Defect inspection system and defect inspection method

Abstract

To provide a defect inspection system and a defect inspection processing method that are capable of obtaining an inspection result with good reproducibility.SOLUTION: A defect inspection system 1 having an imaging device 100 for acquiring an observation image of a sample 107 includes: a learning unit 330a that learns a model 401 for estimating multiple non-defective product images, the model estimating a plurality of non-defective product images of the sample 107 for one input image using a learning image acquired by imaging the sample 107; and a defect inspection unit 330b that inputs an inspection target image 1000 into the trained model 401 for estimating multiple non-defective product images, outputs a plurality of estimated non-defective product images 440 corresponding to the inspection target image 1000, and extracts a defective part using the plurality of estimated non-defective product images 440, the inspection target image 1000 being acquired by imaging an inspection target sample by the imaging device 100.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a defect inspection system and a defect inspection method. [Background technology]

[0002] Patent Document 1 is a background technique of this technical field. Patent Document 1 describes that "a selection of either a learning mode or an operation mode is accepted, and when the learning mode is selected, a learning image including an inspection object is acquired, a setting of a hyperparameter of a first learning model generated by unsupervised learning is accepted, and the first learning model is generated based on the accepted hyperparameter, which outputs a reconstructed image corresponding to the learning image when the acquired learning image is input, an inspection image including an inspection object is acquired, a setting of a threshold of an anomaly degree for verifying the inspection image is accepted, an anomaly degree between the acquired inspection image and a reconstructed image output by the trained first learning model is calculated, and a detected abnormal image is displayed based on the calculated anomaly degree and the accepted threshold of anomaly degree." [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2021-140739 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, in Patent Document 1, a model that outputs a reconstructed image corresponding to an input image is generated by unsupervised learning, and image inspection is performed using the inspection image and the reconstructed image. The deep learning model including a VAE (Variational Auto Encoder) described in the example of Patent Document 1 includes random processing in the process of learning internal parameters. Therefore, even if the same image data set is used for learning, the internal parameters obtained after learning are different for each model. As a result, when an inspection is performed using a trained model, there is a risk that the inspection results will vary depending on the learning results for each model.

[0005] Therefore, the present invention provides a defect inspection system and a defect inspection processing method that are capable of obtaining inspection results with good reproducibility. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, the defect inspection system of the present invention is a defect inspection system having an imaging device that acquires an observation image of a sample, and is characterized in comprising: a learning unit that images the sample to acquire a learning image, and learns a multiple-good image estimation model that estimates multiple good images of the sample for a single input image using the learning image; and a defect inspection unit that images a sample to be inspected using the imaging device to acquire an inspection image, inputs the inspection image into the learned multiple-good image estimation model, outputs multiple estimated good images corresponding to the inspection image, and extracts defective areas using the multiple estimated good images.

[0007] In addition, a defect inspection method according to the present invention is a defect inspection method for a defect inspection system having an imaging device that acquires an observation image of a sample, and is characterized by comprising the steps of: a learning unit imaging the sample to acquire a learning image, and learning a multiple-good image estimation model that uses the learning image to estimate multiple good images of the sample for a single input image; and a defect inspection unit imaging a sample to be inspected using the imaging device to acquire an inspection image, inputting the inspection image into the learned multiple-good image estimation model, outputting multiple estimated good images corresponding to the inspection images, and extracting defective areas using the multiple estimated good images. Effect of the Invention

[0008] According to the present invention, it is possible to provide a defect inspection system and a defect inspection processing method that are capable of obtaining inspection results with good reproducibility. Problems, configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0009] [Figure 1] 1 is a schematic configuration diagram of a defect inspection system according to a first embodiment of the present invention. [Diagram 2] 2 is a flowchart showing a process flow of the defect inspection system shown in FIG. [Diagram 3] 2 is an example showing a structure of a non-defective product image estimation model constituting the defect inspection system according to the first embodiment. [Figure 4] 4 is an example showing a structure of an estimation model for a plurality of non-defective product images constituting the defect inspection system according to the first embodiment. [Diagram 5] 4 is a flowchart showing a learning flow of an estimation model for a plurality of non-defective product images by a learning section shown in FIG. 1. [Figure 6] 2 is a flowchart showing a learning flow of a non-defective product image estimation model by a learning unit shown in FIG. 1. [Figure 7] 2 is an explanatory diagram of a data augmentation process performed by a learning unit shown in FIG. 1. [Figure 8]2 is an explanatory diagram of a noise addition process performed by a learning unit shown in FIG. 1. [Figure 9] 2 is a flowchart showing an inspection process flow by a defect inspection unit shown in FIG. 1. [Figure 10] 2 is an explanatory diagram of integration of inspection results for defect images by the defect inspection unit shown in FIG. 1. [Figure 11] 2 is an explanatory diagram of a false alarm discrimination process performed by the defect inspection unit shown in FIG. 1; [Figure 12A] FIG. 2 is a diagram showing an example of a GUI of the display device shown in FIG. 1, showing the first half of the display screen. [Figure 12B] 2 is a diagram showing an example of a GUI of the display device shown in FIG. 1, showing the latter half of the display screen. FIG. [Figure 13] 13 is an example showing a structure of a multi-head model constituting a defect inspection system according to a second embodiment of the present invention. [Figure 14] 4 is a flowchart showing a flow of learning a multi-head model by a learning unit shown in FIG. 1. [Figure 15] 4 is a flowchart illustrating a learning process of a multi-head model performed by a learning unit shown in FIG. 1. [Figure 16] 10 is a flowchart showing an inspection flow using a multi-head model constituting a defect inspection system according to a third embodiment of the present invention. [Figure 17] 3A and 3B are diagrams illustrating an example of inspection of defect images using integration processing of estimated non-defective images by the defect inspection unit illustrated in FIG. 1. [Figure 18] FIG. 11 is a diagram showing an example of learning of a non-defective image estimation model using maximization of a non-defective conversion amount by a learning unit constituting the defect inspection system according to Example 4 of the present invention. [Figure 19] FIG. 11 is a diagram showing an example of an inspection process using selection of an estimated non-defective image by a defect inspection unit constituting a defect inspection system according to a fifth embodiment of the present invention. [Figure 20] FIG. 2 is a diagram showing an example of a GUI of the display device shown in FIG. 1, showing a display screen of the GUI for selecting an estimated non-defective image. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] In this specification, the imaging device includes a charged particle beam device, an optical inspection device, and an imaging device using a camera, and the charged particle beam device includes an SEM, a review SEM, an FIB, etc. In the following, an SEM will be used as an example of a charged particle beam device. Hereinafter, the embodiments of the present invention will be described with reference to the drawings. Note that the embodiments described below do not limit the invention according to the claims, and all of the elements and combinations thereof described in the embodiments are not necessarily essential to the solution of the invention. EXAMPLES

[0011] Fig. 1 is a schematic diagram of a defect inspection system according to a first embodiment of the present invention. As shown in Fig. 1, the defect inspection system 1 according to this embodiment includes a charged particle beam device 2 and a computer 3. In this embodiment, an example will be described in which the defect inspection system 1 estimates a plurality of corresponding non-defective product images from an inspection image of a sample captured by the charged particle beam device 2, and extracts defective portions through inspection processing using statistical processing.

