Dataset generation device, dataset generation method, imaging device, imaging method, microscope device, dataset generation system, imaging system, and semiconductor device manufacturing system

Abstract

The purpose of the present disclosure is to provide a technique capable of generating a dataset necessary for constructing a discriminator ensuring recognition performance when discriminating a charged particle beam image.　In order to provide the technique, the dataset generation device according to the present invention, which generates a dataset used for training a discriminator in which image recognition of a charged particle beam image is performed, is characterized in that: on the basis of a region of interest (ROI) which is registered to perform image recognition and the charged particle beam image, candidates for a non-ROI, which is a region not including the ROI, are generated; and the non-ROI is generated while the number of the candidates for the non-ROI is reduced on the basis of the ROI.

Description

Dataset generation device, dataset generation method, imaging device, imaging method, microscope device, dataset generation system, imaging system, and semiconductor device manufacturing system

[0001] The present invention relates to a data set generating apparatus, a data set generating method, an imaging apparatus, an imaging method, a microscope apparatus, a data set generating system, an imaging system, and a semiconductor device manufacturing system.

[0002] In semiconductor devices, nanoscale miniaturization and three-dimensional (complex) device structures are progressing due to the demands for lower power consumption and increased memory capacity. In the manufacture of miniaturized devices, not only are pattern dimensions reduced, but also the realization of more precise processed shapes compared to conventional devices is required.

[0003] To achieve the desired processed shape (called the target shape) using plasma etching, it is necessary to set parameters to control the many gas systems, power systems, and radio frequency systems equipped in the etching equipment. To achieve a complex processed shape, many parameters must be adjusted over multiple steps, with settings switched every second, requiring the setting of an enormous number of parameters for processing a single pattern. The processing conditions for etching, or the parameters set to execute the processing conditions, are called an etching recipe.

[0004] When developing an etching recipe, the processed shape is observed each time the parameters are changed. One technique that enables such nanoscale shape observation is cross-sectional scanning electron microscope (X-SEM) observation. X-SEM observation involves three steps: (1) sample formation, (2) X-SEM observation, and (3) image measurement. These steps are performed manually, which requires significant labor and time costs, creating a need for automation of the process. (1) is less expensive than (2) and (3), and automated measurement of (3) is being developed using general-purpose image measurement software and deep learning technology. Regarding the automation of (2) X-SEM observation, various microscope manufacturers have developed automated observation software that is available to users.

[0005] Automating X-SEM observation requires a classifier to recognize the region of interest (ROI) from the SEM image. In advanced CD-SEM technology for automated observation, there are cases where the template matching (TM) method is used as the classifier. The TM method is a method for searching for a field of view that matches a pre-registered image, and is suitable for CD-SEMs used to manage pattern dimensions on production lines. For example, Patent Document 1 discloses a method for automatically correcting the position of the observation field in SEM observation using pattern matching technology. Specifically, Patent Document 1 discloses the following as an invention of a charged particle beam device. "The imaging device is configured to be able to move the sample on at least two drive axes and is equipped with a sample stage that can move the imaging field of view in accordance with the position information of the sample required by the computer system. The computer system is equipped with a classifier that, in response to input of image data of a tilt image captured while the sample is tilted with respect to the charged particle beam, outputs position information of one or more feature parts that exist on the tilt image. The classifier has been trained in advance using training data that takes as input the image data of the tilt image and outputs position information of the feature parts, and the computer system executes a process of outputting position information of the feature parts for new tilt image data input to the classifier."

[0006] The TM method reduces the workload of the operator when aligning the field of view with the position of the object to be observed. However, the cross-sectional shape of X-SEM observations in etching recipe development changes significantly when the etching recipe is changed. Because the image registered for the TM method and the shape of the ROI differ, it remains difficult to identify the ROI using the TM method for the cross-sectional shape of X-SEM observations.

[0007] To address this issue of ROI recognition in X-SEM observation, Patent Document 2 proposes automating X-SEM observation using a classifier trained by machine learning. Patent Document 2 discloses a method for using a trained classifier to recognize a mark on a sample and move to its coordinates. Specifically, Patent Document 2 aims to provide an image display device with reduced power consumption, and discloses the following as an invention of an image display device: "An image display device comprising: first wiring interconnecting a plurality of electron-emitting devices each having an electron-emitting portion between a pair of electrodes; second wiring having a higher resistance than the first wiring; an insulating layer covering the second wiring; and a resistive film connected to the first wiring and covering the insulating layer, the resistive film having a surface resistance of 108 Ω / □ or more, the resistive film connecting to the first wiring at a portion that does not overlap with the second wiring; and the length L between the portion of the resistive film connected to the first wiring and the portion overlapping with the second wiring being five times or more the thickness T of the insulating layer."

[0008] Known image recognition techniques using machine learning include methods using deep learning and cascade classifiers. In these methods, a classifier is trained using a training dataset consisting of a region of interest (ROI) and non-ROIs (non-ROIs). The training dataset is created by an engineer providing ROI and non-ROI information from images captured by an imaging device. Because the recognition success rate of a classifier is affected by the training dataset used, selecting ROI and non-ROI information is important. However, Patent Document 2 does not adequately consider how to select the information. Regarding the creation of a training dataset, Patent Document 2 also discloses a method for creating a style transfer model using a 3D model including an arbitrary cross section generated from a layout pattern. However, etching recipes are often developed using test patterns, and the use of layout patterns is limited.

[0009] International Publication No. 2023 / 021540 Japanese Patent Application Laid-Open No. 2011-129485

[0010] Typically, when training a classifier using machine learning, the recognition success rate of the classifier improves as the number of training images included in the training dataset increases. On the other hand, with charged particle beam images, (1) noise reduction is difficult due to limitations in the brightness of the electron gun, and (2) the structure of the object being observed is monotonous due to the device being used to observe nanoscale structures at a magnified scale. The inventors discovered that, due to reasons (1) and (2), increasing the number of training images does not improve the recognition success rate. Furthermore, they found that reducing images unnecessary for training and carefully examining the training dataset are important for improving the recognition success rate. In particular, since there are countless ways to select non-ROIs compared to ROIs, the selection method can significantly affect the recognition success rate. Therefore, the present disclosure aims to provide a technology capable of generating a dataset necessary for building a classifier with guaranteed recognition performance when identifying charged particle beam images.

[0011] In order to solve the above-mentioned problems, one representative dataset generation device of the present invention is a dataset generation device that generates a dataset used for training a classifier that performs image recognition of a charged particle beam image, characterized in that, based on a region of interest (ROI) registered for the image recognition and the charged particle beam image, non-ROI candidates that are regions that do not include the ROI are generated, and the number of non-ROI candidates is reduced based on the ROI while the non-ROI is generated.

[0012] According to the present invention, it is possible to generate a data set necessary for constructing a classifier with guaranteed recognition performance when classifying a charged particle beam image. Other objects and novel features will become apparent from the description of this specification and the accompanying drawings.

[0013] FIG. 1 is a diagram schematically illustrating the configuration of an imaging device when a scanning electron microscope (SEM) is used as the charged particle beam apparatus. FIG. 2 is a diagram illustrating an example of the system configuration of an imaging system for performing electron microscope observation. FIG. 3 is a flowchart for causing the imaging device to perform observation using a classifier in the imaging system 100. FIG. 4 is a diagram illustrating an example of the configuration of an imaging system using the cloud. FIG. 5 is a diagram illustrating an example of the configuration of an automatic imaging system. FIG. 5 is a diagram illustrating an example of a charged particle beam image 501. FIG. 6 is a schematic diagram illustrating the results of image recognition when a classifier is constructed by selecting non-ROIs. FIG. 7 is a schematic diagram illustrating the results of image recognition when a classifier is constructed without selecting non-ROIs. FIG. 8 is a diagram illustrating a setting method for setting a non-ROI candidate region in the horizontal direction of a charged particle beam image. FIG. 9 is a diagram illustrating a setting method for setting a non-ROI candidate region in the vertical direction of a charged particle beam image. FIG. 10 is a diagram illustrating a setting method for setting a non-ROI candidate region in an oblique direction of a charged particle beam image. FIG. 11 is a diagram showing a case where an ROI and a non-ROI candidate region overlap. This is an example of non-ROI generation using the overlap method. FIG. 12 is a diagram showing a case where an ROI 901 and a non-ROI candidate region are included in a charged particle beam image. FIG. 13 is a diagram showing a case where a non-ROI candidate region is set in a charged particle beam image. FIG. 14 is a diagram showing an example where a non-ROI candidate region is moved randomly. FIG. 15 is a diagram showing an example of a GUI when registering an ROI. FIG. 16 is a diagram showing an example of a GUI when adjusting parameters for non-ROI generation.

