Measurement system

Abstract

The object of this invention is to provide a measurement system that can prepare measurement parameters for measuring the construction of complex patterns at low cost. The measurement system of this invention calculates the measurement sample (see Figure 1) by simultaneously inferring the measurement area from an image of the sample and extracting the characteristic quantities of the inferred measurement area, and using the aforementioned measurement parameters corresponding to the aforementioned characteristic quantities and the characteristics of the aforementioned measurement area.

Description

[Technical Field]

[0001] This invention relates to a measurement system with a measurement structure. [Previous Technology]

[0002] With the miniaturization and / or multilayer construction of semiconductors, the measurement of complex patterns is required. To measure complex patterns, constructing a dedicated measurement method for each pattern requires time and / or human resources costs; therefore, a technology that can perform measurements at low cost is needed. In conventional measurements, measurement points are calculated using pre-prepared measurement parameters.

[0003] Patent Document 1 described below pre-registers the model and parameters of the material and shape of the sample used as the measurement target into a function library, performs matching of the waveform of the captured image with the function library, and uses the matched model and parameters to perform the measurement. [Prior Art Documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2009-198339 [Summary of the Invention]

[0005] [Problem to be solved by the invention] As the complexity of the pattern increases and the characteristics of the measurement location area change, the optimal measurement parameters for each measurement location are different. Therefore, as in Patent Document 1, the technology of pre-registering the measurement parameters into a function library requires the pre-construction of a function library for various sample shapes and parameters in order to cope with the measurement of complex patterns, which is very costly.

[0006] In view of the aforementioned problems, the present invention aims to provide a measurement system that can prepare measurement parameters used for measuring the construction of complex patterns at low cost. [Means for solving the problems]

[0007] The measurement system of the present invention infers the measurement area from the image of the sample while simultaneously extracting the characteristic quantities of the inferred measurement area, and uses the aforementioned measurement parameters corresponding to the aforementioned characteristic quantities, and uses the aforementioned measurement parameters corresponding to the characteristics of the aforementioned measurement area to measure the sample. [Effects of the Invention]

[0008] According to the measurement system of the present invention, measurement parameters used for measuring the structure of complex patterns can be prepared at low cost. Other issues, configurations, advantages, etc. of the present invention are explained by means of the following description of embodiments.

Implementation Method

[0010] <Embodiment 1> FIG1 shows a configuration example of the measurement system 100 according to Embodiment 1 of the present invention. The measurement system 100 is a system for constructing a test sample. The measurement system 100 accepts an input image D101, a measurement region inference model D102, a feature quantity classification model D103, and measurement parameters D104 as inputs. The measurement system 100 includes a measurement region inference unit 101, a feature quantity classification unit 102, a parameter setting unit 103, and a measurement processing unit 104.

[0011] The measurement area inference unit 101 uses the input image D101 and the measurement area inference model D102 to extract the measurement area-related feature quantity D105 of the input image D101, and infers the measurement area D106 based on the extracted feature quantity D105. The reason for inferring the measurement area D106 will be described later.

[0012] The feature classification unit 102 uses feature D105 and feature classification model D103 to classify features according to categories. The input feature is a feature representing the measurement area, and the classification result D107 is classified according to the features of the measurement area.

[0013] The parameter setting unit 103 takes the classification result D107 and the measurement parameter D104 as input and sets the parameters required for measurement processing according to the classification category.

[0014] The measurement processing unit 104 uses the measurement area D106 and the parameters set according to the classification category to perform measurement processing on the measurement area D106 and outputs its measurement result D108.

[0015] Input image D101 displays the image of the target to be measured, including its position and size. Input image D101 can be, for example, an inspection image of a semiconductor taken using a scanning electron microscope (SEM), a medical image taken using computed tomography (CT), an inspection image of an imaging device, or an image of a general road. There are no restrictions on the input image or the target to be measured. The target to be measured may be one in one image or multiple in number; there is no limit to the number. Furthermore, there may be multiple input images instead of just one. The position or size of the target in each image or for each target may vary, and the area to be measured must be set according to the target's parameters.