[0012] <Configuration of defect inspection system> As shown in FIG. 1, the defect inspection system 1 includes a charged particle beam device 2 and a computer 3. The charged particle beam device 1 includes an imaging device 100 and a controller 200. The imaging device 100 includes components such as a stage serving as a sample table and a driving circuit in a housing, in other words, a column or a sample chamber. The controller 200 is a control system that drives and controls the imaging device 100, and can be implemented by a computer or a circuit. The computer 3 is, in other words, a computer system. The defect inspection system 1 includes components necessary for generating a signal waveform or an image based on a detection signal of the imaging device 100 constituting the charged particle beam device 2. An example of the charged particle beam device 2 is a scanning electron microscope (SEM).

[0013] The imaging device 100 of the charged particle beam device 2 outputs a detection signal s1 obtained based on the irradiation of the charged particle beam 103 to a sample 107 on a stage 108. The controller 200 inputs or receives the detection signal s1, processes it, and generates and stores a signal such as an image as a detection signal of the charged particle beam device 2. The controller 200 outputs a signal s2 of the image or the like. The computer (computer system) 3 inputs or receives the signal s2 of the image or the like, and processes it.

[0014] In the imaging device 100 of the charged particle beam device 2, a charged particle beam 103 extracted from an electron source 101 by an extraction electrode 102 is accelerated by an acceleration electrode (not shown). The accelerated charged particle beam 103 is focused by a condenser lens 104, which is a form of a focusing lens. The charged particle beam 103 is deflected by a scanning deflector 105, and one-dimensionally or two-dimensionally scanned on a surface of a sample 107 on a stage 108, which is a sample stage. The charged particle beam 103 is decelerated by a negative voltage (retarding voltage) applied to an electrode built in the stage 108, which is a sample stage, and is focused by the lens action of an objective lens 106, and is irradiated onto the sample 107. The stage 108 may be a mechanism capable of moving in a Z direction perpendicular to the X and Y directions corresponding to the vertical direction, or may be a mechanism capable of rotating and tilting in each axial direction. In addition, although details of the implementation of the charged particle beam device 2 are not limited, a configuration for irradiating multiple beams may be used. Further, although the charged particle beam device 2 is illustrated as having one detector 110, the present invention is not limited thereto and may have a configuration having multiple detectors. For example, the device may have an SE detector for detecting SE (Secondary Electron) and a BSE detector for detecting BSE (Backscattered Electron). For example, the device may have multiple detectors installed at multiple positions. That is, the device may have multiple channels and multiple detection systems as a configuration for capturing an image. When generating an image, multiple images may be processed, such as by accumulating, based on multiple signals repeatedly detected from the same region to generate one image.

[0015] The sample 107 is, for example, a semiconductor wafer. When the charged particle beam 103 is irradiated onto the sample 107, electrons 109 such as secondary electrons (SE) and backscattered electrons (BSE) are emitted from inside the sample 107. The emitted electrons 109 are accelerated by an acceleration action based on a negative voltage (retarding voltage) applied to the sample 107, and are captured by a detector 110. A detection signal s1 output from the detector 110 is sent to a controller 200. The controller 200 receives the detection signal s1 via a communication device 201.

[0016] The controller 200 includes a communication device 201, a processor 202, a memory 203 such as a RAM or a non-volatile storage device, a display device 204, and an input / output device 205. These components are connected to each other by an architecture such as a bus. The controller 200 controls imaging by the imaging device 100 according to a set imaging recipe. The processor 202 generates an observation image (inspection image) as an image whose brightness is determined by the amount of electrons captured by the detector 110 based on a detection signal s1 obtained via the communication device 201. The processor 202 stores data such as the generated observation image (inspection image) in the memory 203. The processor 202 transmits a signal s2 corresponding to data such as the generated observation image (inspection image) to the computer 3 via the communication device 201. The computer 3 receives the signal s2 via the communication device 310. Note that the user U1 may use the controller 200 by operating the display device 204 and the input / output device 205.

[0017] The communication device 201 is a device that is equipped with a communication interface with the imaging device 100 and a communication interface with the computer 3. The communication interface may be, for example, a LAN, but is not limited to this. The input / output device 205 is an input device and an output device. The input device and the output device may be built-in or may be externally connected. Examples of the input device are a keyboard, a mouse, a microphone, etc. Examples of the output device are a display, a printer, a speaker, etc.

[0018] The computer 3 includes a communication device 310, a storage 320, a learning unit 330a, a defect inspection unit 330b, a memory 331 such as a RAM or a non-volatile storage device, a display device 332, and an input / output device 333. These components are connected to each other by an architecture such as a bus. The storage 320 is a memory having a relatively large storage area. A learning image DB 321, an inspection image DB 322, a learned parameter DB 323, and a program 324 are stored in the storage 320. In addition, the memory 331 of the computer 3 stores necessary management information and a database (DB). An external storage device, a server, etc. may be connected to the computer 3, and necessary data and information may be stored in the external storage device, the server, etc., and the data and information may be read and written as appropriate. The learning unit 330a and the defect inspection unit 330b are realized by a processor such as a CPU (not shown), a ROM for storing various programs, a RAM for temporarily storing data in the calculation process, a storage device such as an external storage device, and the processor such as a CPU reads out and executes various programs stored in the ROM, and stores the calculation results, which are the execution results, in the RAM, the external storage device, or in cloud storage via a network connection. Alternatively, the various programs may be stored in a program 324 in the storage 320. The memory 331 may be configured to appropriately store various data and information processed by the learning unit 330a and the defect inspection unit 330b. The memory 331 stores, for example, program execution modules, image data, and processing result information. The storage 320 may store processing result information, history information, screen data, and the like in addition to the DB shown in the figure.