[0014] The following describes in detail embodiments of the present invention with reference to the accompanying drawings. The disclosure of each embodiment is not limited to the description of the embodiment, and configurations in which the elemental technologies disclosed or suggested in each embodiment are appropriately combined within the scope of the knowledge of a person skilled in the art are also included in the scope of embodiments of the present invention.

[0015] In the present disclosure, a description may be given of a region of interest (ROI) included in an image. For example, when information indicating an ROI is stored in a database or the like without limiting the storage format (e.g., image or coordinate information), the term "save an ROI" may also be used. Similarly, expressions such as "save a non-ROI" and "save a non-ROI candidate region" may also be used to refer to regions of an image that do not include a region of interest (non-ROI) and non-ROI candidate regions. Furthermore, a "non-ROI candidate region" may also be simply referred to as a "non-ROI candidate." Furthermore, in the present disclosure, when an ROI is set in association with a certain image, the term "register an ROI in an image" may also be used. When a non-ROI is set in association with a certain image, the term "register a non-ROI in an image" may also be used.

[0016] 1 is a diagram schematically illustrating the configuration of an imaging device when a scanning electron microscope (SEM) is used as the charged particle beam device. The imaging device 200 includes an electron gun 201, a focusing lens 202, a deflection coil 203, an objective lens 204, a secondary electron detector 205, a sample stage 206, an image forming unit 231, a display unit 235, and a control unit 236. Note that the imaging device 200 may also be referred to as an imaging apparatus 2000, which includes the electron gun 201, the focusing lens 202, the deflection coil 203, the objective lens 204, the secondary electron detector 205, the sample stage 206, the image forming unit 231, the display unit 235, and the control unit 236. For example, an imaging device 2000 that images a sample using a charged particle beam is provided with a classifier that performs image recognition of the charged particle beam image, and non-ROI candidates, which are regions that do not include a region of interest (ROI), are generated based on a region of interest (ROI) registered for image recognition and the charged particle beam image, and the number of non-ROI candidates is reduced based on the ROI to generate non-ROIs, thereby generating a data set to be used for training the classifier. Specific processing will be described later.

[0017] The electron gun 201 is a radiation source that emits an electron beam 207 accelerated by a predetermined acceleration voltage. The emitted electron beam 207 is focused by a focusing lens 202 and an objective lens 204 and irradiated onto a sample 220. A deflection coil 203 deflects the electron beam 207 by a magnetic field or an electric field, thereby scanning the surface of the sample 220 with the electron beam 207.

[0018] The sample stage 206 has a function of moving the sample 220 along a predetermined drive axis or tilting or rotating the sample 220 around a predetermined drive axis in order to move the imaging field of view of the imaging device 200. These functions are realized by driving an actuator such as a motor or a piezoelectric element provided on the sample stage 206.

[0019] The secondary electron detector 205 is, for example, an Everhart-Thornley (ET) detector or a semiconductor detector configured with a scintillator, a light guide, and a photomultiplier tube, and detects secondary electrons 221 emitted from a sample 220 irradiated with the electron beam 207. A detection signal output from the secondary electron detector 205 is transmitted to the image forming unit 231. Note that the imaging device 200 may be provided with a backscattered electron detector that detects backscattered electrons and a transmitted electron detector that detects transmitted electrons, in addition to the secondary electron detector 205.

[0020] The image forming unit 231 forms an observation image of the sample 220 using an AD converter 232 that converts the detection signal output from the secondary electron detector 205 into a digital signal, and a calculator 233 that synchronizes the secondary electron signal with the scanning of the electron beam 207 on the sample 220 using the digital signal output from the AD converter 232 and the deflection coil 203. The observation image formed by the image forming unit 231 is sent to a display unit 235 for display, or sent to a computer system 234 for various processing. The computer system 234 includes, for example, a processor, memory, input / output devices, and storage devices. The input / output devices include input devices such as a keyboard and a mouse for inputting user instructions, and output devices such as a display device. For example, a desired portion of the sample 220 can be observed under specific conditions in cooperation with a control unit 236 that controls the sample stage 206, the display unit 235, and the imaging device 200.

[0021] In the following description, a scanning electron microscope (SEM) will be described as an apparatus for imaging a sample (e.g., a semiconductor device) using a charged particle beam, but the present disclosure is not limited to this case. For example, a focused ion beam (FIB) or a transmission electron microscope (TEM) can also be applied as the imaging device 200. In particular, when the imaging device 200 is a microscope, the imaging device 2000 can be a microscope device 2001. The microscope device 2001 includes the imaging device 2000, and the microscope device 2001 detects an ROI using an image of the charged particle beam, thereby moving the field of view. Specific processing will be described later.

[0022] 2 is a diagram showing an example of the system configuration of an imaging system 100 for performing electron microscope observation. The imaging system 100 has a dataset generation system 101, a classifier construction unit 109, a classifier storage unit 110, an observation recipe input unit 111, an observation recipe display unit 112, an input device 113, and an imaging device 200. Note that, since the imaging system 100 is basically capable of performing imaging without requiring user operation, the imaging system 100 can also be referred to as an automatic imaging system 100. The observation recipe input unit 111 and the observation recipe display unit 112 can also be referred to as an automatic observation recipe input unit 111 and an automatic observation recipe display unit 112, respectively.

[0023] The dataset generation system 101 generates a dataset used for learning (training) a classifier that performs image recognition of a charged particle beam image. The dataset generation system 100 generates a dataset used for learning a classifier that performs image recognition of a charged particle beam image, generates non-ROI candidates that are regions that do not include a region of interest (ROI) based on a region of interest (ROI) registered for image recognition and the charged particle beam image, and generates the non-ROIs while reducing the number of non-ROI candidates based on the ROI. Furthermore, the classifier performs image recognition of a charged particle beam image (sometimes simply referred to as a "charged particle beam image"). The dataset generation system 101 includes an image storage unit 102 that stores charged particle beam images, a dataset storage unit 103 that stores information indicating a region of interest (hereinafter also referred to as "ROI") 106 in the charged particle beam image stored in the image storage unit 102 and information indicating a region not including the ROI (hereinafter also referred to as "non-ROI") 107 (the combination of information indicating the ROI and information indicating the non-ROI is also referred to as a "dataset"). A processor 108 that can access the image storage unit 102 and the dataset storage unit 103, and a memory 114. For example, when a user inputs information indicating a charged particle beam image 105 and an ROI 106 into the dataset generation system 101, the processor 108 generates information indicating a non-ROI 107. The derivation of the non-ROI 107 will be described later. It can also be said that the ROI and non-ROI stored in the dataset storage unit 103 are registered in the charged particle beam image stored in the image storage unit 102. It should be noted that when the data set generation system 101 is configured by hardware, for example, it can also be used as a data set generation device.