[0016] The measurement region inference model D102 is a model for inferring the measurement region by exploring the measurement location of the input image D101. The measurement region inference model D102 is constructed using, for example, a convolutional neural network (CNN). The measurement region inference model D102 can be, for example, an object detection AI (artificial intelligence) that detects the position of an object from an image, or a semantic segmentation AI that infers the category of each pixel and detects the position of a specific shape as the position of the object. Alternatively, it can be a model that does not use a CNN, but calculates feature points from features that are invariant to the position and size of objects within the image and detects the feature points of the measurement target. Other models are also possible.

[0017] The measurement region inference unit 101 uses the measurement region inference model D102 to infer the measurement region D106 within the input image D101. At this time, the feature quantity D105 extracted from the input image D101 when inferring the measurement region D106 is also output. The feature quantity D105 is the feature quantity used as the basis for the measurement region inference model D102 to infer the measurement region D106. For example, if an object detection AI or semantic segmentation AI constructed using a CNN is used as the measurement region inference model D102, the feature quantity D105 can be a vector from an intermediate layer of the CNN. If the measurement region D106 is inferred by calculating feature points of the image, the feature quantity D105 can also be set as a feature quantity whose position and size remain unchanged within the image. Other calculation methods or feature quantities can also be used as the feature quantity used as the basis for inferring the measurement region D106. If the measurement area inference model D102 can correctly infer the measurement area D106 from the input image with diverse features, then the extracted feature quantities are different vectors or values ​​represented by the features of the measurement area D106.

[0018] Feature classification model D103 is a model that classifies categories based on feature D105. For example, it can be a classification model composed of a neural network (NN), or it can be a model that compares the distance between features and clusters features with smaller distances, such as the K-means algorithm. There are no restrictions on the composition of the model.

[0019] The feature classification unit 102 inputs feature quantity D105 into feature quantity classification model D103 and outputs classification result D107. As mentioned above, feature quantity D105 presents different features according to the features of measurement region D106, so measurement region D106 can be classified according to the features of measurement region D106.

[0020] The measurement parameter D104 corresponds to the category classified by the feature quantity classification model D103, and is the measurement parameter for each classification category used by the measurement processing unit 104. The measurement parameter D104 is set to a suitable parameter that improves the measurement accuracy of each classification category.

[0021] The parameter setting unit 103 uses the classification result D107 and the measurement parameter D104 to set suitable parameters according to the measurement area D106. The classification result D107 is classified according to the characteristics of the measurement area D106, so suitable parameters can be set according to the measurement area D106.

[0022] The measurement processing unit 104 performs measurement processing using parameters set by the parameter setting unit 103. Measurement processing can be performed using methods such as: for semiconductor images, detecting edges using a threshold method to detect the edges of the measurement target and calculating measurement values ​​from the edges; or other methods. In such cases, there may be parameters that need to be prepared in advance for measurement processing, and the appropriate parameters vary depending on the characteristics of the measurement area. Even in this case, because the parameter setting unit 103 sets appropriate parameters according to the characteristics of the measurement area D106, suitable measurement processing becomes possible.

[0023] Figure 2 shows a flowchart of an example of the steps for preparing the measurement area inference model D102, the feature quantity classification model D103, and the measurement parameters D104. This flowchart can be implemented by any of the functional units of the measurement system 100, or by a computer different from the measurement system 100. For the convenience of the following description, the measurement system 100 is set to implement the steps of the flowchart. The steps in Figure 2 are explained below.

[0024] S201: The measurement system 100 uses the learning data D201 to learn and construct the measurement area inference model D102. The content of the learning data D201 and the steps for learning the measurement area inference model D102 are described below.

[0025] S202: The measurement system 100 uses the learned measurement area inference model D102 and extracts feature quantities from the learning data D202, and learns the feature quantity classification model D103 based on the extracted feature quantities. The content of the learning data D202 and the steps for learning the feature quantity classification model D103 are described below.

[0026] S203: In order to set suitable classification parameters, the measurement system 100 first uses the learned measurement area inference model D102 and extracts the feature quantities of the parameter exploration data D203. The measurement system 100 uses the feature quantity classification model D103 to classify its feature quantities. By classifying the feature quantities, the classification category D204 representing each category can be obtained. The content of the parameter exploration data D203 and the steps to obtain the classification category D204 are described below.

