Image noise reduction method

Abstract

The invention relates to an image noise reduction method for reducing noise on an image generated by a scanning electron microscope. According to the image noise reduction method, a reference image of a sample is generated by using the scanning electron microscope, and artificial noise is added to the reference image to generate an artificial noise image. An anisotropic filter configured to stretch artificial noise in a scanning direction of an electron beam of a scanning electron microscope is applied to an artificial noise image to generate an artificial low-image-quality image, and machine learning is performed using training data including a reference image and the artificial low-image-quality image to generate a de-noising model. An image of a workpiece to be inspected and shape-measured for a semiconductor device is generated by a scanning electron microscope, the image of the workpiece is input to a denoising model, and a denoised image is output from the denoising model.

Description

Technical Field

[0001] The present invention relates to an image noise reduction method for reducing noise on an image generated by a scanning electron microscope. Background Art

[0002] In an image generated by a scanning electron microscope, noise sometimes appears due to various factors such as optical or electrical factors. As a solution for reducing noise on such an image, a denoising model can be used. More specifically, an image containing noise is input to the denoising model, and a denoised image with reduced noise is output from the denoising model. Such a denoising model is a learned model generated by machine learning.

[0003] Figure 4 1 is a schematic diagram showing an example of a conventional method for generating a denoising model by machine learning. Machine learning is performed using a plurality of sets of training data including high-quality images and low-quality images generated by a scanning electron microscope. Each set of high-quality images and low-quality images are images of the same part of a sample such as a wafer. The high-quality images are obtained under the condition of a low scanning rate, and the low-quality images are obtained under the condition of a high scanning rate.

[0004] However, this method requires that images of the same position of the sample be generated under different conditions (i.e., different scanning rates). This method of photographing is very time-consuming for a scanning electron microscope.

[0005] Figure 5 is a schematic diagram showing another example of a conventional method for generating a denoising model by machine learning. The machine learning is performed using training data including a plurality of sets of high-quality images and artificial noise images generated by a scanning electron microscope. Figure 4 The artificial noise image is obtained by adding artificial noise (e.g., Gaussian noise) to the obtained high-quality image. According to this method, the scanning electron microscope can generate images under the same conditions (same scanning rate).

[0006] However, artificial noise such as Gaussian noise does not resemble the actual noise on images generated by scanning electron microscopes, and as a result, denoising models generated by machine learning may not be able to sufficiently reduce the noise. Prior art literature Patent Literature

[0007] Patent Document 1: Japanese Patent Application Publication No. 2021-197144 Summary of the invention Problem that the invention aims to solve

[0008] Therefore, the present invention provides an image noise reduction method capable of reducing noise in an image generated by a scanning electron microscope with high accuracy. Technical means of solving problems

[0009] In one embodiment, an image noise reduction method is provided for reducing noise in an image generated by a scanning electron microscope, wherein a reference image of a sample is generated using the scanning electron microscope, artificial noise is added to the reference image to generate an artificial noise image, an anisotropic filter configured to stretch the artificial noise in a scanning direction of an electron beam of the scanning electron microscope is applied to the artificial noise image, thereby generating an artificial low-quality image, machine learning is performed using training data including the reference image and the artificial low-quality image, thereby generating the denoising model, an image of a workpiece that is an object of inspection and shape measurement of a semiconductor device is generated using the scanning electron microscope, the image of the workpiece is input into the denoising model, and a denoised image is output from the denoising model.

[0010] In one embodiment, the anisotropic filter has at least a scanning rate and a scanning direction of a scanning electron beam and a response characteristic of an electron detector of the scanning electron microscope as parameters, and the parameters are set to be consistent with the parameters when an image of the workpiece to be denoised is taken. In one embodiment, the anisotropic filter has parameters including at least the material and cross-sectional structure of the sample surface and the accelerating voltage and current of the irradiated electron beam, and the parameters are set to match the parameters when capturing an image of the workpiece to be denoised. In one embodiment, the artificial noise is noise that follows a statistical distribution. In one embodiment, the artificial noise is Poisson noise. In one embodiment, the artificial noise is noise following a normal distribution or a log-normal distribution. Effects of the Invention