[0024] The imaging system 100 constructs an image recognition classifier in a classifier construction unit 109 using the datasets of the ROI 106 and the non-ROI 107 in the dataset generation system 101. The image recognition classifier is stored in a classifier storage unit 110. The image recognition classifier (hereinafter also simply referred to as "classifier") is referenced by an observation recipe input unit 111, and the reference can be confirmed in an observation recipe display unit 112. An observation recipe created using the observation recipe display unit 112 to reflect the user's requests enables image classification based on the classifier. The observation recipe is sent to the imaging device 200 (more specifically, the control unit 236 of the imaging device 200), and observation based on the observation recipe is performed in the imaging system 100. Note that the combination of the observation recipe input unit 111 and the observation recipe display unit 112 can also be referred to as an observation recipe generation unit 115 that generates an observation recipe that enables image classification based on the classifier.

[0025] For example, the data set generation system 101, the classifier construction unit 109, the classifier storage unit 110, and the observation recipe input unit 111 of the imaging system 100 can be arranged in the computer system 234 shown in FIG. 1. The memory 114 can store processing instructions that cause the processor 108 to execute the classifier construction unit 109, the observation recipe input unit 111, and the observation recipe display unit 112. These processing instructions may be installed in the computer system 234 as software. Furthermore, the computer system 234 may have a processor and memory in addition to the processor 108 and the memory 114, and may also have other software installed. For example, the observation software may include software that causes the processor 108 or the computer system 234 to generate an observation recipe.

[0026] Furthermore, when the data set generation system 101 is disposed in the computer system 234 shown in FIG. 1, it is possible to display a GUI on an output device of the computer system 234 and to accept user requests via an input device.

[0027] Although the processor 108 is illustrated as being connected between the input device 113 and the dataset generation system 101, the present disclosure is not limited to this case. The processor 108 can also control, for example, components included in the imaging system 100. For example, in the imaging system 100 that images a sample using a charged particle beam, the processor 108 executes a step of performing image recognition on the charged particle beam image and a step of generating a dataset used for learning to perform the image recognition. The dataset is generated by generating non-ROI candidates, which are regions that do not include a region of interest (ROI) based on a region of interest (ROI) registered for image recognition and the charged particle beam image, and generating the non-ROIs while reducing the number of non-ROI candidates based on the ROI. Specific processing will be described later.

[0028] The processor 108 includes, for example, a CPU (Central Processing Unit) or a GPU (Graphical Processing Unit). Although the case where there is one processor 108 is shown, the processor 108 may include one or more processors. The dataset storage unit 103 and the classifier storage unit 110 include an HDD (Hard Disk Drive), an SSD (Solid State Drive), and memory. The dataset storage unit 103 and the classifier storage unit 110 are also not limited to being configured with a single storage device, and may have a storage device arranged for each type of data, for example.

[0029] Although the imaging system 100 has been described above, it is also possible to include the imaging system 100 in the manufacturing apparatus 10 that manufactures semiconductor devices, thereby calling the imaging system 100 a semiconductor device manufacturing system 1. In the semiconductor device manufacturing system 1, the charged particle beam image 105 can also be acquired by imaging the semiconductor device manufactured by the manufacturing apparatus 10 with the imaging device 200. For example, the imaging device 200 images the semiconductor device manufactured by the manufacturing apparatus 10 based on an observation recipe described below, and acquires a charged particle beam image. Each time the etching recipe in the manufacturing apparatus 10 is changed, a charged particle beam image is acquired with the imaging device 200 and the structure of the semiconductor device is analyzed, thereby developing an etching recipe suitable for obtaining a desired structure.

[0030] 3 is a flowchart showing how the imaging device 200 performs observation using a classifier in the imaging system 100. The processor 108 executes input and output of data to and from the data set generation system 101 and controls the entire flowchart.

[0031] First, the processor 108 performs a charged particle beam image and ROI determination 301. For example, a user selects a charged particle beam image 105 and an ROI 106 to be observed via the input device 113. At this time, a region in the charged particle beam image 105 that is to be image-recognized during observation is also saved (registered) as the ROI 106. The ROI 106 may be saved by saving an image file of the ROI 106 or by saving a text file specifying coordinates within the charged particle beam image 105. The processor 108 stores the charged particle beam image 105 input from the input device 113 in the image storage unit 102, and also stores the ROI 106 in the dataset storage unit 103.

[0032] Next, the processor 108 performs non-ROI candidate region calculation 302 using the charged particle beam image 105 and the ROI 106. In non-ROI candidate region calculation 302, the processor 108 generates multiple non-ROI candidate regions based on the charged particle beam image 105 and the ROI 106. Specifically, the candidate region for the non-ROI 107 is automatically calculated by the processor 108, basically without requiring user specification or the like. The result of non-ROI candidate region calculation 302 may be temporarily stored in the memory 114 or may be saved in the dataset storage unit 103. The processor 108 performs non-ROI generation 303 using the result of non-ROI candidate region calculation 302, and information indicating the generated non-ROI 107 is saved in the dataset storage unit 103.

[0033] Next, the processor 108 determines whether the number of generated non-ROIs 107 is greater or less than a threshold value (306). If the number of non-ROIs 107 is greater than the reference value, the processor 108 adjusts the parameters in non-ROI generation parameter adjustment (309) and again executes non-ROI generation (303). The processor 108 generates a non-ROI by reducing the number of multiple non-ROI candidate regions based on the ROI 106.

[0034] If the number of non-ROIs 107 is appropriate, the processor 108 executes classifier construction 304. In classifier construction 304, the non-ROIs 107 and ROIs 106 stored in the dataset storage unit 103 are used, and a classifier required for image recognition is constructed in a classifier construction unit 109. The constructed classifier is stored in a classifier storage unit 110. The classifier stored in the classifier storage unit 110 is referenced from an observation recipe input unit 111. The observation recipe includes, for example, observation conditions ranging from low-magnification observation to high-magnification observation, and the observation conditions are displayed in an observation recipe display unit 112.

[0035] (Configuration example utilizing cloud computing) It is also possible to configure the imaging system 100 shown in Fig. 2 as an observation system in which the observation functions excluding the imaging device 200 are implemented as applications on a platform. An example of such implementation is shown in Fig. 4.

[0036] 4 is a diagram showing an example of the configuration of an image capturing system 100A using the cloud. The image capturing system 100A includes an image capturing device 200, a platform 400, and a terminal 410.

[0037] The platform 400 is implemented with an application for generating a dataset used for training a classifier that performs image recognition of a charged particle beam image. Specifically, the platform 400 includes a database 401, an OS 402, middleware 403, an observation recipe creation application 405, and a dataset generation application 406. The platform 400 is built on the cloud, and applications that execute processing run on the OS 402 and the middleware 403. In the platform 400, the observation recipe creation application 405 executes processing corresponding to the classifier construction unit 109, the observation recipe input unit 111, and the observation recipe display unit 112 of the imaging system 100, and the dataset generation application 406 executes processing corresponding to the dataset generation system 101 of the imaging system 100. A user can access the platform from a terminal 410 via a network and use the functions of the application built on the platform 400. The platform 400 includes a database 401, which stores data necessary for executing the application. Furthermore, the imaging device 200 is also connected via a network to enable data exchange with the platform 400. Note that the imaging system 100A and the observation recipe creation application 405 can also be referred to as the automatic imaging system 100A and the automatic observation recipe creation application 405. The platform 400 may be a server. Furthermore, the semiconductor device manufacturing apparatus 1 may be configured to include the platform 400 on which an application for generating a data set used for training a classifier that performs image recognition of charged particle beam images is implemented.

[0038] The ROI 502 and non-ROI 503 in the charged particle beam image 501 will be described using FIG. 5 . FIG. 5 is a diagram showing an example of the charged particle beam image 501. The charged particle beam image 501 is, for example, an image of a cross section of a semiconductor device, and shows two insulating layers 505 extending in the horizontal direction (x-axis direction), ten trenches 504 extending in the vertical direction (y-axis direction), and fourteen electrode portions 506. The ROI 502 is shown surrounded by a solid-line frame, and the non-ROI 503 is shown surrounded by a dashed-line frame. Note that the horizontal direction (x-axis direction) and the vertical direction (y-axis direction) are set for convenience in describing the charged particle beam image. Furthermore, descriptions of the trenches 504, insulating layers 505, and electrode portions 506 may be omitted to facilitate understanding of the present disclosure.