[0027] S204: The measurement system 100 uses classification category D204 and index values ​​representing measurement accuracy. By exploring classification category parameters to improve the index values, measurement parameters D104 are obtained. The index values ​​can be calculated according to index value calculation method D205. The content of the index values ​​and the steps for exploring measurement parameters D104 are described below.

[0028] The learning data D201 can correctly infer the data of the measurement area D106, such as the pre-annotated image and measurement area, using the measurement area inference model D102. The form of the learning data D201 varies according to the measurement area inference model D102.

[0029] The learning data D202 is an image group used to learn the feature classification model D103. In the case of supervised learning, the learning data D202 may contain data representing categories as training data, while in the case of unsupervised learning such as KMeans, it may only contain images, depending on the learning method and the data format.

[0030] Parameter exploration data D203 is data used to explore suitable parameters for each classification category. Parameter exploration data D203 represents the form of the exploration method for the corresponding parameters, such as consisting solely of images.

[0031] The image data contained in each of the learning data D201, learning data D202 and parameter exploration data D203 may be repeated or may be other image data.

[0032] The index value calculated by the index value calculation method D205 is, for example, an index representing the measurement accuracy. In terms of the index value, considerations can be made such as the error between the measured value calculated based on the actual measured value and the inferred information, and the correlation coefficient between the true value calculated using most data and the inferred value.

[0033] If the feature classification model D103 is a model constructed under unsupervised learning, then among the data (D201, D202, D203, D205) used in Figure 2, the data other than the images are only the measurement area and index value of the learning data D201. Therefore, the cost required to prepare such data is small.

[0034] Figure 3 illustrates an example of the method for learning the measurement region inference model D102 in S201. If the measurement region inference model D102 is an object detection AI using CNN, then the learning data D201 consists of image D301 and measurement region D302. Measurement region D302 is the measurement region from which the measurement value of image D301 is calculated. Measurement region D302 can be prepared by, for example, a user manually setting measurement region D302, actually performing measurement processing, and confirming that the processing result is correct. Other methods can also be used to prepare it.

[0035] The measurement region inference unit 301 of the measurement system 100 infers the measurement region of the image D301 using the measurement region inference model D102 to obtain the measurement region D304. The measurement region inference model update unit 302 of the measurement system 100 compares the measurement region D304 with the measurement region D302 in the learning data D201, and updates the weight parameters of the CNN of the measurement region inference model D102 to reduce the difference between the two.

[0036] Regarding the method by which the measurement region inference model update unit 302 obtains the difference between measurement region D302 and measurement region D304, methods such as comparing the centroid positions of the regions, using generalized intersection over union (GIoU), or combining these methods are not limited. Furthermore, regarding the method of updating the weight parameters of the CNN in the measurement region inference model D102, stochastic gradient descent (SGD) can be considered, but the method is not limited.

[0037] Through the above steps, the measurement region (i.e., D304) inferred by the measurement region inference model D102 is close to the measurement region D302 of the learning data D201, and can be correctly inferred. In order to improve the generalization performance of the measurement region inference model D102, it is possible to consider enhancing the data and learning by inverting and / or scaling up or down the images in the learning data D201. The preprocessing of the learning data D201 is not limited to this.

[0038] Figure 4 illustrates an example of the method for learning the feature quantity classification model D103 in S202. The feature quantity extraction unit 401 of the measurement system 100 extracts feature quantity D401 from the learning data D202 using the measurement region inference model D102. The feature quantity classification model learning unit 402 of the measurement system 100 uses feature quantity D401 to learn the feature quantity classification model D103. The learning method varies depending on the form of the feature quantity classification model D103. For example, if the feature quantity classification model D103 is a model using NN (Neural Network), it can be a method that compares the training category with the inferred category and updates the weight parameters of the NN. If it is a method such as KMeans, it can be a corresponding learning method. There is no restriction on the learning method. If it is an unsupervised learnable classification model such as KMeans, the learning data D202 can be just images, and the pre-preparation cost is small. Furthermore, to improve generalization performance, the learning unit 402 of the feature classification model can consider adding noise or other preprocessing to the feature D401. However, there are no restrictions on the preprocessing method.