[0011] According to the present invention, since an anisotropic filter that simulates the noise peculiar to a scanning electron microscope is applied to an artificial noise image, an artificial low-quality image similar to an image containing actual noise can be generated. By using training data containing such an artificial low-quality image and a reference image (high-quality image) for machine learning, a denoising model with high noise reduction accuracy can be generated. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a schematic diagram showing one embodiment of an image generation system. Figure 2This is a diagram illustrating an embodiment of a method for generating a denoising model through machine learning. Figure 3 : is a diagram showing an example of a reference image, an artificial noise image, and an artificial low-quality image. Figure 4 This is a diagram illustrating an example of a conventional method for generating a denoising model through machine learning. Figure 5 This is a diagram illustrating another example of a conventional method for generating a denoising model through machine learning. DETAILED DESCRIPTION

[0013] Hereinafter, embodiments of the present invention will be described with reference to the drawings. Figure 1 1 is a schematic diagram showing an embodiment of an image generation system. The image generation system includes a scanning electron microscope 1 for generating an image of a workpiece W and a processing system 5 for processing the image generated by the scanning electron microscope 1. Examples of the workpiece W include wafers, masks, panels, substrates, etc., which are targets for inspection and shape measurement of semiconductor devices.

[0014] The processing system 5 is composed of at least one computer. The processing system 5 includes a storage device 5a storing a program and a computing device 5b that performs operations according to commands included in the program. The storage device 5a includes a main storage device such as a random access memory (RAM), a hard disk drive (HDD), a solid state drive (SSD) and other auxiliary storage devices. Examples of the computing device 5b include a CPU (central processing unit) and a GPU (graphics processing unit). However, the specific structure of the processing system 5 is not limited to this embodiment.

[0015] The processing system 5 may be an edge server connected to the scanning electron microscope 1 through a communication line, may be a cloud server connected to the scanning electron microscope 1 through a communication network such as the Internet or a local area network, or may be a fog computing device (gateway, fog server, router, etc.) installed in a network connected to the scanning electron microscope 1. The processing system 5 may be a combination of multiple servers. For example, the processing system 5 may be a combination of an edge server and a cloud server connected to each other through a communication network such as the Internet or a local area network. In another example, the processing system 5 may include multiple servers (computers) that are not connected through a network.

[0016] The scanning electron microscope 1 includes an electron gun 15 for emitting an electron beam, a focusing lens 16 for focusing the electron beam emitted from the electron gun 15, an X deflector 17 for deflecting the electron beam in the X direction, a Y deflector 18 for deflecting the electron beam in the Y direction, an objective lens 20 for focusing the electron beam on a workpiece W as an example of a sample, a workpiece stage 31 for supporting the workpiece W, and a stage moving device 35 for moving the workpiece stage 31 in translation. The configuration of the electron gun 15 is not particularly limited. For example, a field emission type electron gun or a semiconductor photocathode type electron gun can be used as the electron gun 15.

[0017] The electron beam emitted from the electron gun 15 is focused by the focusing lens 16, deflected by the X deflector 17 and the Y deflector 18, and focused by the objective lens 20 to irradiate the surface of the workpiece W. When the primary electrons of the electron beam are irradiated to the workpiece W, electrons such as secondary electrons and reflected electrons are emitted from the workpiece W. The electrons emitted from the workpiece W are detected by the electron detector (scintillator) 26. The electron detection signal of the electron detector 26 is input to the image acquisition device 28 and converted into an image. In this way, the scanning electron microscope 1 generates an image of the surface of the workpiece W. The image acquisition device 28 is connected to the processing system 5.

[0018] The processing system 5 has a program in the storage device 5a for generating a denoising model for reducing noise from an image generated by the scanning electron microscope 1. Figure 2 An embodiment of a method for generating a denoising model through machine learning is described.

[0019] The processing system 5 sends instructions to the scanning electron microscope 1 to generate a reference image of the sample. Figure 1 The workpiece W shown in the figure may have the same structure as the workpiece W, or may have other structures. Examples of samples include wafers, masks, panels, substrates, etc., which are objects of inspection and shape measurement of semiconductor devices. The reference image is an image used as a correct label for machine learning and is a high-quality image. More specifically, the scanning electron microscope 1 can generate a reference image as a high-quality image by scanning the surface of the sample with an electron beam at a low scanning rate.