[0039] The ROI 502 includes a structure to be image-recognized in the imaging system 100, and in FIG. 5 , as an example, the ROI 502 is an area including the tips of two trenches 504. There are five ROIs 502 in FIG. 5 . Non-ROIs 503 are indicated by areas surrounded by dotted lines in FIG. 5 . The non-ROIs 503 are areas in the charged particle beam image 501 excluding the ROIs 502, and FIG. 5 shows 10 areas as an example. The method for setting the non-ROIs 503 is not limited to the case shown in FIG. 5 , and other areas may be set. In conventional methods, when the classifier construction unit 109 constructs a classifier using many non-ROI 503 areas, there is a possibility that a classifier with a low recognition success rate will be constructed.

[0040] The recognition success rate will be described with reference to FIGS. 6 and 7 . FIG. 6 is a schematic diagram 601 showing the results of image recognition when a classifier is constructed by selecting non-ROIs. ROIs are detected using the classifier, and four detection regions 602, which are detected regions, are generated for a certain charged particle beam image. On the other hand, FIG. 7 is a schematic diagram 603 showing the results of image recognition when a classifier is constructed without selecting non-ROIs. As a result of detecting ROIs using the classifier, only one detection region 602 is generated for the same charged particle beam image as in FIG. 6, and the three detection regions 602 found in FIG. 6 are not included. It can be said that the recognition success rate is high for the classifier in FIG. 6 and low for the classifier in FIG. 7 .

[0041] (Method for generating non-ROI candidate regions) There are countless ways to select non-ROIs 503, but a high recognition success rate can be achieved by using an appropriate generation method. In this disclosure, a method is described in which multiple non-ROI candidate regions are first generated, and then three processes are combined to select a non-ROI from the multiple non-ROI candidate regions. Note that, in the following description, a case is described in which all non-ROI candidate regions have a common size, but the present disclosure is not limited to this case. The present disclosure can also be applied to cases in which the non-ROI candidate regions do not have a common size.

[0042] (Calculation of Non-ROI Candidate Region) A method for determining a non-ROI candidate region will be described with reference to FIGS. 8 to 10 . FIGS. 8 to 10 are diagrams illustrating calculation of the non-ROI candidate region 302 in FIG. 3 . FIG. 8 is a diagram illustrating a setting method for setting a non-ROI candidate region in the horizontal direction of a charged particle beam image. In FIG. 8 , non-ROI candidate regions in a charged particle beam image 700 are surrounded by dotted lines. Let us assume that there are a first candidate region 701x, a second candidate region 702x, and a third candidate region 703x. If rx is the horizontal size of the non-ROI and dx is the distance from the right end (the side on the positive x-axis direction) of the first candidate region 701 to the left end (the side on the negative x-axis direction) of the second candidate region 702, then the horizontal distance Lx1 from the first candidate region 701x to the second candidate region 702x is Lx1 = rx + dx.

[0043] Here, the horizontal distance Lx2 from the second candidate area 702x to the third candidate area 703x is defined using Lx1 as Lx2 = a × Lx1 (a ≥ 1). For example, when a = 1 and dx = 0, the first candidate area 701x to the third candidate area 703x are set adjacent and equidistant from each other. When a = 1 and dx ≠ 0, the candidate areas are set dx apart.

[0044] 8 shows three non-ROI candidate regions, but if the non-ROI candidate regions are small relative to the charged particle beam image 700, the non-ROI candidate image can be positioned anywhere between the third candidate region 703x and the horizontal end of the charged particle beam image 700. Therefore, the distance Lx(m+1) from the mth (m is an integer equal to or greater than 2) candidate region (the third or subsequent candidate region) to the (m+1)th candidate region can be defined as Lx(m+1)=a×Lxm (a>0). By changing the parameters dx and a, the non-ROI region can be set in the horizontal direction.

[0045] Next, a method for setting non-ROI candidate regions in the vertical direction will be described using FIG. 9 . FIG. 9 is a diagram showing a setting method for setting non-ROI candidate regions in the vertical direction of a charged particle beam image. Non-ROI candidate regions can also be set in the vertical direction using a method similar to that for the horizontal direction. Let us assume that the non-ROI candidate regions are a first candidate region 701y, a second candidate region 702y, and a third candidate region 703y. If ry is the vertical size of the non-ROI and dy is the distance from the top end (the side on the positive y-axis direction) of the first candidate region 701y to the bottom end (the side on the negative y-axis direction) of the second candidate region 702y, then the distance Ly1 from the first candidate region 701y to the second candidate region 702y and the distance Ly2 from the second candidate region 702y to the third candidate region 703y can be defined as Ly1 = ry + dx and Ly2 = b × Ly1 (b≧1), respectively. Similarly to the horizontal direction, the distance Ly(n+1) from the nth (n is an integer equal to or greater than 2) candidate region (the third or subsequent candidate region) to the (n+1)th candidate region can be defined as Ly(n+1) = b × Lyn (b > 0). By changing the parameters dy and b, non-ROI candidate regions can be set in the vertical direction.

[0046] Fig. 10 is a diagram showing a setting method for setting non-ROI candidate regions in diagonal directions on a charged particle beam image. While Fig. 8 shows the horizontal shift amount separately, and Fig. 9 shows the vertical shift amount separately, a method for setting non-ROI candidate regions in diagonal directions by combining the horizontal and vertical directions, as shown in Fig. 10, is also conceivable. Fig. 10 shows an example of setting non-ROI candidate regions in diagonal directions, in which a first candidate region 701d, a second candidate region 702d, and a third candidate region 703d are set.

[0047] The non-ROI candidate regions are calculated by the processor 108 of the dataset generation system 101 and stored in the dataset storage unit 103 or the memory 114. The format of the information indicating the candidate regions can be an image file or a text file using coordinate information. When a text file is used, the name of the charged particle beam image (e.g., file name) and the coordinate information of the non-ROI candidate regions are stored in a format that can be referenced in non-ROI generation (non-ROI generation 303 in FIG. 3) (the reference format will be described later using FIGS. 15 and 16).

[0048] (Non-ROI Generation) After setting the non-ROI candidate regions, the processor 108 generates the non-ROI (non-ROI generation 303 in FIG. 3 ). Three methods are used to generate the non-ROI: the overlap method, the ROI correlation method (ROI-non-ROI correlation coefficient method), and the non-ROI correlation method (non-ROI-non-ROI correlation coefficient method). In generating the non-ROI, the non-ROI is selected from the non-ROI candidates based on the ratio of the overlapping area between each ROI and each non-ROI candidate (overlap rate OR). In addition, the selection of the non-ROI from the non-ROI candidates is performed based on the first correlation coefficient (correlation coefficient of the ROI-non-ROI correlation coefficient method) associated with each ROI and each non-ROI candidate. In other words, the selection of the non-ROI from the non-ROI candidates is performed by selecting two non-ROI candidates from the non-ROI candidates and based on a second correlation coefficient (correlation function of the non-ROI-non-ROI correlation coefficient method) between one of the two selected non-ROI candidates and the other of the two selected non-ROI candidates. Hereinafter, the generation of the non-ROI (non-ROI generation 303 and determination 306 in FIG. 3) will be described with reference to FIGS. 11 and 12 .