[0039] Figure 5 shows an example of the method for exploring measurement parameters D104 by classification category in S204. The feature quantity extraction unit 501 of the measurement system 100 uses the measurement region inference model D102 to extract the measurement region D501 and its feature quantity D502 from the parameter exploration data D203. The feature quantity classification unit 102 uses the feature quantity classification model D103 to classify the feature quantity D502. By classifying the feature quantity D502, the classification category D503 representing each category is obtained. The measurement parameter exploration unit 503 of the measurement system 100 uses the classification category D503 and the measurement region D501 to explore the measurement parameter D104.

[0040] Figure 6 shows an example of the configuration of the measurement parameter exploration unit 503. First, the measurement processing unit 104 performs measurement processing using the measurement area D501 and the classification category D503. The measurement processing method varies depending on the measurement target; for example, if it is an inspection image of a semiconductor, a method of calculating the measurement value using a threshold method may be considered.

[0041] The index value calculation unit 5032 of the measurement parameter exploration unit 503 calculates the index value using the index value calculation method D205. As an example of the index value, the average error between the measured value and the true value when measuring multiple images, the correlation coefficient between the true value and the measured value, etc., can be considered. This index value can be freely determined according to the purpose of the measurement. The parameter setting unit 5033 of the measurement parameter exploration unit 503 sets the parameter D601 according to the index value. As an example of the parameter D601 setting method, if the function of the index value is differentiable, the parameter for improving the index value can be analytically calculated; if the function of the index value is not differentiable, a method called black-box optimization, such as Bayes optimization, can be used to explore the parameter, or other methods can be used. When using methods such as Bayes optimization, the parameter that improves the index value as much as possible can be calculated through multiple explorations.

[0042] Figure 7 shows an example of the steps by which the measurement processing unit 104 calculates the measurement value. In this example, the line width of the vertical line shown in D701 is calculated. In order to calculate the line width, measurement areas are respectively arranged around the left and right ends of the vertical line as areas for exploring measurement points. Next, the measurement points in the measurement areas are calculated to obtain measurement points as shown in D702. Then, the line width is calculated by calculating the distance between the measurement points.

[0043] Figure 8 shows an example of the steps by which the measurement processing unit 104 in Figure 7 calculates the measurement points. First, a line profile is created from the measurement area shown by the dashed line D701. Regarding the method of creating the line profile, for example, the distribution of brightness values ​​along a direction with a large brightness gradient and easily calculated edges (in this example, the horizontal direction) can be used as the line profile. In cases where the brightness values ​​contain a lot of noise, smoothing processing can be performed. As a smoothing method, applying an averaging filter or a Hamming window in the vertical or horizontal direction can be considered; the method is not limited. The parameters of the smoothing method need to be determined in advance based on the pattern.

[0044] Next, the measurement points are determined according to the line profile. One example of this method is the threshold method. The threshold method sets the brightness value of the line profile at a specific ratio between the maximum and minimum values ​​as the threshold (Equation 1), and sets the coordinates of values ​​belonging to the threshold as measurement points. The specific ratio is a parameter determined in advance. In addition to the threshold method, there is also a method of setting the coordinates of the line profile at its maximum inclination as the measurement point; there are no restrictions on the method of determining the measurement points. Threshold = Minimum value + (Maximum value - Minimum value) × Ratio (Equation 1)

[0045] The smoothing intensity or threshold ratio when the line profile is created is a parameter determined in advance, but the method in Figure 6 can automatically explore the method that best improves the index value. In addition, various line profile smoothing methods and the method of determining the measurement points can also be considered and prepared to explore the method that best improves the index value.

