Training method, inference model, noise reduction method, noise reduction system, and noise reduction program

Abstract

The present invention appropriately reduces flicker due to random noise in an analysis image used for failure analysis of a semiconductor device.　This training method generates an inference model used for reducing flicker due to random noise in an analysis image used for failure analysis of a semiconductor device. The training method comprises: a training purpose acquisition step (S01) for acquiring a training analysis image including random noise and a target signal pertaining to the semiconductor device, and a training noise image not including the target signal and including random noise; noise addition steps (S02, S03) for reducing the contrast of the training analysis image, adding random noise of a training noise image to the training analysis image with the reduced contrast, and generating a noise-added training image; and a training step (S04) for training machine learning of the inference model by using the noise-added training image.

Description

Learning method, inference model, noise reduction method, noise reduction system, and noise reduction program

[0001] The present invention relates to a learning method for generating an inference model used to reduce flicker caused by noise from an image, the generated inference model, and a noise reduction method, noise reduction system, and noise reduction program for reducing flicker caused by noise from an image using the generated inference model.

[0002] Patent Document 1 discloses that noise is reduced from an image using an inference model generated by machine learning.

[0003] JP 2023-139782 A

[0004] Analysis images, such as light emission images, used in failure analysis of semiconductor devices contain random noise. In failure analysis of semiconductor devices, a pattern image of the semiconductor device and an analysis image are sometimes superimposed and displayed. However, when the pattern image of the semiconductor device and the analysis image are superimposed and displayed, flickering caused by random noise can make the pattern image of the semiconductor device difficult to view. Therefore, when performing failure analysis using an analysis image, it is necessary to reduce the flickering caused by random noise.

[0005] Conventional inference models, including that shown in Patent Document 1 above, are not specialized for reducing flicker caused by random noise from analyzed images, and there is a risk that they may not be able to adequately reduce flicker caused by random noise.

[0006] One embodiment of the present invention has been made in consideration of the above, and aims to provide a learning method, an inference model, a noise reduction method, a noise reduction system, and a noise reduction program that can appropriately reduce flicker caused by random noise from analysis images used in failure analysis of semiconductor devices.

[0007] In order to achieve the above object, a learning method according to one embodiment of the present invention is a learning method for generating an inference model used to reduce flicker caused by random noise from an analysis image used in failure analysis of a semiconductor device, and includes a learning acquisition step for acquiring a learning analysis image that includes a target signal related to the semiconductor device and also includes random noise, and a learning noise image that does not include the target signal and also includes random noise; a noise addition step for reducing the contrast of the learning analysis image acquired in the learning acquisition step, and adding random noise from the learning noise image to the contrast-reduced learning analysis image, thereby generating a learning noise-added image; and a training step for training machine learning of the inference model using the learning noise-added image generated in the noise addition step.

[0008] In a learning method according to one embodiment of the present invention, random noise from a learning noise image is added to a contrast-reduced analysis image for learning to generate a noise-added image for learning. An inference model generated using such a noise-added image for learning can appropriately reduce flickering caused by random noise from the analysis image. Therefore, according to the learning method according to one embodiment of the present invention, flickering caused by random noise can be appropriately reduced from the analysis image.

[0009] In the training step, the contrast-reduced analysis image for learning may be an image that does not contain random noise, and the noise-added image for learning may be an image that contains random noise, and machine learning training of the inference model may be performed using this configuration. This configuration allows for more appropriate machine learning training of the inference model, thereby generating a more appropriate inference model. As a result, flicker due to random noise can be more appropriately reduced in the analysis image.

[0010] In the noise addition step, a pixel value corresponding to the noise image for training may be subtracted from the pixel value of each pixel in the analysis image for training, the pixel value of each pixel after subtraction may be multiplied by a preset coefficient less than 1, and the subtracted pixel value may be added to the pixel value of each pixel after multiplication to generate an analysis image for training with reduced contrast. This configuration makes it possible to generate a more appropriate analysis image for training with reduced contrast. As a result, flicker due to random noise can be more appropriately reduced in the analysis image.