[0049] (Overlap Method) Generation of a non-ROI using the overlap method will be described using FIG. 11 . FIG. 11 is a diagram illustrating a case where an ROI and a non-ROI candidate region overlap. In a charged particle beam image 800, an ROI 801 is indicated by a solid-line frame, and a non-ROI candidate region 802 is indicated by a dotted-line frame. Here, the overlapping region between the ROI 801 and the non-ROI candidate region 802 is referred to as an overlap region 803. If the horizontal length of the overlap region 803 is ex and the vertical length is ey, the area SO of the overlap region is S = ex × ey. If the horizontal length of the ROI 801 is rx and the vertical length is ry, the area SR of the ROI 801 is SR = rx × ry. Here, if the non-ROI candidate region does not overlap with the ROI 801, SO = 0, and the overlap rate is 0. When the number of ROIs registered in the charged particle beam image 800 is NR (NR is an integer) and the number of non-ROI candidate regions is NS (NS is an integer), the number NCO of combinations of both for calculating the overlap rate OR is NCO = (NR × NS) / 2. The NCO overlap rates are calculated, and a threshold TOL for the overlap rate OR is set as a parameter to generate non-ROIs.

[0050] The overlap rate OR ranges from 0 to 1, and when OR = 1, the ROI is covered by the non-ROI candidate region. The non-ROI candidate region for which OR = 1 is not included in the generated non-ROI. The overlap rate threshold TOL is adjusted in the non-ROI generation parameter adjustment 309 of FIG. 3 based on the target number of non-ROIs to be generated. For example, OR = 0.5 is used as the initial value. The overlap rate threshold TOL is adjusted in the non-ROI generation parameter adjustment 309 so as to reduce the difference between the number of non-ROIs calculated in the non-ROI generation 303 and the target number of non-ROIs to be generated.

[0051] The calculation result of the overlap rate OR is saved, for example, in a text file, and is held in a format that can be referenced when generating a non-ROI, together with the name of the charged particle beam image and coordinate information of the non-ROI candidate region.

[0052] (ROI-Non-ROI Correlation Coefficient Method) A method for selecting a non-ROI from non-ROI candidate regions using a correlation coefficient will be described with reference to FIG. 12 . FIG. 12 is a diagram showing a case where a charged particle beam image 900 includes an ROI 901 and non-ROI candidate regions 902 and 903. First, a selection method using the correlation coefficient between the ROI 901 and the non-ROI candidate region 902 (hereinafter referred to as the "ROI-non-ROI correlation coefficient method") will be described, and then a method for selecting and selecting two non-ROI candidate regions 902 and 903 (hereinafter referred to as the "non-ROI-non-ROI correlation coefficient method") will be described.

[0053] First, in the ROI-non-ROI correlation coefficient method, a correlation coefficient is calculated between an ROI 901 registered in a charged particle beam image 900 and a non-ROI candidate region 902. Here, a case will be described in which the non-ROI candidate region 902 and the ROI 901 have the same number of pixels, but the present disclosure can also be applied to cases in which the number of pixels differs. For example, if there is a difference between the two, one is converted to the same number of pixels as the other before calculation.

[0054] There are several methods for calculating the correlation coefficient, with four typical methods being the sum of squared differences, the sum of absolute differences, the normalized correlation coefficient, and the average normalized correlation coefficient. The method using the sum of squared differences calculates the sum of squared errors of pixel values ​​for each cell in two regions. Let T(x, y) be the pixel value at horizontal position x and vertical position y in the ROI 901, I(x, y) be the pixel value at horizontal position x and vertical position y in the non-ROI candidate region 902, rx be the number of pixels in the horizontal direction of the ROI 901, and ry be the number of pixels in the vertical direction. The correlation coefficient RSSD using the sum of squared differences is expressed by the following equation: In RSSD, when a non-ROI candidate region 902 completely matches the ROI 901, RSSD=0. Since it is necessary to select non-ROIs generated by the processor 108 that do not match the ROI 901, the RSSD threshold for selecting non-ROI candidate regions is set to a value greater than 0.

[0055] Next, while the method of calculating the sum of squared differences involves squaring, there is also a method of adding up the absolute value differences. The sum of absolute differences RSAD is expressed by the following formula: The RSAD method requires no square calculations and therefore requires a short calculation time. As with RSSD, when the ROI 901 and the non-ROI candidate region 902 completely coincide with each other, RSAD=0. Therefore, the threshold value for selecting the non-ROI candidate region is set to a value greater than 0.

[0056] It is also possible to use normalized cross-correlation RNCC as the correlation function. The RNCC of the ROI 901 and the non-ROI candidate region 902 is expressed by the following equation: The RNCC value falls within the range of 0 to 1, and the closer the value is to 1, the more similar the images are determined to be.

[0057] In addition, among the normalized cross-correlation methods, an average normalized cross-correlation (RZNCC) may be used, which calculates a correlation coefficient by subtracting the average of the pixel values ​​of the entire region from each pixel value. When the average of all pixel values ​​of the ROI 901 is MI and the average of all pixel values ​​of the non-ROI candidate region 902 is MT, RZNCC is expressed by the following formula: RZNCC is a value ranging from -1 to 1.

[0058] As with the calculation of the overlap rate OR, the correlation coefficient derived by the ROI-non-ROI correlation coefficient method (hereinafter also referred to as the "ROI-non-ROI correlation coefficient") is calculated for only NCO results. In selecting non-ROIs, it is necessary to select images that are different from non-ROIs from the ROI candidate regions. The threshold value TCR1 for the ROI-non-ROI correlation coefficient is set to a value greater than 0 for RSSD and RSAD. It is also set to a value close to 0 for RNCC and close to -1 for RZNCC. By setting the threshold value TCR1 in accordance with the calculation method, non-ROIs suitable for constructing a classifier are selected from the non-ROI candidate regions.

[0059] As a result of the calculation of the correlation coefficients, NCO correlation coefficients are stored in the data set storage unit 103. As with the overlap ratio, the calculation results of the correlation coefficients are stored in a text file, and are held in the data set storage unit 103 in a format that can be referenced, together with the name of the charged particle beam image and coordinate information of the non-ROI candidate region.

[0060] (Non-ROI-Non-ROI Correlation Coefficient Method) Next, a method of selecting two non-ROI candidate regions, calculating a correlation coefficient, and sorting non-ROI regions, i.e., the non-ROI-non-ROI correlation coefficient method, will be described. The calculation of the correlation coefficient derived by the non-ROI-non-ROI correlation coefficient method (hereinafter also referred to as the "non-ROI-non-ROI correlation coefficient") is the same as that of the ROI-non-ROI correlation coefficient, and is performed using a non-ROI candidate region 902 and another non-ROI candidate region 903 as the calculation targets. Furthermore, similar to the ROI-non-ROI correlation coefficient, RSSD (sum of squared differences), RSAD (sum of absolute differences), RNCC (normalized correlation coefficient), and RZNCC (mean normalized correlation coefficient) can be applied to the non-ROI-non-ROI correlation coefficient. In the case of the ROI-non-ROI correlation coefficient method, the correlation coefficient is calculated between the ROI 901 and the non-ROI candidate region 902, but in the case of the non-ROI-non-ROI correlation coefficient method, the correlation function is calculated by using the number of pixels in the non-ROI candidate region 903 instead of the ROI 901.

[0061] When two non-ROI candidate regions are selected from multiple (NS) non-ROI candidate regions, the number of combinations NCC is NCC = (NS × (NS - 1)) / 2, and therefore the non-ROI-non-ROI correlation coefficients for the NCC combinations are calculated. The threshold TCR2 of the non-ROI-non-ROI correlation coefficient for sorting out non-ROIs differs depending on the method of calculating the correlation coefficient, and is a value greater than 0 for RSSD and RSAD, a value between 0 and 1 for RCNN, and a value between -1 and 1 for RZNCC. Setting the threshold TNCC makes it possible to reduce the number of non-ROIs, thereby preventing a decrease in the recognition success rate for charged particle beam images of the classifier constructed by the classifier construction unit 109 depending on the number of non-ROIs generated.