[0020] The processing system 5 acquires a reference image from the scanning electron microscope 1. Further, the processing system 5 adds artificial noise to the reference image to generate an artificial noise image. Artificial noise is noise that is statistically independent between pixels of the reference image. In other words, artificial noise is noise that has no correlation between pixels of the reference image. More specifically, artificial noise is noise that follows a statistical distribution, and specific examples thereof include Poisson noise, Gaussian noise, and noise that follows a log-normal distribution.

[0021] In this embodiment, Poisson noise is used as artificial noise. This is because Poisson noise is similar to the noise that appears on the image generated by the scanning electron microscope 1. Therefore, it can be expected that the noise on the image generated by the scanning electron microscope 1 can be reduced with high accuracy by the denoising model constructed by machine learning described later.

[0022] The processing system 5 generates an artificial low-quality image by applying an anisotropic filter to the artificial noise image generated as described above. The anisotropic filter is configured to make the artificial noise on the artificial noise image close to the actual noise on the image generated by the scanning electron microscope 1. The actual noise on the image is affected by the scanning operation of the electron beam, the afterglow caused by the electron detector (scintillator) 26, and the charging of the photographed object. The anisotropic filter is a filter for simulating the noise unique to the scanning electron microscope 1.

[0023] The anisotropic filter is arranged to stretch the artificial noise in the scanning direction of the electron beam of the scanning electron microscope 1. In one embodiment, the anisotropic filter is represented by the following formula. [Formula 1] Here, v(x, y) represents the brightness value of the pixel at the coordinate (x, y) after the anisotropic filter is applied, sampled (x, y) represents the brightness value of the pixel at the coordinate (x, y), e represents a natural constant, τ represents an attenuation constant, and the symbol * represents a convolution operation. The attenuation constant τ is a theoretical value determined by the scanning direction and scanning speed of the electron beam, the material of the sample, and the response characteristics of the electron detector (scintillator) 26, and can be calculated. In this embodiment, the scanning direction of the electron beam is equivalent to the X direction.

[0024] In one embodiment, the anisotropic filter has at least the scanning rate and scanning direction of the scanning electron beam and the response characteristics of the electron detector 26 as parameters, and these parameters are set to be consistent with the parameters when the image of the denoising object is captured.

[0025] In one embodiment, the anisotropic filter has at least the material and cross-sectional structure of the sample surface, and the acceleration voltage and current of the irradiated electron beam as parameters, and these parameters are set to be consistent with the parameters when the image of the denoising object is captured.

[0026] Figure 3 The figure shows an example of a reference image, an artificial noise image obtained by adding artificial noise to the reference image, and an artificial low-quality image obtained by applying an anisotropic filter to the artificial noise image. The artificial low-quality image is an image close to an actual low-quality image generated by the scanning electron microscope 1 .

[0027] The anisotropy in the actual image (SEM image) generated by the scanning electron microscope 1 is also caused by the charging of the sample. In this case, a transfer function that depends on the material of the sample surface, the cross-sectional structure of the sample, the acceleration voltage and current of the electron beam, and the scanning rate of the electron beam can be used instead of the above formula (1).

[0028] The artificial low-quality image generated by applying an anisotropic filter to the artificial noise image and the above-mentioned reference image generated by the scanning electron microscope 1 are used as training data for machine learning. That is, the artificial low-quality image is used as an explanatory variable and the reference image is used as a target variable (correct label). The processing system 5 performs machine learning of the denoising model so that when the artificial low-quality image is input to the denoising model, the difference between the image output from the denoising model and the reference image is minimized.

[0029] Examples of machine learning include SVR (Support Vector Regression), PLS (Partial Least Squares), deep learning, random forest, and decision tree. In one example, the denoising model is composed of a neural network constructed by a deep learning method.

[0030] The denoising model generated by machine learning using training data including a reference image and an artificial low-quality image is stored in the storage device 5 a of the processing system 5 as a learned model.

[0031] The denoising model generated in this way can reduce with high accuracy the noise on the image generated by the scanning electron microscope 1. That is, the scanning electron microscope 1 generates an image of the workpiece W that is the object of inspection and shape measurement of the semiconductor device, the processing system 5 acquires the image (SEM image) of the workpiece W from the scanning electron microscope 1, inputs the image of the workpiece W into the denoising model, and outputs a denoised image as an image with reduced noise from the denoising model.