[0011] In the noise addition step, multiple contrast-reduced analysis images for learning may be generated using multiple coefficients that are different from each other, thereby generating multiple noise-added images for learning. This configuration allows multiple contrast-reduced analysis images for learning to be generated from a single analysis image for learning, thereby increasing the variety of images used for machine learning training of the inference model. This allows for more appropriate machine learning training of the inference model and the generation of a more appropriate inference model. As a result, flickering due to random noise can be more appropriately reduced in the analysis image.

[0012] In the acquisition step for learning, an original image larger in size than the analysis image for learning, including a portion that can become the analysis image for learning, may be acquired, a portion of the acquired original image may be cropped, a value based on pixel values ​​of the cropped image may be compared with a value corresponding to pixel values ​​of the noise image for learning, and the cropped image may be used as the analysis image for learning according to the comparison result.Furthermore, in the acquisition step for learning, a value based on pixel values ​​of the cropped image may be compared with an average or variation value of pixel values ​​of the noise image for learning, and the cropped image may be used as the analysis image for learning according to the comparison result.With these configurations, an analysis image for learning can be acquired easily and appropriately.

[0013] The inference model generated by the learning method according to one embodiment of the present invention is itself an invention having a novel configuration. That is, the inference model according to one embodiment of the present invention is an inference model for inputting information based on an analysis image used in failure analysis of a semiconductor device, performing calculations according to the input, and outputting information, and is generated by the learning method described above.

[0014] In order to achieve the above object, a noise reduction method according to one embodiment of the present invention is a noise reduction method for reducing flickering due to random noise from an analysis image used in failure analysis of a semiconductor device, and includes a reduction acquisition step for acquiring an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise, and an image generation step for generating an inference image in which flickering due to random noise has been reduced from the analysis image for reduction acquired in the reduction acquisition step, using an inference model generated in advance by machine learning training.

[0015] According to the noise reduction method according to one embodiment of the present invention, flicker caused by random noise can be appropriately reduced from the analysis image.

[0016] The inference model may be an inference model generated by the above-described learning method. With this configuration, flicker caused by random noise can be appropriately reduced from the analysis image.

[0017] In the reduction acquisition step, a pattern image of the semiconductor device corresponding to the reduction analysis image may also be acquired, and in the image generation step, a composite image may be generated by superimposing the generated inference image and the pattern image acquired in the reduction acquisition step. With this configuration, a composite image may be obtained by superimposing the inference image in which flicker due to random noise has been appropriately reduced and the pattern image.

[0018] One embodiment of the present invention can be described not only as a noise reduction method invention as described above, but also as a noise reduction system and a noise reduction program invention as described below. These are essentially the same invention, and achieve similar actions and effects, just in different categories.

[0019] That is, a noise reduction system according to one embodiment of the present invention is a noise reduction system for reducing flicker caused by random noise from an analysis image used in failure analysis of a semiconductor device, and includes: a reduction acquisition means for acquiring an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise; and an image generation means for generating an inference image in which flicker caused by random noise has been reduced from the analysis image for reduction acquired by the reduction acquisition means, using an inference model generated in advance by machine learning training. The inference model may be an inference model generated by the learning method described above.

[0020] Furthermore, a noise reduction program according to one embodiment of the present invention is a noise reduction program for reducing flicker caused by random noise from an analysis image used in failure analysis of a semiconductor device, and causes a computer to execute a reduction acquisition process for acquiring an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise, and an image generation process for generating an inference image in which flicker caused by random noise has been reduced from the analysis image for reduction acquired by the reduction acquisition process using an inference model generated in advance by machine learning training. The inference model may be an inference model generated by the learning method described above.

[0021] According to one embodiment of the present invention, flicker caused by random noise can be appropriately reduced from an analysis image used in failure analysis of a semiconductor device.