[0046] Figure 9 is a scatter plot illustrating an example of an index value using the true value and the measured value. In Figure 9, the true value is set as the horizontal axis, and the measured value is set as the vertical axis. For this scatter plot, 'a', when linearly approximated in the form of y=ax+b, is called the slope, and is one of the index values. The slope represents the average ratio of the change in the measured value to the change in the true value; the closer it is to 1, the better the accuracy. 'b' is called the offset, which is also one of the index values; the closer it is to 0, the better the accuracy. Furthermore, the coefficient of determination of the approximate straight line is also one of the index values. The coefficient of determination represents the goodness of the fit of the approximate straight line; the closer it is to 1, the better the accuracy. Furthermore, the average and distribution of the error between the true value and the measured value are also one of the index values. Other values ​​can also be set as index values. The parameter setting unit 5033 explores the parameter D601 for index values ​​that meet the user's measurement purpose to see improvement. Alternatively, a multi-purpose optimization technique, known as optimization, can be used to improve most indicator values ​​to explore parameter D601. There is no limit to the number of indicator values ​​that can be improved.

[0047] Figure 10 shows examples of using the same value for a specific ratio in the threshold method in various measurement areas. In Figure 10, there is one vertically extending pattern with a slightly brighter brightness value and three horizontally extending patterns with even brighter brightness values ​​in a dark background. The positions of the individual horizontally extending patterns relative to the vertically extending patterns are different. As far as the measurement target is concerned, the measurement area is arranged around the left end of the horizontal side of the horizontally extending pattern (1)(2)(3), and the line outline is the central figure. When all the specific ratios of the threshold method shown in Figure 8 are set to 0.5, the measurement points are determined as shown in the right figure. (1)(2) The measurement points are determined in approximately the correct positions, but (3) the measurement point cannot be determined at the left end of the horizontally extending pattern. It can be seen that a ratio of 0.5 is not suitable when observing the outline of (3).

[0048] Figure 11 shows an example where only the ratio of (3) in Figure 10 is set to 0.8. In this case, if the right figure is checked, it can be seen that the measurement point of (3) can be placed at the left end of the horizontally extended pattern. In the measurement system 100, since (1), (2), and (3) have different characteristics, they are classified into different categories, and other ratios can be used for measurement. In addition, the value of the ratio can be automatically explored using the method shown in Figure 2, so suitable measurement can be performed with less preparation cost.

[0049] <Embodiment 2> In embodiment 2 of the present invention, a configuration example with the following function is described: using the feature quantity calculated in embodiment 1, a feature quantity classification model is used to detect the measurement area of ​​abnormal features and to issue a warning as an outlier.

[0050] Figure 12 is a configuration diagram of the measurement system 100 according to Embodiment 2. In Embodiment 2, in addition to the configuration described in Embodiment 1, an outlier warning unit 1201 is included. The outlier warning unit 1201 uses feature quantity D105 and feature quantity classification model D103 to detect measurement areas with abnormal features and warn that the measurement area is an outlier. Other configurations are the same as in Embodiment 1.

[0051] Figure 13 shows an example of a specific method by which the outlier warning unit 1201 warns of outliers. The feature classification model D103 is configured to calculate the center position of each cluster, similar to KMeans, and to ensure that each measurement region belongs to the cluster whose center position is closest to the feature. A threshold is set based on the distance from the center position of the cluster. Measurement regions whose distance from any cluster center exceeds the threshold can be considered outliers. This threshold can be automatically set by increasing the data distribution of each cluster during the learning of the feature classification model D103 by a factor of 3, etc. The outlier warning unit 1201 can detect outliers through the steps described above.

[0052] <Embodiment 3> In embodiment 3 of the present invention, an example is described that displays the measurement area, measurement point, and classification category to which the measurement area belongs, calculated in embodiments 1 to 2, and displays the classification category parameter information. Other configurations are the same as in embodiments 1 to 2.

[0053] Figure 14 shows an example of the GUI provided by the measurement system 100. The measurement area and / or measurement point display section shows the classification category to which each measurement area belongs, as well as the measurement area and measurement points calculated using the methods of Embodiments 1 to 2. The parameter display section displays the set classification category parameters. The measurement value display section displays the measurement values. Through these displays, the user can confirm, based on the flowchart in Figure 2, that the automatically explored classification category parameters are possible.