[0062] Furthermore, in selecting non-ROIs, in order to avoid the influence of noise when constructing the classifier, it is necessary to avoid including similar images among the generated non-ROIs. Taking such influences into consideration, the threshold TCR2 for the non-ROI-non-ROI correlation coefficient is set to a value greater than 0 for RSSD and RSAD, 0 for RNCC, and close to -1 for RZNCC, just as in the case of the threshold TCR1.

[0063] The number of candidate non-ROI regions NS to be generated is generally set to approximately two to five times the number of ROIs. This number is affected by the amount of noise in the charged particle beam image and the number of gradations in the charged particle beam image. The less noise there is and the wider the gradations in the charged particle beam image, the more non-ROIs can be used to train the classifier. The number of non-ROIs to be generated is adjusted using parameters including the overlap rate threshold TOL, the ROI-non-ROI correlation coefficient threshold TCR1, and the non-ROI-non-ROI correlation coefficient threshold TCR2.

[0064] 3 is repeatedly performed using the processor 108 until the number of non-ROIs becomes appropriate, and the parameters and the number of generated non-ROIs at that time are stored as history in the memory 114 and the data set storage unit 103. In adjusting the parameters, for example, the steepest descent method can be applied based on changes in parameter setting values ​​and the number of non-ROIs generated by those parameters until the number of non-ROIs reaches a target number.

[0065] Conventionally, non-ROIs have been created by utilizing the experience and know-how of a user, but in the present disclosure, by setting parameters, it becomes possible to create a non-ROI without relying on experience or know-how. Furthermore, by learning (training) a classifier using the non-ROI created in this way, it becomes possible to build a classifier that ensures a certain level of classification performance.

[0066] (Classifier Construction) A system for learning a classifier from a dataset generated by the dataset generation system 101 will be described. Classifier construction 304 shown in FIG. 3 is performed using the generated non-ROIs and registered ROIs. The classifier construction method can be set depending on the type of charged particle beam image, the structure of the semiconductor device, and the like. For example, a cascade classifier may be used as the classifier, or RCNN (Region Based Convolutional Neural Networks) or Yolo (You Only Look Once), which utilize deep learning, may be used to construct the classifier. The classifier construction unit 109 constructs the classifier using the processor 108. The device used to construct the classifier varies depending on the construction method, and may be a central processing unit (CPU) or a graphical processing unit (GPU). The classifier constructed by the CPU or GPU is stored in the classifier storage unit 110. The classifier storage unit 110 may include, for example, a hard disk drive (HDD), a solid state drive (SSD), or memory, or may be cloud storage.

[0067] (Observation Recipe Creation and Observation) An imaging system equipped with a classifier learning system will be described. The classifier constructed by the learning system (dataset generation system 101 and classifier construction unit 109) is stored in the classifier storage unit 110, referenced in the observation recipe, and used to classify images acquired by the imaging device 200. With regard to observation recipe creation 307 shown in FIG. 3 , the observation recipe can be created using observation software installed in the computer system 234. The observation software also includes a GUI (graphical user interface), which is displayed on the observation recipe display unit 112. In creating the observation recipe, condition settings for causing the control unit 236 of the imaging device 200 to perform functions required for observation, such as stage movement, focus adjustment, magnification setting, and image shift movement, are performed in the observation recipe input unit 111 and the observation recipe display unit 112. As will be described later, by pressing a recipe execution button displayed on the observation recipe display unit 112, the condition settings for the various functions of the observation recipe are transferred to the control unit 236 of the imaging device 200, and imaging of the sample based on the observation recipe is performed in the imaging device 200. The charged particle beam image captured by the imaging device 200 is subjected to image recognition using the constructed classifier, and an ROI in the charged particle beam image is detected. According to the position of the detected ROI, stage movement, focus adjustment, magnification setting, image shift movement, etc. are performed in the imaging device 200, and observation is performed under conditions suitable for analyzing the ROI.

[0068] In this way, in the present disclosure, an identifier having a certain level of identification performance can be used, making it possible to extract an ROI from a charged particle beam image, and basically making it possible to observe a charged particle beam image without requiring user operation.

[0069] In the first embodiment, a method for setting a non-ROI candidate region in the horizontal direction, the vertical direction, or a diagonal direction that is a combination of these directions was described for acquiring a non-ROI candidate region (calculating a non-ROI candidate region 302 in FIG. 3 ). In contrast, in the second embodiment, a method for using a random variable to set a non-ROI candidate region will be described with reference to FIGS. 13 and 14 . In the following description, components that are the same as or equivalent to those in the first embodiment above will be assigned the same reference numerals, and their description will be simplified or omitted.

[0070] FIG. 13 is a diagram illustrating the setting of a non-ROI candidate region in a charged particle beam image 1000. As shown in FIG. 13, a first non-ROI candidate region 1001 is set in the charged particle beam image 1000. The non-ROI candidate region 1001 is a rectangular region surrounded by dotted lines with a horizontal length rx and a vertical length ry. First, non-ROI candidate regions are set in a grid pattern within the charged particle beam image 1000, with intervals of hx in the horizontal direction and hy in the vertical direction. A second non-ROI candidate region 1002 is one of the non-ROI candidate regions arranged in the grid pattern. Next, each non-ROI candidate region is moved by a length sx in the horizontal direction and a length sy in the vertical direction. sx and sy are set for each non-ROI candidate region using, for example, random variables. The random variables can be generated, for example, from a library stored in the computer system 234.

[0071] FIG. 14 is a diagram showing an example of randomly moving non-ROI candidate regions. FIG. 14 shows the distribution of non-ROI candidate regions after the movement, with the first non-ROI candidate region 1001 in FIG. 13 becoming the first non-ROI candidate region 1003 after the movement in FIG. 14, and the second non-ROI candidate region 1002 in FIG. 13 becoming the second non-ROI candidate region 1004 after the movement in FIG. 14. While FIGS. 13 and 14 illustrate the case where the distances hx and hy are set so that the non-ROI candidate regions do not overlap, hx and hy may also be set to a size such that the non-ROI candidate regions overlap, and this method also allows for the setting of sufficient non-ROI candidate regions. The generated non-ROI candidate regions are stored in the dataset storage unit 103 in the form of a text file. By calculating the overlap rate OR, the ROI-non-ROI correlation coefficient, and the non-ROI-non-ROI correlation coefficient, it is possible to generate a non-ROI by selecting from the generated non-ROI candidate regions.

[0072] In Example 3, a GUI of the dataset generation system 101 will be described. The GUI is displayed, for example, on an output device in the computer system 234 of FIG. 1. When generating a non-ROI in the dataset generation system 101, a charged particle beam image and an ROI are registered by the user. FIG. 15 is a diagram showing an example of a GUI for registering an ROI. In the following description, a case will be described in which an input form into which information is input by the user, and check boxes and buttons that can be selected by clicking a mouse are included. In the following description, components that are the same as or equivalent to those in Example 1 above will be assigned the same reference numerals, and their description will be simplified or omitted.

[0073] The ROI data registration window 1100 includes an upload button 1102. Clicking this button displays a charged particle beam image in a charged particle beam image display area 1103. The file name of the currently displayed charged particle beam image is displayed in an input image name display area 1101. If the charged particle beam image has observation magnification information, the magnification is displayed in a magnification display area 1111. An example of a registered ROI area is the registration area 1106, which is highlighted with a dotted frame and has a registered number written in the upper left corner of the frame. In this example, two ROIs identified by numbers 1 and 2 are registered along with the charged particle beam image. To register an area, for example, the user drags a guide area 1104 displayed in a thick frame with the mouse cursor 1105 positioned therein, places the guide area 1104 at a desired position on the charged particle beam image, and clicks (selects) the guide area 1104. This registers the area specified by the guide area 1104 as an ROI. To the right of the ROI data registration window 1100 is a region size adjustment function display 1110, which allows the size of the region to be registered to be changed in the horizontal and vertical directions.