[0032] According to the present embodiment, since an anisotropic filter that simulates the noise peculiar to the scanning electron microscope 1 is applied to the artificial noise image, it is possible to produce an artificial low-quality image similar to an image containing actual noise. By performing machine learning using training data containing such an artificial low-quality image and a reference image (high-quality image), a highly accurate denoising model can be generated.

[0033] The following describes an example of applying the above-mentioned denoising effect to the measurement of semiconductor devices. Here, the example describes the measurement of the roughness of the edges of each line of a simple stripe pattern in which linear patterns are arranged side by side.

[0034] The scanning electron microscope 1 generates a plurality of SEM images scanned at a scanning rate of 3 MHz. The processing system 5 applies the anisotropic filter represented by the above equation (1) to these images to generate a plurality of artificial low-quality images to which artificial noise equivalent to a scanning rate of 100 MHz is added.

[0035] Next, the processing system 5 uses the pairing of the SEM image (reference image) with a scanning rate of 3 MHz and the artificial low-quality image as training data, performs machine learning to generate the SEM image (reference image) from the artificial low-quality image, and generates a denoising model.

[0036] The processing system 5 uses the obtained denoising model to perform denoising on the SEM image actually taken at a scanning rate of 100 MHz. When the edge roughness measured under the high-quality image condition of the scanning rate of 3 MHz is used as the reference, the correlation with the edge roughness measured from the above denoised image is about 0.8. This is the same as the result measured under the scanning rate of 25 MHz, that is, it is confirmed that the shooting speed can be increased by 4 times by using this denoising process.

[0037] The above-mentioned embodiments are recorded for the purpose of enabling a person with common knowledge in the technical field to which the present invention belongs to implement the present invention. For those skilled in the art, various modifications of the above-mentioned embodiments are certainly possible, and the technical idea of ​​the present invention can also be applied to other embodiments. Therefore, the present invention is not limited to the described embodiments, but should be interpreted in the broadest scope of the technical idea defined by the technical solution. Possibility of Industrial Application

[0038] The present invention can be used in an image noise reduction method for reducing noise on an image generated by a scanning electron microscope. Explanation of symbols

[0039] 1 Scanning electron microscopy 5. Processing System 15 Electron Gun 16 Focusing lens 17 X Deflector 18 Y deflector 20 Objective lens 26 Electronic Detector 28 Image acquisition device 31 Workpiece platform 35 Platform moving device.

Claims

1. An image noise reduction method for reducing noise in an image generated by a scanning electron microscope, characterized in that generating a reference image of the sample using the scanning electron microscope, adding artificial noise to the reference image to generate an artificial noise image, applying an anisotropic filter to the artificial noise image, thereby generating an artificial low-quality image, the anisotropic filter being constructed in such a way as to stretch the artificial noise in the scanning direction of the electron beam of the scanning electron microscope, The denoising model is generated by performing machine learning using training data including the reference image and the artificial low-quality image, generating an image of a workpiece to be inspected and the shape of a semiconductor device by using the scanning electron microscope, inputting the image of the workpiece into the denoising model, A denoised image is output from the denoising model.

2. The image noise reduction method according to claim 1, wherein: The anisotropic filter has at least a scanning rate and a scanning direction of a scanning electron beam and a response characteristic of an electron detector of the scanning electron microscope as parameters, and the parameters are set to be consistent with the parameters when an image of the workpiece to be denoised is captured.

3. The image noise reduction method according to claim 1, wherein: The anisotropic filter has at least the material and cross-sectional structure of the sample surface and the accelerating voltage and current of the irradiated electron beam as parameters, and the parameters are set to match the parameters when the image of the workpiece to be denoised is captured.

4. The image noise reduction method according to claim 1, wherein: The artificial noise is noise that follows a statistical distribution.

5. The image noise reduction method according to claim 4, characterized in that: The artificial noise is Poisson noise.

6. The image noise reduction method according to claim 4, characterized in that: The artificial noise is noise that follows a normal distribution or a log-normal distribution.

Images