[0054] Figure 15 shows an example of the GUI provided by the measurement system 100. In addition to the GUI illustrated in Figure 14, a warning is displayed if the measurement area contains outliers. The user can correct the measurement area in the measurement area correction unit. Not only can the warning be issued for measurement areas with outliers, but other measurement areas can also be corrected. Furthermore, the measurement system 100 can be instructed to relearn the measurement area inference model D102 using the corrected measurement area. This makes it more difficult for the measurement area inference model D102 to output erroneous measurement areas, thereby improving the performance of the measurement area inference unit 101 in a short time.

[0055] <Modifications of the Invention> The present invention is not limited to the embodiments described above, but includes various modifications. For example, the embodiments described above are given in detail for the purpose of explaining the invention in an easily understandable manner, and do not necessarily include all the described components. Furthermore, a part of one embodiment may be replaced with the components of another embodiment. Furthermore, the components of another embodiment may be added to the components of one embodiment. Furthermore, a part of the components of another embodiment may be added to, removed from, or replaced with a part of the components of each embodiment.

[0056] In the above embodiments, the measurement processing unit 104 performs measurement using measurement parameters of the feature quantities corresponding to the measurement area. Alternatively, measurement can be performed using a measurement method of the feature quantities corresponding to the measurement area. The measurement method referred to here is an algorithm, etc., that performs measurement processing on measurement points on an image.

[0057] In the above embodiments, the various functional units (measurement area inference unit 101, feature quantity classification unit 102, parameter setting unit 103, measurement processing unit 104, and other functional units described in the figures) of the measurement system 100 can be composed of hardware such as a loop device with such functions installed, or can be composed of software with such functions installed by a computing device such as a central processing unit (CPU). [Simplified Explanation of the Diagram]

[0009] [Figure 1] Presents an example of the configuration of the measurement system 100 in Embodiment 1. [Figure 2] Presents a flowchart of an example of the steps for preparing the measurement area inference model D102, the feature quantity classification model D103, and the measurement parameter D104. [Figure 3] Presents an example of the method for learning the measurement area inference model D102 in S201. [Figure 4] Presents an example of the method for learning the feature quantity classification model D103 in S202. [Figure 5] Presents an example of the method for exploring the measurement parameter D104 according to the classification category in S204. [Figure 6] Presents an example of the configuration of the measurement parameter exploration unit 503. [Figure 7] Presents an example of the steps for the measurement processing unit 104 to calculate the measurement value. [Figure 8] Presents an example of the steps for the measurement processing unit 104 to calculate the measurement point in Figure 7. [Figure 9] A diagram illustrating an example of the index value using a scatter plot of the true value and the measurement value. [Figure 10] Showing examples of using the same value for a specific ratio in the threshold method in various measurement areas. [Figure 11] Showing an example of setting only the ratio (3) in Figure 10 to 0.8. [Figure 12] A configuration diagram of the measurement system 100 in Embodiment 2. [Figure 13] Showing an example of a specific method for the outlier warning unit 1201 to warn of outliers. [Figure 14] Showing an example of a GUI provided by the measurement system 100. [Figure 15] Showing an example of a GUI provided by the measurement system 100.

Claims

1. A measurement system for measuring the structure of a test sample, comprising: a measurement region inference unit that infers a measurement region in an image of the aforementioned sample as a measurement target; and a measurement processing unit that performs measurement on the aforementioned measurement region using measurement parameters corresponding to the characteristics of the aforementioned measurement region inferred by the measurement region inference unit; wherein the measurement region inference unit infers the aforementioned measurement region from the aforementioned image while simultaneously extracting characteristic quantities of the inferred aforementioned measurement region; and the measurement processing unit performs the aforementioned measurement using the aforementioned measurement parameters corresponding to the aforementioned characteristic quantities and the aforementioned measurement parameters corresponding to the characteristics of the aforementioned measurement region.

2. The measurement system as described in claim 1, wherein, The aforementioned measurement area inference unit obtains a measurement area inference model pre-constructed by machine learning to infer the aforementioned measurement area from the aforementioned image while extracting the feature quantities of the inferred measurement area. The aforementioned measurement area inference unit extracts the feature quantities of the aforementioned measurement area from the aforementioned image by inputting the aforementioned image into the aforementioned measurement area inference model.