[0074] The magnification link button 1107 is a button for automatically changing the size of the guide region of the registered region from the magnification of the previous charged particle beam image and the magnification of the new charged particle beam image when there are two or more charged particle beam images. When the magnification link button 1107 is selected, if the guide region 1104 before the observation magnification of the imaging device 200 is changed is set to p1x pixels in the horizontal direction, p1y pixels in the vertical direction, and m1 times the magnification, and if the guide region after the observation magnification is changed is set to p2x pixels in the horizontal direction, p2y pixels in the vertical direction, and m2 times the magnification, then the guide region 1104 is enlarged or reduced by calculating p2x = m2 / m1 × p1x and p2y = m2 / m1 × p1y.

[0075] When the Square button 1108 is clicked, the shape of the guide region 1104 is unified to a square. The training data tag region 1109 indicates tag information assigned as information in the registration region 1106. Examples of tag information that can be used include information about the structure of the semiconductor device, such as trenches or fins, information about the type of semiconductor device, such as logic or memory, and information about the cross-sectional location, such as the top or bottom. In the case of the imaging system 100A shown in FIG. 4 , by storing the region of the ROI registered in the database 401 in the dataset generation application 406 on the platform 400, it becomes possible to search for the region of the ROI and the charged particle beam image based on the tag information. By constructing a classifier using registered images searched from the database 401 based on information such as logic and fins, it is possible to construct a classifier with a high recognition success rate even when the number of charged particle beam image samples is small.

[0076] In addition, the user may have knowledge regarding the generation of non-ROIs necessary for constructing a classifier, for example, knowledge regarding non-ROIs to be selected from the information on the charged particle beam image and the registered ROIs. In this case, the ROI data registration window 1100 shown in FIG. 15 can also be used to select candidate non-ROI regions. For example, by displaying candidate non-ROI regions on the charged particle beam image in the charged particle beam image display area 1103 and allowing the user to select them, the user's knowledge regarding non-ROIs can be reflected in the generation of non-ROIs.

[0077] In Example 4, another example of the GUI of the dataset generation system 101 will be described. The GUI is displayed, for example, on an output device in the computer system 234 of FIG. 1. In the non-ROI generation parameter adjustment 309 of FIG. 3, as a method for adjusting the threshold TOL of the overlap rate OR, the threshold TCR1 of the ROI-non-ROI correlation coefficient, and the threshold TCR2 of the non-ROI-non-ROI correlation coefficient so as to achieve a target number of non-ROIs, a method in which the user uses a GUI to perform adjustments is also available. Adjustment of generation parameters using a GUI will be described using FIG. 16. In the following description, components that are the same or equivalent to those in Example 1 above will be assigned the same reference numerals, and their description will be simplified or omitted.

[0078] 16 is a diagram showing an example of a GUI for adjusting parameters for non-ROI generation. A non-ROI data generation parameter adjustment window 1200 includes a "show" button for calling up a charged particle beam image for registering a non-ROI, and a file name display section 1201 for displaying the name of the called file. For example, a user selects the "show" button and selects a charged particle beam image to register according to the displayed file dialog. The name of the selected charged particle beam image is displayed in the file name display section 1201, and the image itself is also displayed on the GUI. A base non-ROI display window 1202 displays the selected charged particle beam image and a non-ROI region 1203 generated using the reference parameters within a dotted frame.

[0079] The reference parameter display area 1205 displays the overlap rate OR threshold TOL, the ROI-non-ROI correlation coefficient threshold TCR1, and the non-ROI-non-ROI correlation coefficient threshold TCR2, and also displays the number of non-ROIs calculated in each case. Furthermore, the new parameter display area 1204 displays the non-ROI regions generated using the parameters adjusted by the user. The user adjusts the values ​​of TOL, TCR1, and TCR2 using the adjustment buttons in the parameter adjustment area 1207. When the apply button is selected, the non-ROI regions generated using the adjusted parameters are displayed in the new parameter display area 1204, allowing the user to confirm the results. Furthermore, the number of non-ROIs calculated using the adjusted parameters is displayed in the new parameter non-ROI count display area 1206.

[0080] When the Save button 1208 is selected, the adjusted parameters are reflected in the reference parameters. If you are having trouble setting the initial parameters, you can enter the number of non-ROIs you want to generate in the target non-ROI number display area 1210 and click the Auto button 1209 to view the automatically generated non-ROI areas in the new parameter non-ROI number display area 1206. The algorithm for automatic generation can be, for example, a method of adjusting three thresholds using the steepest descent method until the target number of non-ROIs is reached.

[0081] (Actions and Effects) As described above, according to the present disclosure, when identifying a charged particle beam image, it is possible to generate a data set necessary for constructing a classifier with guaranteed recognition performance. In the present disclosure, by setting parameters, it is possible to create a non-ROI without relying on the user's experience or know-how. Furthermore, since a classifier with a certain level of recognition performance can be used, it is possible to extract an ROI from a charged particle beam image, and it is basically possible to observe a charged particle beam image without requiring user operation.

[0082] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described examples, and various modifications are possible within the scope of the gist of the present invention.

[0083] The following are aspects that may be included in the present invention, but the present invention is not limited to these. (Aspect 1) A dataset generation device that generates a dataset used for training a classifier that performs image recognition of a charged particle beam image, the dataset generation device being characterized in that: Regions of Interest (ROIs) registered for the image recognition are generated based on the ROIs and the charged particle beam images, and non-ROI candidates that are regions that do not include the ROIs are generated based on the ROIs, and the non-ROIs are generated while reducing the number of non-ROI candidates based on the ROIs. (Aspect 2) The dataset generation device described in Aspect 1, the dataset generation device being characterized in that the non-ROIs are selected from the non-ROI candidates based on the proportion of overlapping regions between each of the ROIs and each of the non-ROI candidates for each of the ROIs. (Aspect 3) The dataset generation device according to Aspect 1 or Aspect 2, wherein the non-ROI is selected from the non-ROI candidates based on a first correlation coefficient associated with each of the ROIs and each of the non-ROI candidates. (Aspect 4) The dataset generation device according to any one of Aspects 1 to Aspect 3, wherein two non-ROI candidates are selected from the non-ROI candidates, and the non-ROI is selected from the non-ROI candidates based on a second correlation coefficient associated with one of the two selected non-ROI candidates and the other of the two selected non-ROI candidates. (Aspect 5) The dataset generation device according to any one of Aspects 1 to Aspect 4, wherein the classifier is a cascade classifier. (Aspect 6) A dataset generation method for generating a dataset used to train a classifier that performs image recognition of a charged particle beam image, the dataset generation method comprising: a step of generating non-ROI candidates, which are regions that do not include a Region of Interest (ROI) registered for the image recognition, based on the charged particle beam image; and a step of generating non-ROIs based on the ROI while reducing the number of non-ROI candidates.(Aspect 7) An imaging device for imaging a sample using a charged particle beam, comprising: a classifier for performing image recognition of the image of the charged particle beam; and based on a region of interest (ROI) registered for the image recognition and the image of the charged particle beam, non-ROI candidates, which are regions that do not include the ROI, are generated, and the number of non-ROI candidates is reduced based on the ROI to generate non-ROIs, thereby generating a dataset used for training the classifier. (Aspect 8) An imaging method for imaging a sample using a charged particle beam, comprising the steps of: performing image recognition of the image of the charged particle beam; and generating a dataset to be used for learning for performing the image recognition, wherein the dataset is generated by generating non-ROI candidates, which are regions that do not include a Region of Interest (ROI) registered for the image recognition, based on the image of the charged particle beam and the ROI, and generating non-ROIs while reducing the number of non-ROI candidates based on the ROI. (Aspect 9) A microscope apparatus equipped with the imaging device according to Aspect 7, wherein a field of view is moved by detecting the ROI using the image of the charged particle beam. (Aspect 10) A dataset generation system for generating a dataset used to train a classifier that performs image recognition of a charged particle beam image, characterized in that the following steps are executed by a processor: generating a region of interest (ROI) registered for the image recognition and non-ROI candidates that are regions that do not include the ROI based on the charged particle beam image; and generating non-ROIs based on the ROI while reducing the number of non-ROI candidates.(Aspect 11) An imaging system for imaging a sample using a charged particle beam, wherein a step of performing image recognition of the image of the charged particle beam and a step of generating a dataset to be used for learning to perform the image recognition are executed by a processor, and the dataset is generated by generating non-ROI candidates, which are regions that do not include a Region of Interest (ROI) registered for the image recognition and the image of the charged particle beam, and generating non-ROIs while reducing the number of non-ROI candidates based on the ROI. (Aspect 12) A semiconductor device manufacturing system including a platform on which an application is implemented for generating a dataset to be used for training a classifier that performs image recognition of charged particle beam images, wherein the application executes the following steps: generating non-ROI candidates that are regions that do not include a Region of Interest (ROI) registered for the image recognition and based on the charged particle beam image; and generating non-ROIs based on the ROI while reducing the number of non-ROI candidates. (Aspect 13) The semiconductor device manufacturing system according to Aspect 12, wherein the platform is a server. (Aspect 14) The semiconductor device manufacturing system according to Aspect 12 or Aspect 13, wherein the classifier is a cascade classifier.