[0022] FIG. 1 is a diagram illustrating the configuration of a learning system and a noise reduction system according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an inference model generated and used in the embodiment, together with example images to be input and output. FIG. 3 is an image used in the embodiment. FIG. 4 is a graph illustrating an image used in the embodiment. FIG. 5 is a diagram illustrating an overview of machine learning training of the inference model in the embodiment. FIG. 6 is a flowchart illustrating a learning method which is a process executed in the learning system according to an embodiment of the present invention. FIG. 7 is a flowchart illustrating a noise reduction method which is a process executed in the noise reduction system according to an embodiment of the present invention. FIG. 8 is an example of noise-induced flicker reduction performed by the embodiment. FIG. 9 is an example of noise-induced flicker reduction performed by the embodiment. FIG. 10 is an example of a composite image (superimposed image) generated by the embodiment. FIG. 11 is a diagram illustrating the configuration of a learning program and a noise reduction program according to an embodiment of the present invention, together with a recording medium.

[0023] Hereinafter, embodiments of a learning method, an inference model, a noise reduction method, a noise reduction system, and a noise reduction program according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are given the same reference numerals, and duplicated explanations will be omitted.

[0024] Fig. 1(a) shows a learning system 10 that executes the learning method according to this embodiment. Fig. 1(b) shows a noise reduction system 20 that executes the noise reduction method according to this embodiment. The learning system 10 is a system (apparatus) that generates an inference model used to reduce (remove) flicker caused by random noise from an analysis image used in failure analysis of a semiconductor device. The noise reduction system 20 is a system (apparatus) that reduces flicker caused by random noise from the analysis image using the inference model generated by the learning system 10.

[0025] The analysis image to be subjected to the reduction of flicker due to random noise is, for example, a light emission image used in light emission analysis. The light emission image is obtained by capturing an image of a semiconductor device to be subjected to failure analysis using an imaging device such as an InGaAs camera. The light emission image captures light emission at locations corresponding to the presence or absence of a failure in the semiconductor device and the location of the failure. In the light emission image, areas other than the light-emitting locations of the semiconductor device can be considered as the background.

[0026] By analyzing the light emission in the light emission image, it is possible to perform an analysis (failure analysis) of the semiconductor device to determine whether or not there is a fault, the location of the fault, etc. The failure analysis of the semiconductor device using the noise-reduced light emission image may be performed in the same manner as conventional methods. The failure analysis of the semiconductor device is performed, for example, during the inspection of semiconductor devices manufactured on a production line. Furthermore, the failure analysis of the semiconductor device is performed on semiconductor devices that have been found to be faulty or operating abnormally through separate inspections. Examples of semiconductor devices that are the subject of failure analysis include microcontrollers, integrated circuits (ICs), analog ICs, hybrid ICs, and other such devices, memory devices such as volatile memories or nonvolatile memories, power semiconductor devices, optical semiconductor devices, and discrete devices. The analysis image for which flicker due to random noise is reduced may be an image other than a light emission image, as long as it is used for the failure analysis of semiconductor devices. For example, the analysis image may be one obtained by imaging with a thermal camera (heat image), one obtained by imaging with a SI-CCD camera, or one obtained by OBIRCH (Optical Beam Induced Resistance Change) analysis, OBIC (Optical Beam Induced Current) analysis, or EOFM (Electro Optical Frequency Mapping) analysis.

[0027] The analysis image includes random noise, which is the target of reduction in this embodiment. Random noise is noise that occurs irregularly in time and space depending on, for example, the imaging conditions and the imaging device. In particular, because random noise occurs as a spatial flicker, the entire analysis image appears to flicker due to the random noise. This reduces visibility when the analysis image is displayed superimposed on other images, such as a pattern image of a semiconductor device, and reduces the accuracy of failure analysis. An analysis image with reduced random noise flicker improves visibility when displayed superimposed on other images. Therefore, using an analysis image with reduced random noise flicker can improve the accuracy of failure analysis. When the analysis image is an emission image, emission analysis performance can be improved.

[0028] FIG. 2 shows an example of a luminescence image 31 containing random noise, and an example of a luminescence image 32 (denoise image) in which random noise has been reduced by using an inference model.

[0029] The learning system 10 and the noise reduction system 20 are configured to include a conventional computer including hardware such as a processor, a memory, a communication module, etc. The processor is, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit).

