,
And the teacher data generating means for generating teaching data including information and information of the secondary candidate categories of the primary candidate category,
Classification means for classifying the region to be classified using the teacher data;
With
The classification means regards the teacher data having the information of the primary candidate category and the information of the secondary candidate category as two teacher data existing at the same position in the feature amount space, and becomes the classification target. A classification apparatus characterized by classifying regions . The device of claim 1, wherein the two training data is characterized in that different weights are set. The classifier, classification apparatus according to claim 1 or claim 2, characterized in that for classifying based on a plurality of training data located proximate data to be classified regions in the feature value space. The classification device is any one of the preceding claims, characterized in that for classifying on the basis of the distance between the same category of the training data distribution representative data and data to be classified regions in the feature value space Classification device according to.

Images