[0019] The communication device 310 is a device that is equipped with a communication interface with the controller 200. An external device may be connected to the computer (computer system) 3 via a network such as a LAN. The input / output device 333 is an input device or an output device. The input device or the output device may be built-in or may be externally connected. Examples of the input device are a keyboard, a mouse, a microphone, etc. Examples of the output device are a display, a printer, a speaker, etc.

[0020] A user U1 of the defect inspection system 1 operates the display device 332, the input / output device 333, etc. to use the computer (computer system) 3. In this way, the user U1 uses the defect inspection system 1. A screen that serves as a user interface of the defect inspection system 1, in other words, a screen with a graphical user interface (GUI), is displayed on the display screen of the display device 332. The computer 3 may be configured as a client-server system. In this case, the user U1 operates a PC or the like that serves as a client terminal, and the PC or the like accesses the computer (computer system) 3 that serves as a server through communication.

[0021] The computer (computer system) 3 is a part that executes the processing characteristic of this embodiment. The computer (computer system) 3 is connected to the controller 200. The controller 200 may be configured as a part of the computer (computer system) 3, and the computer 3 may be configured as a part of the controller 200. The controller 200 and the computer (computer system) 3 may be configured as an integrated computer system. The controller 200 may execute at least a part of the processing characteristic of this embodiment. The entire system including the computer (computer system) 3 may be the charged particle beam device 2.

[0022] In addition, when the computer 3 (computer system) is in the form of a client-server system, the operation is as follows, for example. A user U1 accesses the computer (computer system) 3, which is a server, from a client terminal such as a PC via a network. The server provides a screen with a graphical user interface (GUI) to the client terminal. The server transmits GUI screen data (for example, a Web page) for that purpose to the client terminal. The client terminal displays a GUI screen on a display based on the received screen data. The user U1 looks at the GUI screen and inputs instructions and settings. The client terminal transmits the input information to the server. The server executes processing according to the received input information. For example, the server evaluates the captured image and estimates the device status, stores the processing results, and transmits GUI screen data (or only update information) for displaying the processing results to the client terminal. The client terminal updates the display of the GUI screen based on the received screen data. The user U1 can check the processing results, for example, the estimation results of the device status, by looking at the GUI screen.

[0023] <Processing flow of defect inspection system> Fig. 2 is a flowchart showing a process flow of the defect inspection system 1 shown in Fig. 1. As shown in Fig. 2, in step S1, an observation region of a sample 107 is scanned with an electron beam (charged particle beam) 103, and a detection signal s1 captured by a detector 110 is sent to a controller 200 via a communication device 201. An observation image generated by a processor 202 based on the sent detection signal s1 is sent to a computer (computer system) 3 via a communication device 310. A learning unit 330a stores the sent observation image in a learning image DB 321.

[0024] In step S2, the learning unit 330a determines internal parameters of a multiple-good image estimation model that generates multiple good image corresponding to the input image by learning using the training image DB 321, and stores the determined internal parameters in the learned parameter DB 323. In step S3, the electron beam (charged particle beam) 103 is scanned over an area on the sample 107 to be inspected for defects, and a detection signal s1 captured by the detector 110 is sent to the controller 200 via the communication device 201. An observation image generated by the processor 202 based on the sent detection signal s1 is sent to the computer 300 via the communication device 310. The defect inspection unit 330b stores a signal s2 corresponding to data such as the sent observation image (inspection image) in the inspection image DB 322.

[0025] In step S4, the defect inspection unit 330b inputs images to be inspected among the images stored in the inspection image DB 322 as input images to an estimation model 401 (described later in FIG. 4) of multiple good images whose internal parameters have been read from the learned parameter DB 323, and outputs multiple good images corresponding to the input images. The inspection result is output using the input images and the multiple output good images, and the process ends.

[0026] <Example of inspection using estimation models from multiple good product images> 3 to 12A and 12B, a specific example of the defect inspection process will be described assuming a case where defects on a semiconductor wafer, which is sample 107, are inspected by defect inspection system 1 according to this embodiment.

[0027] FIG. 3 is an example showing the structure of a good product image estimation model constituting the defect inspection system 1 according to the present embodiment. As shown in FIG. 3, the good product image estimation model 400 is configured based on a machine learning model, such as an autoencoder, U-Net, or GAN (generative adversarial networks). When the amount of defect images used for learning the model is insufficient, such as in the early stages of operation of the defect inspection system 1, these models are trained to estimate a good product image corresponding to an input image 430 using only good product images as learning data, thereby enabling defect inspection by comparing the input image 430 with the estimated good product image (estimated good product image 440). As an example of a learning method, the internal parameters are repeatedly updated so that the error between the input image 430 and the output image of the model is minimized, and the model learns to reconstruct a good product image from a subspace constructed by the feature representation.

[0028] In this embodiment, a good-quality image estimation model 400 based on U-net is used as an example. U-net is composed of a feature extraction unit 410 in the first half and an image generation unit 420 in the second half. The feature extraction unit 410 has a structure in which processing is repeated using a convolution layer that extracts image features and a pooling layer that reduces the size of the output vector of the convolution layer, and local features and global features of the input image 430 are obtained. In the image generation unit 420 in the second half, the feature obtained in the feature extraction unit 410 in the first half is repeatedly processed using a convolution layer and an upsampling layer that increases the size of the output vector of the convolution layer, thereby obtaining an estimated good-quality image 440 of the same size as the input image 430. In addition, a skip connection is used to connect the output of the feature extraction unit 410 to the feature obtained by upsampling, and a clearer output image can be obtained. This good product image estimation model 400 is trained using good product images as training data, thereby learning internal parameters that enable the characteristic expression of good product images and the generation of good product images from the extracted feature amounts.

[0029] 4 is an example showing the structure of a multiple-good-image estimation model 401 constituting the defect inspection system 1 according to this embodiment. The multiple-good-image estimation model 401 is composed of multiple good-image estimation models 400, and by inputting one input image 430, it outputs the same number of estimated good-images 440 as the number of the configured good-image estimation models 400. Although the respective good-image estimation models 400 have the same structure, they output different estimated good images 440 for the same input image because their internal parameters determined by a learning process including randomness are different. It is sufficient that the multiple-good-image estimation model 401 includes at least two good-image estimation models 400.