[0030] Furthermore, the learning system 10 and the noise reduction system 20 may include an imaging device such as a camera for acquiring images to be used for processing. Alternatively, the learning system 10 and the noise reduction system 20 may be included in an imaging device such as a camera for acquiring images to be used for processing. Although the learning system 10 and the noise reduction system 20 are shown as separate systems (devices) in Fig. 1, they may also be realized by the same system (device).

[0031] The computers that make up the learning system 10 and the noise reduction system 20 may be computer systems including multiple computers. Furthermore, the computers may be configured using cloud computing or edge computing. The functions of the learning system 10 and the noise reduction system 20, which will be described later, are realized by these components operating through programs or the like. For example, the functions of the learning system 10 and the noise reduction system 20, which will be described later (particularly the functions of the noise reduction system 20), may be realized by a dedicated application compatible with the imaging device.

[0032] Next, the functions of the learning system 10 and noise reduction system 20 according to this embodiment will be described. As shown in FIG. 1, the learning system 10 includes a learning acquisition unit 11, a noise addition unit 12, and a training unit 13.

[0033] The learning system 10 generates an inference model through machine learning training. The inference model includes, for example, a neural network. The neural network may be multi-layered. That is, the inference model may be generated through deep learning. The neural network may also be a convolutional neural network (CNN). The format of the inference model may be similar to an inference model generated through training using conventional machine learning (e.g., Noise2Clean).

[0034] The inference model is used to reduce flicker caused by random noise from an analysis image. For example, the inference model inputs an analysis image for which flicker caused by random noise is to be reduced, and outputs the analysis image after the flicker caused by random noise has been reduced. This inference model is a denoising model (noise removal model), which is a filter that improves flicker caused by random noise in the analysis image. In this case, the inference model has neurons in its input layer for inputting the analysis image for which flicker caused by random noise is to be reduced. For example, the information input to the inference model is the pixel value (luminance value) of each pixel in the analysis image for which flicker caused by random noise is to be reduced. In this case, the input layer has neurons equal to the number of pixels in the analysis image, and each neuron receives the pixel value of the corresponding pixel.

[0035] The inference model also has neurons in its output layer for outputting an analysis image after flickering due to random noise has been reduced. For example, the information output from the inference model is the pixel value of each pixel in the analysis image after flickering due to random noise has been reduced. In this case, the output layer has neurons equal to the number of pixels in the analysis image, and each neuron outputs the pixel value of the corresponding pixel.

[0036] Alternatively, the inference model may input an analysis image of a target for reducing flicker due to random noise and output an image (data) of random noise contained in the analysis image. In this case, the output image of random noise can be removed from the analysis image of a target for reducing flicker due to random noise to obtain an analysis image after reducing flicker due to random noise. The inference model may also be other than those described above, as long as it is used to reduce flicker due to random noise from the analysis image of a target for reducing flicker due to random noise.

[0037] An inference model is a model that functions to input information based on an analysis image, perform calculations according to the input, and output information. It is anticipated that the inference model will be used as a program module that is part of artificial intelligence software. The inference model is used, for example, in a computer having a processor and memory, and the computer's processor operates according to instructions from the model stored in the memory. For example, the computer's processor operates in accordance with the instructions to input information to the model, perform calculations according to the model, and output results from the model. Specifically, the computer's processor operates in accordance with the instructions to input information to the input layer of a neural network, perform calculations based on parameters such as learning weighting coefficients in the neural network, and output results from the output layer of the neural network. Note that the inference model may be configured using something other than a neural network.

[0038] The learning acquisition unit 11 is a learning acquisition means that acquires a learning analysis image that includes a target signal related to a semiconductor device and also includes random noise, and a learning noise image that does not include the target signal and also includes random noise. The learning acquisition unit 11 may acquire an original image that includes a portion that can become the learning analysis image and is larger in size than the learning analysis image, crop out a portion of the acquired original image, compare a value based on pixel values ​​of the cropped image with a value corresponding to pixel values ​​of the learning noise image, and use the cropped image as the learning analysis image depending on the comparison result. The learning acquisition unit 11 may also compare a value based on pixel values ​​of the cropped image with an average or variation value of pixel values ​​of the learning noise image, and use the cropped image as the learning analysis image depending on the comparison result.