3. The measurement system as described in claim 2, further comprising a measurement region inference model updating unit for updating the measurement region inference model, wherein the measurement region inference unit infers the learning measurement region contained in the learning image data by inputting the learning image data into the measurement region inference model, the measurement region inference model updating unit calculates the difference between the two by comparing the learning measurement region contained in the learning image data with the learning measurement region inferred by the measurement region inference unit, and the measurement region inference model updating unit updates the measurement region inference model to reduce the difference.

4. The measurement system as described in claim 1, wherein the measurement system further comprises a feature quantity classification unit for classifying the aforementioned feature quantity, and the aforementioned measurement processing unit performs the aforementioned measurement by using the aforementioned measurement parameters corresponding to the features of the aforementioned measurement region, using the aforementioned measurement parameters corresponding to the classification category classified by the aforementioned feature quantity classification unit.

5. The measurement system as described in claim 4, wherein, The aforementioned feature classification unit obtains a feature classification model configured to classify the aforementioned features into the aforementioned classification categories. The aforementioned feature classification unit classifies the aforementioned features into the aforementioned classification categories by inputting the aforementioned features into the aforementioned feature classification model.

6. The measurement system as described in claim 5, wherein, The aforementioned measurement area inference unit uses a measurement area inference model pre-constructed through machine learning to infer the aforementioned measurement area from the aforementioned image while simultaneously extracting the feature quantities of the inferred measurement area. This constitutes a system that extracts the feature quantities of the inferred measurement area while inferring the aforementioned measurement area. The aforementioned measurement system further includes a feature quantity classification model learning unit that constructs the aforementioned feature quantity classification model. The aforementioned feature quantity classification model learning unit constructs the aforementioned feature quantity classification model and uses the aforementioned measurement area inference model to classify the feature quantities extracted from the learning data.

7. The measurement system as described in claim 1, the aforementioned measurement system further includes a parameter setting unit for setting the aforementioned measurement parameters, the aforementioned parameter setting unit sets the aforementioned measurement parameters corresponding to the aforementioned feature quantity, and the aforementioned measurement processing unit performs the aforementioned measurement by using the aforementioned measurement parameters set by the aforementioned parameter setting unit and using the aforementioned measurement parameters corresponding to the features of the aforementioned measurement area.

8. The measurement system as described in claim 7, further comprising a measurement parameter exploration unit for exploring the aforementioned measurement parameters, wherein the aforementioned measurement processing unit performs measurement processing using the parameter exploration image data and the result of classifying the aforementioned parameter exploration image data according to feature quantities, wherein the aforementioned measurement parameter exploration unit calculates an index value representing the accuracy of the measurement processing of the aforementioned parameter exploration image data, and wherein the aforementioned measurement parameter exploration unit obtains the aforementioned measurement parameters used by the aforementioned measurement processing unit by exploring the aforementioned measurement parameters that improve the aforementioned index value.

9. The measurement system as described in claim 1, wherein, In addition to the aforementioned measurement parameters corresponding to the aforementioned characteristic quantities, the aforementioned measurement processing unit performs the aforementioned measurement by using the measurement method corresponding to the aforementioned characteristic quantities.

10. The measurement system as described in claim 1, further comprising an outlier warning unit for detecting outliers of the aforementioned feature quantity, wherein the measurement processing unit uses the aforementioned measurement parameters corresponding to the aforementioned feature quantity to remove the outliers detected by the outlier warning unit.

11. The measurement system as described in claim 1, wherein the measurement system provides a user interface for displaying the measurement results of the aforementioned measurement processing unit.

12. The measurement system as described in claim 4, wherein the aforementioned measurement system provides a user interface for displaying category parameters that identify the aforementioned classification categories.

13. The measurement system as described in claim 10, wherein the measurement system provides a user interface for displaying the aforementioned measurement area belonging to the aforementioned outliers.

14. The measurement system as described in claim 3, wherein the measurement system provides a user interface for users to modify the measurement area inferred by the measurement area inference unit, and the measurement area inference model update unit updates the measurement area inference model using the modified measurement area.

Images