[0030] Fig. 5 is a flowchart showing a learning flow of the estimation model 401 of multiple good images by the learning unit 330a shown in Fig. 1. An example will be described in which the estimation model 401 of multiple good images is learned, the estimation model 401 including n good image estimation models 400. First, in step S201, the learning unit 330a reads images used for learning the estimation model 401 of multiple good images from the training image DB 321, and creates a learning dataset.

[0031] Thereafter, in step S202, the learning unit 330a trains each of the multiple (i=n, 1, i+1) good image estimation models 400 in order to generate the estimation models 401 for multiple good image. Before starting the learning of each good image estimation model 400, a random initial seed is set for each good image estimation model 400 so that there is no overlap (step S203). As an example of a method for determining the initial seed, it can be determined based on the call time of a function that sets the initial seed of a program. Next, in step S204, the learning unit 330a sets the initial values ​​of the internal parameters of the good image estimation model 400(i). Since the initial values ​​are determined by referring to a random number table, different values ​​are set when the initial seeds are different (step S205).

[0032] Next, in step S206, the learning unit 330a performs a learning process to determine internal parameters of the good product image estimation model 400(i) with the initial values ​​set, using a learning data set, to create a trained good product image estimation model 400(i) (step S207). The process from setting the initial seed to the learning process is repeated at least two times (step S208), to create a trained multiple good product image estimation model 401 (step S209). After creating an arbitrary number of trained multiple good product image estimation models 401, all internal parameters are set as internal parameters of one multiple good product image estimation model 401, and are stored in the trained parameter DB 323.

[0033] FIG. 6 is a flowchart showing a learning flow of the good image estimation model by the learning unit shown in FIG. 1. As shown in FIG. 6, in step S301, learning loop 1 is started (i=n, 1, i+1). First, in step S302, the learning unit 330a executes learning mini-batch division of the learning data set. In the learning mini-batch division, the images included in the learning data set are divided into an arbitrary number of images, and each of the images is set as a learning mini-batch. For each divided learning mini-batch, data augmentation (step S305), noise addition (step S306), learning loss calculation (step S307), and internal parameter update (step S308) are executed. In the data augmentation process (step S305), the learning unit 330a processes each of the images included in the learning mini-batch with luminance conversion, contrast conversion, and distortion addition at random intensities within a certain range based on reference values ​​in a random number table, thereby creating an input image for the good image estimation model 400. In order to train the good image estimation model 400 to perform estimation processing from a non-good image to a good image, noise is added to the input image as a non-good feature to create a non-good image. The created non-good image is input to the good image estimation model 400 to obtain an estimated good image 440 as an output. At this time, the average of squared errors of the luminance of each pixel between the estimated good image and the input image 430, which is the image before noise is added, is calculated as the learning loss. The learning loss may include SSIM (Structural Similarity), which is an index value related to image quality, or a regularization term that imposes a penalty on the internal parameters to prevent excessive fitting to the learning data set. Based on the calculated learning loss, an optimization method such as gradient descent is used to search for values ​​of the internal parameters that reduce the learning loss, and the internal parameters of the good image estimation model 400 are updated based on the search result. The process from data augmentation (step S305) to internal parameter update (step S308) is repeated the same number of times as the number of training mini-batches, and then the training data set is divided into training mini-batches again, and the internal parameters are updated for each training mini-batch. By repeating this training loop an arbitrary number of times, the internal parameters of the non-defective image estimation model are determined.

[0034] FIG. 7 is an explanatory diagram of the data augmentation process by the learning unit 330a shown in FIG. 1. As shown in FIG. 7, in the data augmentation process 700, the learning unit 330a processes contrast conversion 701, brightness conversion 702, and distortion addition 703 for each of the learning images included in the learning mini-batch. In this way, the input image 430 is created. In the contrast conversion 701, the contrast of the entire image to be processed is changed randomly within a set range. For example, the contrast can be changed randomly by updating the contrast with a value obtained by multiplying the brightness value of each pixel by a random value selected from the range of 0.8 to 1.2. If the contrast is less than 1.0, the contrast is lowered, and if the contrast is greater than 1.0, the contrast is increased. In the brightness conversion process 702, a randomly determined brightness value is added to each pixel of the image, or a random value is determined for each pixel and added to the brightness value. The value to be added is a value randomly selected from a set range. The range is, for example, ±10% of the average pixel value of the entire image to be processed. In the distortion addition 703, the image to be processed is randomly distorted using a transformation process. The image can be distorted, for example, by applying an affine transformation to the vertical and horizontal directions of the image. By learning images in which the luminance change, contrast, and image distortion have changed within a certain range as good products, it is possible to improve the robustness of the inspection process against distortion caused by the imaging device 100 and changes in image quality such as luminance and contrast caused by changes in the state of the imaging device 100 during operation.

[0035] FIG. 8 is an explanatory diagram of the noise addition process 800 by the learning unit 330a shown in FIG. 1. In the noise addition process 800, the learning unit 330a adds noise as an image feature that is not a good product to the input image 430 created by the data extension process. The noise addition process 800 is performed probabilistically using random numbers. For example, a random value is obtained from the range of 0 to 1 as a random number value, and the noise addition process (802) is performed only if the random number value is smaller than a set value (801). If the set value is 0.8, the noise addition process 802 is performed with a probability of 80%. By performing the noise addition process 800 probabilistically, the learning mini-batch includes images to which no noise has been added and images to which noise has been added, and learning is possible to leave good images as good images and convert non-good images to good images. Methods for generating noise to be added as image features that are not good products include Perlin noise and wavelet noise, for example. Either noise generation method can generate an image with a random pattern, and the generated image is multiplied by a random coefficient within a certain range and the resultant value is added to the good image to create a noise image 803 having non-good characteristics. In addition to the above-mentioned noise generation method, a process such as pasting an image obtained by cutting out a part of an image or another image in the learning data set onto the input image 430 or drawing a line with a random position, angle, length, thickness, and brightness may be performed as a non-good characteristic.