[0039] The learning analysis image and learning noise image acquired by the learning acquisition unit 11 are for generating an inference model. The learning acquisition unit 11 acquires the learning analysis image and learning noise image, for example, as follows. The learning analysis image and learning noise image are images of the same format as the analysis image that is the target of reducing flickering due to random noise. For example, the learning analysis image and learning noise image are images of the same size as the analysis image that is the target of reducing flickering due to random noise, and are luminescent images similar to the analysis image that is the target of reducing flickering due to random noise.

[0040] The analysis image for training and the noise image for training contain random noise. The analysis image for training contains a target signal related to a semiconductor device. The target signal related to a semiconductor device is a signal used for failure analysis of the semiconductor device in the analysis image. For example, if the analysis image is an image of light emission, the target signal is a signal related to light emission. Also, if the analysis image is an image of heat generation, the target signal is a signal related to heat generation (e.g., heat generation due to operation of the semiconductor device). The background portion of the analysis image that does not emit light or generate heat is not related to the target signal. The noise image for training does not contain a target signal related to a semiconductor device. For example, the noise image for training is an image that captures only the portion corresponding to the background of the analysis image for training. If the analysis image for which flicker due to random noise is reduced is an image of light emission, the analysis image for training is an image with light emission related to a semiconductor device, and the noise image for training is an image without light emission related to a semiconductor device. If the analysis image for which flicker due to random noise is reduced is an image of heat generation, the analysis image for training is an image with heat generation related to a semiconductor device, and the noise image for training is an image without heat generation related to a semiconductor device.

[0041] The analysis image for learning may have a larger pixel value (higher brightness) of the target signal related to the semiconductor device, in order to generate an inference model more appropriately.

[0042] The learning acquisition unit 11 acquires an original image that is larger in size than the analysis image for learning and includes a portion that can become the analysis image for learning. The original image including the analysis image for learning is, for example, an image obtained by capturing an image of a semiconductor device for learning prepared in advance with an imaging device in a state in which random noise occurs in the captured image. The semiconductor device for learning may be a semiconductor device of the same type as the analysis image for which flicker due to random noise is to be reduced, or may be a semiconductor device of a different type. The original image is acquired, for example, by receiving it from the imaging device that captured the original image, or by accepting a user's input operation of the original image to the learning system 10.

[0043] As shown in FIG. 3 , the learning acquisition unit 11 cuts out an image of a region (a part of the original image) of a predetermined size smaller than the original image 41 from the original image 41, and sets the cut-out image as a learning analysis image 42. The cut-out portion may be a random portion or a predetermined portion. The cut-out portion is assumed to include a target signal related to a semiconductor device. The detection of a portion of the original image that includes a target signal related to a semiconductor device may be performed by any method as needed. The learning acquisition unit 11 may acquire multiple learning analysis images 42 from multiple different portions of one original image 41.

[0044] For example, the analysis image 42 for learning may be cut out to a size that is 10% to 20% of the size (field of view size) of the original image 41. If the size of the original image 41 is 1000 pixels x 1000 pixels, the size of the cut-out analysis image 42 for learning may be 128 pixels x 128 pixels.

[0045] The learning acquisition unit 11 acquires an original image that includes a portion that may become a learning noise image and is larger in size than the learning noise image. The original image including the learning noise image is, for example, an image obtained by capturing an image with an imaging device in a state in which random noise occurs in the captured image and in which an imaging target (semiconductor device) is not present. The original image is acquired, for example, by receiving it from the imaging device that captured the original image or by accepting a user's input operation of the original image to the learning system 10.

[0046] 3 , the learning acquisition unit 11 cuts out an image of a region (a part of the original image) having a predetermined size smaller than the original image 51 from the original image 51, and sets the cut-out image as a learning noise image 52. The cut-out portion may be a random portion or a predetermined portion. The learning acquisition unit 11 may acquire a plurality of learning noise images 52 from a plurality of different portions of one original image 51.