[0084] 1: semiconductor device manufacturing system, 10: manufacturing apparatus, 100: imaging system, 100A: imaging system, 101: dataset generation system, 102: image storage unit, 103: dataset storage unit, 105: charged particle beam image, 106: ROI, 107: non-ROI, 108: processor, 109: classifier construction unit, 110: classifier storage unit, 111: observation recipe input unit, 112: observation recipe display unit, 113: input device, 114: memory, 115, observation recipe generation unit, 200: imaging device, 201: electron gun, 202: focusing lens, 203: deflection coil, 204 : Objective lens, 205: Secondary electron detector, 206: Sample stage, 207: Electron beam, 220: Sample, 221: Secondary electrons, 231: Image forming unit, 232: AD converter, 233: Calculator, 234: Computer system, 235: Display unit, 236: Control unit, 400: Platform, 401: Database, 402: OS, 403: Middleware, 405: Observation recipe creation application, 406: Data set generation application, 410: Terminal, 501: Charged particle beam image, 502: ROI, 503: Non-ROI, 504: Trench, 601: Non-ROI Schematic diagram of image recognition when the number of non-ROIs is selected, 602: detection region, 603: schematic diagram of image recognition when the number of non-ROIs is not selected, 700: charged particle beam image, 701x: first candidate region, 702x: second candidate region, 703x: third candidate region, 701y: first candidate region, 702y: second candidate region, 703y: third candidate region, 701d: first candidate region, 702d: second candidate region, 703d: third candidate region, 800: charged particle beam image, 801: ROI, 802: non-ROI candidate region, 803: overlap region, 900: charged particle beam image , 901: ROI, 902: Non-ROI candidate region, 1000: Charged particle beam image, 1001: First non-ROI candidate region, 1002: Second non-ROI candidate region, 1003: First non-ROI candidate region after movement, 1004: Second non-ROI candidate region after movement, 1100: ROI data registration window, 1101: Input image name display region, 1102: Upload button, 1103: Charged particle beam image display region, 1104: Guide region, 1105: Mouse cursor, 1106: Registration region, 1107: Magnification link button, 1108: Square button,1109: Learning data tag area, 1110: Area size adjustment function display, 1111: Magnification display area, 1200: Non-ROI data generation parameter adjustment window, 1201: File name display section, 1202: Base non-ROI display window, 1203: Non-ROI area, 1204: New parameter display area, 1205: Reference parameter display area, 1206: New parameter non-ROI number display area, 1207: Parameter adjustment area, 1208: Save button, 1209: Auto button, 1210: Target non-ROI number display area, 2000: Imaging device, 2001: Microscope device,

Claims

1. A dataset generation device that generates a dataset used to train a classifier that performs image recognition of charged particle beam images, characterized in that: a Region of Interest (ROI) registered for the image recognition is generated based on the ROI and the charged particle beam image, and non-ROI candidates, which are areas that do not include the ROI, are generated based on the ROI, while reducing the number of non-ROI candidates.

2. A data set generating device according to claim 1, characterized in that the non-ROIs are selected from the non-ROI candidates based on the proportion of overlapping area between each of the ROIs and each of the non-ROI candidates.

3. A data set generating device according to claim 1, characterized in that the non-ROIs are selected from the non-ROI candidates based on a first correlation coefficient between each of the ROIs and each of the non-ROI candidates.

4. A data set generating device according to claim 1, wherein two non-ROI candidates are selected from the non-ROI candidates, and the non-ROI is selected from the non-ROI candidates based on a second correlation coefficient between one of the two selected non-ROI candidates and the other of the two selected non-ROI candidates.

5. A data set generating device according to any one of claims 2 to 4, characterized in that the classifier is a cascade classifier.

6. A dataset generation method for generating a dataset used to train a classifier that performs image recognition of charged particle beam images, the dataset generation method comprising: a step of generating non-ROI candidates, which are regions that do not include a Region of Interest (ROI) registered for the image recognition and the charged particle beam image, based on the ROI; and a step of generating non-ROIs while reducing the number of non-ROI candidates based on the ROI.

7. An imaging device for imaging a sample using a charged particle beam, comprising: a classifier for performing image recognition of the image of the charged particle beam; and based on a region of interest (ROI) registered for the image recognition and the image of the charged particle beam, non-ROI candidates, which are regions that do not include the ROI, are generated, and the number of non-ROI candidates is reduced based on the ROI to generate non-ROIs, thereby generating a dataset to be used for training the classifier.

8. An imaging method for imaging a sample using a charged particle beam, comprising the steps of: performing image recognition of the image of the charged particle beam; and generating a dataset to be used for learning to perform the image recognition, wherein the dataset is generated by generating non-ROI candidates, which are regions that do not include a Region of Interest (ROI) registered for the image recognition and the image of the charged particle beam, and generating non-ROIs while reducing the number of non-ROI candidates based on the ROI.

9. A microscope apparatus equipped with the imaging device according to claim 7, characterized in that the field of view is shifted by detecting the ROI using an image of the charged particle beam.

10. A dataset generation system for generating a dataset used to train a classifier that performs image recognition of charged particle beam images, characterized in that the following steps are executed by a processor: generating a region of interest (ROI) registered for the image recognition and non-ROI candidates that are regions that do not include the ROI based on the charged particle beam image; and generating non-ROIs based on the ROI while reducing the number of non-ROI candidates.

11. An imaging system for imaging a sample using a charged particle beam, wherein a step of performing image recognition of the image of the charged particle beam and a step of generating a dataset to be used for learning to perform the image recognition are executed by a processor, and the dataset is generated by generating non-ROI candidates, which are regions that do not include a Region of Interest (ROI) registered for the image recognition and the image of the charged particle beam, and generating non-ROIs while reducing the number of non-ROI candidates based on the ROI.

12. A semiconductor device manufacturing system having a platform on which an application is implemented for generating a dataset used to train a classifier that performs image recognition of charged particle beam images, wherein the application executes the following steps: generating Regions of Interest (ROIs) registered for the image recognition and non-ROI candidates, which are regions that do not include the ROIs, based on the charged particle beam images; and generating non-ROIs based on the ROIs while reducing the number of non-ROI candidates.

13. A semiconductor device manufacturing system according to claim 12, wherein the platform is a server.

14. A semiconductor device manufacturing system according to claim 12, wherein the classifier is a cascade classifier.

Images