[0036] FIG. 9 is a flow chart showing an inspection process flow by the defect inspection unit 330b shown in FIG. 1. As shown in FIG. 9, first, in step S401, the inspection area of ​​the sample 107 to be inspected is scanned with the charged particle beam 103, and the acquired inspection image is stored in the inspection image DB 322. Next, in step S402, the defect inspection unit 330b reads out an image to be inspected from the inspection image DB 322. After that, in step S403, the defect inspection unit 330b reads out the internal parameters of the estimation model 401 of the multiple good images from the learned parameter DB 323. In step S404, the defect inspection unit 330b inputs the inspection image 1000 (FIG. 10) to the estimation model 401 of the multiple good images into which the internal parameters have been read, and outputs multiple estimated good images 440 (FIG. 10). When the inspection image 1000 including a defect is input to the estimation model 401 of the multiple good images, an image in which the defective portion is converted to look like a good product is output. Therefore, in the comparison process (step S405), the defect inspection unit 330b compares the inspection image 1000 with each of the estimated good images 440 to extract defective parts, and creates a difference image in which each pixel indicates the difference in pixel value between the inspection image 1000 and the estimated good image 440. At this time, since multiple estimated good images 440 are output for one inspection image 1000, the same number of difference images 1002 (FIG. 10) as the estimated good images 440 are created. In the false alarm discrimination process (step S406), the defect inspection unit 330b removes false alarms, which are parts where a difference has occurred but are not actually defective, based on the image features of the difference image 1002, and outputs a defect map 1004 (FIG. 10) showing the defective parts. The number of defect maps 1004 output is the same as the estimated good images 440. In the integration process (step S407), the defect inspection section 330b performs statistical processing on the multiple defect maps 1004, and outputs one integrated defect map 1006 (FIG. 10) as an inspection result (step S408).

[0037] FIG. 10 is an explanatory diagram of integration of inspection results for defect images by the defect inspection unit 330b shown in FIG. 1. As shown in FIG. 10, the estimation model 401 of multiple good images includes three good image estimation models (not shown), and an example will be described in which three estimated good images 440 are output for one inspection image 1000. The inspection image 1000 is an example of an image of a semiconductor circuit pattern captured using the charged particle beam device 2. The defect inspection unit 330b obtains three estimated good images 440 by inputting the inspection image 1000 to the estimation model 401 of multiple good images that has read internal parameters from the learned parameter DB 323. Since the three good image estimation models 400 (not shown) have different internal parameters, they output different estimated good images 440 even when the same inspection image 1000 is input. Next, a difference image 1002 is created by comparing each of the three estimated good images 440 with the inspection image 1000. The area where a difference occurs in the comparison process 1001 is an area where the inspection image 1000 is converted into a different shape, and therefore is likely to be a defect. Next, a false alarm discrimination process 1003 is performed on the difference image 1002 to exclude areas where a difference occurs but are not actually defective, and a defect map 1004 showing the defective portion is created. Next, an integrated defect map 1006 showing the median of pixel values ​​at the same position of the three defect maps 1004 is created as a statistical process for the three defect maps 1004. By using the median of the multiple defect maps 1004 (median process 1005), it is possible to exclude false alarms that occur when the multiple non-defective image estimation model 401 includes a non-defective image estimation model 400 with a large estimation error of a specific non-defective pattern due to the influence of randomness in learning. In addition, by using the defective portions extracted from the multiple defect maps 1004 as the final inspection result, it is possible to suppress inspection variations such as extraction being possible in one model but not in another model, and improve the reproducibility of the inspection result. The statistical procedure for integrating the defect map 1004 may be the average, mode, or maximum.

[0038] FIG. 11 is an explanatory diagram of the false alarm discrimination process by the defect inspection unit 330b shown in FIG. 1. The target of the false alarm discrimination process is a difference image 1002 (FIG. 10) created by a comparison process 1001 (FIG. 10) between an inspection image 1000 (FIG. 10) and an estimated non-defective image 440 (FIG. 10). As shown in FIG. 11, first, the defect inspection unit 330b performs a grouping process 1101 of defect candidates on the difference image 1002. Since the pixel value of a defective portion is locally high in the difference image 1002, the grouping process 1101 creates a defect candidate map 1102 in which an area with a pixel value of a certain value or more is set to 1 and the other areas are set to 0, and a labeling process is used to assign the same identification number to connected pixels to treat the connected area as one defect candidate and give the corresponding defect candidate ID. Next, the defect inspection unit 330b calculates 1103 the feature amount of each defect candidate using the difference image 1002 and the defect candidate map 1102. The feature amount is, for example, the coordinates of the center of gravity, the area on the defect candidate map 1102, and the maximum, average, and variance of the difference value of the brightness. The feature amount of the defect candidate may be, for example, a minimum value, circularity, edge strength, or the like. Next, the defect inspection unit 330b judges whether each defect candidate is a false alarm or a real alarm by the false alarm judgment process 1105. The conditions for judging as a false alarm are, for example, the area is equal to or less than a threshold value, the maximum value of the difference value of the brightness is equal to or less than a threshold value, the average value of the difference value of the brightness is equal to or less than a certain value, and the variance of the difference value of the brightness is equal to or less than a certain value. Here, the variance and maximum value of the brightness are values ​​for each pixel, and the average is the average value of the brightness of all pixels in the region. Also, a combination of a plurality of conditions may be used. An image in which the pixel value of the defect candidate on the defect candidate map corresponding to the defect candidate ID judged as a false alarm is changed to 0 is output as the defect map 1004.

[0039] 12A and 12B are diagrams showing an example of the GUI of the display device shown in FIG. 1, a diagram showing the first half of the display screen, and a diagram showing an example of the GUI of the display device shown in FIG. 1, a diagram showing the second half of the display screen, respectively. As shown in FIG. 12A, the GUI according to this embodiment has an input field 1210 for learning parameters of the estimation model 401 of multiple good images and a display field 1220 for displaying the learning status. The input field 1201 for learning parameters includes input fields for parameters related to data augmentation processing 1212, noise addition 1213, the number of times of learning 1214, the number of image generation copies 1215, and learning images 1216 as learning parameters 1211. In the input field for data augmentation processing 1212, a check box is used to select an augmentation processing to be executed during learning. In the input field for noise addition 1213, a value between 0.0 and 1.0 is input as the probability of executing noise addition during learning. In the input field for the number of image generation copies 1215, input is input as to how many estimated good images the estimation model 401 of multiple good images outputs for one input image. Depending on the number of image generation units 1215, it is determined how many good image estimation models 400 are to be used to configure the multiple good image estimation model 401. In the input field for training image 1216, the file path of the image to be used for training the multiple good image estimation model 401 is input. During training, the image corresponding to the specified file path is used. After inputting each parameter, the training of the multiple good image estimation model 401 is executed by pressing the training start button 1217 (clicking with a mouse or the like). The training status display field 1220 has a display field where the training loss 1222 display field of the multiple good image estimation model 401 and the estimated good image 1223 display field can be confirmed for each image generation unit 1221. The training loss 1222 display field displays the training loss of the internal parameters of the multiple good image estimation model 401 at each training count. The display field for the estimated good image 1223 has an input field for the number of training counts. After inputting an arbitrary number of learning times in the input field for the number of learning times, pressing the display button (clicking with a mouse or the like) displays the learning images and the corresponding estimated good product images 440 for the number of learning times of the internal parameters of the estimation model 401 for the corresponding multiple good product images.