[0047] Note that the imaging device for acquiring the learning analysis images and the learning noise images may capture images such that the learning analysis images and the learning noise images contain random noise similar to that of the analysis images for which random noise flickering is to be reduced. By including similar random noise in these images, the generated inference model can more appropriately reduce random noise flickering from the analysis images for which random noise flickering is to be reduced. For example, the imaging device for acquiring the learning analysis images and the learning noise images may capture images using the same type of imaging device and under the same imaging conditions as the imaging device for acquiring the analysis images for which random noise flickering is to be reduced.

[0048] Furthermore, when cutting out the learning analysis image 42 from the original image 41, the learning acquisition unit 11 may determine, based on the learning noise image, whether or not to adopt the cut-out image as the learning analysis image 42. In this case, the learning acquisition unit 11 compares a value based on the pixel values ​​of the cut-out image with a value corresponding to the pixel values ​​of the learning noise image, and designates the cut-out image as the learning analysis image according to the comparison result. For example, the learning acquisition unit 11 determines whether to adopt the cut-out image as the learning analysis image 42 as follows:

[0049] The learning acquisition unit 11 calculates the average pixel values ​​of the learning noise image 52. This average pixel value corresponds to the average pixel values ​​of the background in the learning noise image 52. The learning acquisition unit 11 calculates a value by multiplying the calculated average pixel values ​​by a preset coefficient greater than 1 (for example, 3). The learning acquisition unit 11 compares the calculated value with the pixel values ​​of each pixel of the partial image cut out from the original image 41.

[0050] If the partial image cut out from the original image 41 includes pixels having pixel values ​​equal to or greater than the calculated value, the learning acquisition unit 11 determines that the cut-out partial image should be used as the learning analysis image 42. This is because pixels having pixel values ​​that significantly exceed the average pixel value of the learning noise image 52 (average pixel values ​​of the background) are considered to be related to a target signal related to the semiconductor device. Alternatively, if the number of pixels having pixel values ​​equal to or greater than the calculated value in the partial image cut out from the original image 41 is equal to or greater than a preset threshold, the learning acquisition unit 11 may determine that the cut-out partial image should be used as the learning analysis image 42.

[0051] Furthermore, instead of or in addition to the above-described method, the learning acquisition unit 11 may make the above-described determination using the following method. The learning acquisition unit 11 calculates the variance (e.g., standard deviation (S.D.)) of pixel values ​​of the learning noise image 52. When acquiring multiple learning noise images 52, the learning acquisition unit 11 may calculate the average of the variances of the multiple learning noise images 52. The learning acquisition unit 11 calculates the variance of pixel values ​​of a partial image cut out from the original image 41. The calculated variance of pixel values ​​of the partial image is the same as the variance of pixel values ​​of the learning noise image 52 (e.g., standard deviation).

[0052] The learning acquisition unit 11 compares these variations. If the variation in pixel values ​​of the partial image cut out from the original image 41 is greater than the variation in pixel values ​​of the learning noise image 52, the learning acquisition unit 11 determines to use the cut-out partial image as the learning analysis image 42. This is because the variation in pixel values ​​of an image including a portion corresponding to a target signal related to a semiconductor device is considered to be greater than the pixel values ​​of the learning noise image 52. For example, the standard deviation of the pixel values ​​of the partial image including a portion corresponding to the target signal is 11.2, and the standard deviation of the pixel values ​​of the noise image 52 is 4.7.

[0053] Alternatively, the learning acquisition unit 11 may extract a plurality of (e.g., 100) partial images and determine that, among the partial images having a pixel value variation greater than the pixel value variation of the learning noise image 52, the partial image having the largest variation should be adopted as the learning analysis image 42. Furthermore, a value obtained by multiplying the pixel value variation of the learning noise image 52 by a preset coefficient greater than 1 may be used in the above determination.

[0054] Note that the learning acquisition unit 11 may acquire learning analysis images and learning noise images by methods other than those described above. The number of acquired learning analysis images and learning noise images is sufficient for processing to generate an inference model. The learning acquisition unit 11 outputs the acquired learning analysis images and learning noise images to the noise addition unit 12.

[0055] The noise addition unit 12 is a noise addition means that reduces the contrast of the learning analysis image acquired by the learning acquisition unit 11 and adds random noise of the learning noise image to the contrast-reduced learning analysis image to generate a learning noise-added image. The noise addition unit 12 may subtract a pixel value corresponding to the learning noise image from the pixel value of each pixel of the learning analysis image, multiply the pixel values ​​of each pixel after the subtraction by a predetermined coefficient less than 1, and add the subtracted pixel value to the pixel values ​​of each pixel after the multiplication to generate a learning analysis image with reduced contrast. The noise addition unit 12 may generate a plurality of learning analysis images with reduced contrast using a plurality of mutually different coefficients to generate a plurality of learning noise-added images.

[0056] The contrast of an image is defined, for example, as the difference between the maximum and minimum pixel values ​​of pixels contained in the image. The contrast-reduced learning analysis image and the noise-added learning image generated by the noise addition unit 12 are used for machine learning training to generate an inference model. Reducing the contrast of the learning analysis image reduces flickering caused by random noise contained in the analysis image. Although the random noise contained in the analysis image cannot be reduced to zero, flickering caused by random noise can be reduced to close to zero. In this embodiment, the contrast-reduced learning analysis image is treated as an image that does not contain random noise. The noise-added learning image is an image that contains random noise corresponding to the contrast-reduced learning analysis image (an image that does not contain random noise).

[0057] In this embodiment, the inference model is trained so that when a noise-added learning image (an image containing random noise) is input to the inference model, an analysis image for learning (an image containing no random noise) with reduced contrast is output. The noise adding unit 12 generates the analysis image for learning with reduced contrast and the noise-added learning image, for example, as follows:

[0058] The noise addition unit 12 receives the learning analysis image 42 and the learning noise image 52 from the learning acquisition unit 11. Fig. 4 shows a graph for explaining the generation of the learning analysis image with reduced contrast and the learning noise-added image. The graph shown in Fig. 4 corresponds to the image. In the graph shown in Fig. 4, the horizontal axis indicates each pixel, and the vertical axis indicates the pixel value of each pixel.

[0059] Graph G1 shown in Fig. 4 is a graph of the analysis image 42 used for learning. The peak portions in graph G1 are portions corresponding to the target signal related to the semiconductor device. If the analysis image is an luminescence image, these portions are the luminescent parts. The portions other than the peak portions in graph G1 are background portions that do not correspond to the target signal related to the semiconductor device. All portions in graph G1 contain random noise.

[0060] Graph G2 shown in Fig. 4 is a graph of the training noise image 52. Graph G2 does not have any peaks, and all of the graph is background that does not correspond to the target signal related to the semiconductor device. All of the graph G2 contains random noise.

[0061] The noise adding unit 12 calculates the average of the pixel values ​​of the learning noise image 52. If the average of the pixel values ​​of the learning noise image 52 has been calculated by the learning acquisition unit 11, the noise adding unit 12 may use that calculation. This average of the pixel values ​​corresponds to the average of the pixel values ​​of the background in the learning noise image 52.

[0062] As shown in Fig. 3, the noise addition unit 12 subtracts the calculated average pixel value (background average) from the pixel value of each pixel in the analysis image 42 for learning. Graph G3 shown in Fig. 4 is a graph of the image obtained in this manner. As shown in graph G3 in Fig. 4, this image is a graph in which the average pixel value of the background portion is zero. Note that in this image, the average pixel value of the background portion does not need to be strictly zero, as long as the pixel value of the background portion is roughly close to zero.