[0040] As shown in FIG. 12B, the GUI according to this embodiment has an input field 1230 for the inspection parameters of the estimation model 401 of multiple good images and a display field 1240 for the inspection results. The input field 1230 for the inspection parameters includes an input field 1232 for the estimated good product integration, an input field 1233 for the learned parameters of the estimation model 401 of multiple good images, and an input field 1234 for the image to be inspected. In the input field 1232 for the estimated good product integration, a check box is used to select a statistical processing method to be used in the integration process. In the input field 1233 for the learned parameters, a path to a file in which the internal parameters of the estimation model 401 of multiple good images are saved is input. In the input field 1234 for the image to be inspected, a file path to the image to be inspected is input. The display field 1240 for the inspection results has an image display field 1241 and a defect distribution display field 1242. In the image display field 1241, the image to be inspected, a defect map, and each estimated good product image are displayed as the inspection target. A defect distribution display field 1242 displays a wafer map 1243 showing the defect density for each region on a semiconductor wafer, which is the sample 107, and the number of defects and the defect density (color bar display) in each region. The regions are, for example, semiconductor chips manufactured on a semiconductor wafer.

[0041] As described above, according to this embodiment, it is possible to provide a defect inspection system and a defect inspection processing method capable of obtaining inspection results with good reproducibility. EXAMPLES

[0042] A second embodiment will be described with reference to Figs. 13 to 15. The basic configuration of the second embodiment is similar to and common to that of the first embodiment, and the following mainly describes components of the second embodiment that are different from those of the first embodiment. In this embodiment, the estimation model 401 of multiple good product images is not composed of a plurality of good product image estimation models, but is composed of a common feature extraction unit 410 and a plurality of image generation units. A specific example will be described below.

[0043] FIG. 13 is an example showing the structure of a multi-head model constituting a defect inspection system according to the second embodiment. As shown in FIG. 13, the multi-head model is an example of a non-defective product image estimation model 400a, and is composed of a feature extraction unit 410 in the first half and a plurality of image generation units in the second half. The plurality of image generation units in the second half specifically include a first image generation unit 420a and a second image generation unit 420b. The feature extraction unit 410 has a structure in which processing is repeated using a convolution layer that extracts image features and a pooling layer that reduces the size of the output vector of the convolution layer, and inputs the extracted feature amount to each of the image generation units in the second half (the first image generation unit 420a and the second image generation unit 420b). The first image generation unit 420a and the second image generation unit 420b obtain an output of the same size as the input image by repeating processing using a convolution layer and an upsampling layer that increases the size of the output vector of the convolution layer for the feature amount obtained from the feature extraction unit 410. With this model structure, when one input image 430 is input to the multi-head model, the same number of estimated good images as the number of image generating units are output. Specifically, a first estimated good image 1301 is output from the first image generating unit 420a, and a second estimated good image 1302 is output from the second image generating unit 420b. In this way, the input image multi-head model is composed of at least two or more image generating units.

[0044] FIG. 14 is a flowchart showing a learning flow of the multi-head model by the learning unit shown in FIG. 1. A case where a multi-head model having n image generating units is learned will be described as an example. As shown in FIG. 14, first, the learning unit 330a reads images used for learning the multi-head model from the learning image DB 321 and creates a learning data set (step S501). Next, the learning unit 330a randomly sets an initial seed (step S502). Next, in step S503, the learning unit 330a randomly initializes the internal parameters of the multi-head model. The learning unit 330a saves all the internal parameters of the learned multi-head model obtained by executing the learning process of the initialized multi-head model (steps S504 and S505) in the learned parameter DB 323 (step S506), and ends the process.

[0045] FIG. 15 is a flowchart for explaining the learning process of the multi-head model by the learning unit 330a shown in FIG. 1. First, the learning unit 330a divides the learning data set into learning mini-batches (step S603). For each divided learning mini-batch, data augmentation (step S606), noise addition (step S607), and learning loss calculation (step S608) are processed. When i in the learning loop 1 in the flowchart of FIG. 15 is 1, that is, when the process corresponds to the first time of the learning loop 1 (step S609), the internal parameters of the feature extraction unit 410 and the first and second image generation units 420a and 420bi are updated. When i is other than 1 (step S609), the internal parameters of the first and second image generation units 420a and 420bi are updated. After updating the internal parameters using all of one learning mini-batch in the learning loop 3, the learning data set is divided into learning mini-batches again, and the learning loop 3 is executed (steps S604 and S612). In learning loop 2, learning loop 3 is executed m times (steps S602, S613), where m may be any number equal to or greater than 1. In learning loop 1, learning loop 2 is executed the same number of times as n, which is the number of image generators (steps S601, S614).

[0046] As described above, according to this embodiment, in addition to the effects of the first embodiment, the number of parameters of the model can be reduced by having multiple image generation units for a common image feature extraction unit in the multi-head model, and the processing time and resource consumption for learning and inspection can be reduced. EXAMPLES

[0047] A third embodiment will be described with reference to Figs. 16 and 17. The basic configuration of the third embodiment is similar to and common to that of the first embodiment, and the following mainly describes components of the third embodiment that are different from the first embodiment. In the third embodiment, in the inspection flow, the integration process is not performed on a defect map, but on an estimated non-defective image to extract a defective portion. A specific example will be described below.

[0048] FIG. 16 is a flowchart showing an inspection flow using a multi-head model constituting a defect inspection system according to the third embodiment of the present invention. First, the inspection area of ​​the sample 107 to be inspected is scanned with the charged particle beam 103, and the acquired inspection image is stored in the inspection image DB 322 (FIG. 1) (step S701). Next, in step S702, the defect inspection unit 330b reads out an image to be inspected from the inspection image DB 322 (FIG. 1). After that, in step S703, the defect inspection unit 330b reads out the internal parameters of the estimation model 401 of multiple good images from the learned parameter DB 323 (FIG. 1). In step S704, the defect inspection unit 330b inputs the inspection image to the estimation model 401 of multiple good images into which the internal parameters have been read, and outputs multiple estimated good images. In step S705, the defect inspection unit 330b performs statistical processing on the multiple estimated good images 440, and creates one integrated estimated good image. In step S706, the defect inspection unit 330b creates a difference image by comparing the inspection image with the integrated estimated non-defective image in order to extract the defective parts. In step S707, the defect inspection unit 330b executes false alarm discrimination processing on the difference image to output a defect map showing the defective parts (step S708).

[0049] FIG. 17 is a diagram showing an example of defect image inspection using the integration process of estimated good images by the defect inspection unit 330b shown in FIG. 1. The estimation model 401 of multiple good images includes three good image estimation models 400 (not shown), and an example will be described in which three estimated good images 440 are output for one inspected image 1000. The defect inspection unit 330b obtains three estimated good images 440 by inputting the inspected image 1000 to the estimation model 401 of multiple good images that has read internal parameters from the learned parameter DB 323. Since the three good image estimation models 400 (not shown) have different internal parameters, three different estimated good images 440 are output. Next, as a statistical process for the three estimated good images 440, an integrated estimated good image 1702 is created that indicates the median of pixel values ​​at the same position of the three estimated good images 440 (median process 1701). Next, the inspection image 1000 is compared (1703) with the integrated estimated good image 1702 to generate a difference image 1002. A false alarm discrimination process 1704 is executed on the difference image 1002 to output a defect map 1004 indicating the defective portion. The statistical process used when integrating the estimated good images may be the average value, the mode value, or the maximum value.

[0050] As described above, according to the present embodiment, in addition to the effects of the first embodiment, the comparison process and the false alarm discrimination process are performed using the integrated image, so that the false alarm discrimination process can be performed using reproducible results, thereby enabling higher accuracy. EXAMPLES

[0051] A fourth embodiment will be described with reference to FIG. 18. The basic configuration of the fourth embodiment is similar to and common to the first embodiment, and the following mainly describes the components of the fourth embodiment that are different from the first embodiment. In the fourth embodiment, not only the reconstruction error but also the non-defective conversion amount is used in the calculation of the learning loss 1801 in the learning flow of the non-defective image estimation model 400. A specific example will be described below.

[0052] FIG. 18 is a diagram showing an example of learning of the good image estimation model 400 using maximization of the good conversion amount by the learning unit 330a constituting the defect inspection system 1 according to the fourth embodiment of the present invention. As shown in FIG. 18, the learning of the internal parameters of the estimation model 401 of the multiple good images is performed by minimizing the square error between the input image 430 and the estimated good image 440 and maximizing the square error between the noise image and the estimated good image. The reconstruction error 1803 is learned to output the estimated good image 440 of the input image 430. However, when the defect of the inspected image is minute during inspection, even if the good image can be correctly estimated, the difference value with the inspected image is small, so that the image is erroneously judged as a false alarm in the false alarm discrimination process. If the good image estimation model 400 is trained to maximize the square error between the noise image 803 and the estimated good image 440, it will output an image in which the noise image is converted as much as possible. However, if learning is performed only by maximizing the non-defective conversion amount 1802, an attempt is made to convert the noise image 803 as much as possible, and learning ends up outputting an estimated non-defective image 440 that is far removed from the input image 430. By learning by combining minimizing the reconstruction error 1803 and maximizing the non-defective conversion amount 1802, it is possible to learn to output an image in which non-defective parts remain correctly non-defective and defective parts have difference values ​​that are larger than the difference values ​​from the corresponding original non-defective patterns. This increases the difference values ​​of defective parts in the comparison process, improving the stability of extraction of defective parts and leading to improved reproducibility of the final inspection results.

[0053] As described above, according to this embodiment, in addition to the effects of the first embodiment, the difference value of the defective portion in the comparison process becomes larger, the stability of the extraction of the defective portion increases, and the reproducibility of the final inspection result can be improved. EXAMPLES

[0054] A fifth embodiment will be described with reference to Figs. 19 and 20. The basic configuration of the fifth embodiment is similar to and common to that of the first embodiment, and the following mainly describes the components of the fifth embodiment that are different from those of the first embodiment. In the fifth embodiment, in the inspection flow of the non-defective image estimation model 400, not all of the multiple estimated non-defective images 440 are used, but only some of the estimated non-defective images 440 are selectively used. A specific example will be described below.

[0055] FIG. 19 is a diagram showing an example of an inspection process using the selection of the estimated good image 440 by the defect inspection unit 330b constituting the defect inspection system 1 according to the fifth embodiment of the present invention. The estimation model 410 of multiple good images includes three good image estimation models 400 (not shown), and an example will be described in which three estimated good images 440 are output for one inspection image 1000. The defect inspection unit 330b obtains three estimated good images 440 by inputting the inspection image 1000 to the estimation model 401 of multiple good images that has read internal parameters from the learned parameter DB 323. Since the three good image estimation models 400 (not shown) have different internal parameters, three different estimated good images 440 are output. Next, the inspection image 1000, all the estimated good images 440, and parameters 1900 indicating the image generation unit to be used are input to the comparison process 1902. In the comparison process 1902, the estimated good image 440 output from the image generating unit other than the designated one is excluded from the processing target, and a difference image 1002 between the inspection image 1000 and the estimated good image 440 is created. When parameters indicating 1 and 2 are input as the image generating unit to be used, the estimated good image 440 output from the image generating unit 420 is excluded from the processing target. As a statistical process, an integrated difference image 1904 indicating the average value of pixel values ​​at the same position in the difference image 1002 is created for the difference image 1002 created by the comparison process 1902. Next, a false alarm discrimination process 1003 is executed for the integrated difference image 1904, and a defect map 1004 indicating a defective portion is output. The image generating unit to be used 1900 may be switched for each inspection area. When a specific image generating unit included in the estimation model 401 of multiple good images cannot output an appropriate estimated good image 440 for a pattern in a part of the inspection area, false alarms in the inspection can be suppressed.

[0056] Fig. 20 is a diagram showing an example of the GUI of the display device shown in Fig. 1, and is a diagram showing a display screen of the GUI for selecting an estimated good image. As shown in Fig. 20, the GUI according to this embodiment has an input field 2001 for inspection parameters. The input field 2001 for inspection parameters includes an input field 1232 for estimated good integration, an input field 1233 for learned parameters of the estimation model 401 for multiple good images, an input field 2002 for an image generation unit to be used, and an input field 1234 for an image to be inspected. In the input field 2002 for the image generation unit to be used, the number of an image generation unit to be used in the comparison process of the inspection process flow is input.

[0057] As described above, according to this embodiment, in addition to the effects of the first embodiment, if some models included in the estimation model of multiple non-defective product images have failed to learn, the accuracy of the final inspection result decreases. Therefore, by using the output of only models that have been well trained, the accuracy can be further improved. The present invention is not limited to the above-mentioned embodiment, but includes various modified examples. For example, the above-mentioned embodiment has been described in detail to easily explain the present invention, and the present invention is not necessarily limited to those including all of the configurations described. In addition, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment.

[0058] In addition, the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. In addition, the above-mentioned configurations, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as the program, table, file, etc. that realizes each function can be stored in a memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD. In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0059] 1. Defect inspection system 2...Charged particle beam device 3...Computer (computer system) 100...imaging device 101...electron source 102...Extraction electrode 103...Charged particle beam 104…Condenser lens 105...Scanning deflector 106…Objective lens 107...Sample 108…Stage 109...electronic 110…Detector 200…Controller 201,310…Communication equipment 202…Processor 203,331…Memory 204,332…Display device 205,333...I / O devices 320…Storage 321…Learning image DB 322...Inspection image DB 323…Trained parameter DB 324…Program 330a…Learning Department 330b…Defect inspection department 400, 400a…Good product image estimation model 401…Estimation model for multiple good images 410…Feature extraction unit 420...Image generation unit 420a...first image generating unit 420b...Second image generating unit 430...Input image 440… Estimated good product image 700 Data Augmentation Processing 701...Contrast conversion processing 702…Brightness conversion processing 703…Distortion processing 800…Noise addition processing 803...Noise image 1000...Image to be inspected 1001...Comparison processing 1002...Differential image 1003, 1704…False report discrimination processing 1004…Defect map 1005,1701...Median processing 1006…Integrated defect map 1102…Defect candidate map 1105…Disinformation judgment 1301…First estimated good product image 1302…Second estimated good product image 1702…Integrated estimated good image 1902...Selective comparison processing 1903...Averaging process 1904…Merge differential image

Claims

1. A defect inspection system having an imaging device for acquiring an observation image of a sample, a learning unit that images the sample to obtain learning images, and learns an estimation model of a plurality of good images that estimates a plurality of good images of the sample for one input image using the learning images; a defect inspection unit that images a sample to be inspected using the imaging device to obtain an inspection image, inputs the inspection image into a trained estimation model of the multiple good product images, outputs a plurality of estimated good product images corresponding to the inspection image, and extracts defective areas using the multiple estimated good product images.

2. 2. The defect inspection system according to claim 1, The defect inspection system according to claim 1, wherein the estimation model for the plurality of non-defective product images includes a plurality of non-defective product image estimation models.

3. 2. The defect inspection system according to claim 1, A defect inspection system characterized in that the estimation model for the multiple good product images has a common image feature extraction unit and multiple image generation units that generate a single good product image based on the output of the image feature extraction unit.

4. 3. The defect inspection system according to claim 2, The defect inspection system is characterized in that the defect inspection unit creates a single integrated estimated good image from the multiple estimated good images using statistical processing, and compares the inspection image with the integrated estimated good image to extract defective areas.

5. 3. The defect inspection system according to claim 2, The defect inspection system is characterized in that the defect inspection unit compares the inspection image with each of the multiple estimated good product images, creates a defect map which is an image showing defective areas, performs statistical processing on the multiple defect maps, and outputs an integrated defect map.

6. 6. The defect inspection system according to claim 4 or 5, A defect inspection system characterized in that the statistical processing creates an image showing at least one of the average, median, mode, and maximum of pixel values ​​at the same position on multiple images to be processed.

7. 2. The defect inspection system according to claim 1, The defect inspection system is characterized in that the learning unit is a defect inspection system in which the learning images are defect images of good products, and the learning unit repeatedly updates internal parameters of the estimation model of the multiple defect images so as to minimize a reconstruction error, which is the difference between the learning image and an estimated defect image estimated from a noise image in which non-good product features have been added to the learning image.

8. 8. The defect inspection system according to claim 7, a learning unit that repeatedly updates internal parameters of the estimation model of the multiple good product images so as to minimize the reconstruction error and maximize the amount of good product conversion from the noise image to the estimated good product image.

9. 2. The defect inspection system according to claim 1, A defect inspection system characterized in that the learning unit performs at least one of brightness conversion, contrast conversion, and distortion addition on the training images before inputting the multiple non-defective product images into an estimation model.

10. 2. The defect inspection system according to claim 1, The defect inspection system according to the present invention is characterized in that the defect inspection unit extracts defective portions by using some of the plurality of estimated good images.

11. 7. The defect inspection system according to claim 6, A display device is provided, a display device that displays on a screen thereof an input field for inputting the number of estimated good product images to be output based on the estimated model of the plurality of good product images;

12. A defect inspection method for a defect inspection system having an imaging device for acquiring an observation image of a sample, comprising: a learning unit capturing an image of the sample to obtain a learning image, and learning a multiple-good-image estimation model that uses the learning image to estimate multiple good-quality images of the sample for one input image; a defect inspection unit capturing an image of a sample to be inspected using the imaging device to obtain an image to be inspected, inputting the image to be inspected into a trained estimation model of the multiple good product images, outputting a plurality of estimated good product images corresponding to the image to be inspected, and extracting defective areas using the multiple estimated good product images.

13. The defect inspection method according to claim 12, The defect inspection method according to the present invention, wherein the estimation model for the plurality of non-defective product images includes a plurality of non-defective product image estimation models.

14. The defect inspection method according to claim 13, a step in which the defect inspection unit creates one integrated estimated good image from the plurality of estimated good images using statistical processing; a step in which the defect inspection unit compares the inspection image with the integrated estimated non-defective image to extract a defective portion.

15. The defect inspection method according to claim 13, a step in which the defect inspection unit compares the inspection image with each of the plurality of estimated non-defective images to create a defect map which is an image showing defective portions; a step of the defect inspection unit performing statistical processing on a plurality of defect maps and outputting an integrated defect map.

Images