[0063] Next, as shown in FIG. 3 , the noise adding unit 12 multiplies the pixel value of each pixel in the analysis image 42 for learning, from which the background average has been subtracted, by a brightness multiplier, which is a preset coefficient less than 1. Graph G4 in FIG. 4 is a graph of the image obtained in this manner. Multiplying by the brightness multiplier can reduce the contrast of the image. As a result, noise in the background portion of the image can be reduced. The brightness multiplier may be randomly selected, for example, from a range of 0.01 (lower limit) to less than 1 (upper limit) (excluding 1). The brightness multiplier may also be preset so that the variation in pixel values ​​in the background portion of the obtained image is equal to or less than a certain value. For example, the brightness multiplier may be set so that the standard deviation (σ value) of the pixel values ​​in the background portion is equal to or less than 1.

[0064] Furthermore, the coefficient (3 in the above example) by which the average pixel value is multiplied by the learning acquisition unit 11 to determine whether or not to use the extracted partial image as the learning analysis image 42 may be a value corresponding to the luminance magnification. For example, the coefficient may be a value that prevents the target signal from being buried in noise even when the lower limit of the luminance magnification is used. From this perspective, the luminance magnification may be a value corresponding to the coefficient.

[0065] Next, as shown in FIG. 3 , the noise adding unit 12 adds the background average subtracted above to the pixel value of each pixel in the image obtained by multiplying the pixel value by the brightness magnification factor, i.e., the image with reduced contrast. The image obtained in this way is a learning analysis image with reduced contrast. Because the analysis image of a semiconductor device contains noise other than random noise, adding the background average to the contrast-reduced image makes it possible to obtain a learning analysis image in which only the flicker caused by random noise is reduced from the actual analysis image. Graph G5 shown in FIG. 4 is a graph of the image obtained in this way (learning analysis image with reduced contrast). The learning analysis image with reduced contrast is an image in which the contrast of the analysis image is reduced without changing the background average pixel value (based on the background average pixel value). If the analysis image is an luminescence image, the image is an luminescence image with reduced contrast.

[0066] The analysis image for learning with reduced contrast has smaller pixel values ​​in the area corresponding to the target signal, but as described above, flickering due to random noise in the background area is reduced. Therefore, as described above, the analysis image for learning with reduced contrast can be treated as an image that does not contain random noise.

[0067] The noise addition unit 12 may use a plurality of brightness magnifications to generate a plurality of analysis images for training, in which the contrast is reduced by different rates, for one analysis image for training 42. This increases the variety of images used for training the machine learning of the inference model.

[0068] Furthermore, as shown in Fig. 3 , the noise adding unit 12 subtracts the calculated average pixel value (background average) of the training noise image 52 from the pixel value of each pixel of the training noise image 52. Graph G6 shown in Fig. 4 is a graph of the image obtained in this manner. As shown in graph G6 in Fig. 4 , this image is a graph in which the average pixel value is zero. Note that in this image, the average pixel value does not need to be strictly zero, as long as the pixel value is roughly close to zero. The image obtained in this manner is a zero-average noise image.

[0069] Next, the noise addition unit 12 adds, for each pixel at the same position, the pixel value of each pixel in the analysis image for learning with reduced contrast (e.g., a luminescence image with reduced contrast) and the pixel value of each pixel in the zero-mean noise image. The image obtained in this way is the noise-added image for learning. Graph G7 shown in FIG. 4 is a superposition of graph G5 of the analysis image for learning with reduced contrast and graph G6 of the zero-mean noise image.

[0070] As described above, by adding a zero-mean noise image to a learning analysis image with reduced contrast, it is possible to add pixel values ​​of only the random noise portion of the learning noise image, excluding pixel values ​​when there is no random noise in the learning analysis image with reduced contrast.

[0071] The generated noise-added learning image may be generated by adding random noise of the noise learning image to the contrast-reduced analysis image for learning, and may be any image other than the above as long as it can be used for training the machine learning of an inference model. Furthermore, noise addition unit 12 may generate the noise-added learning image from analysis image for learning 42 and noise learning image 52 by a method other than the above.

[0072] The noise addition unit 12 outputs to the training unit 13 a combination of the generated contrast-reduced analysis image for learning and a noise-added image for learning generated from the contrast-reduced analysis image for learning.

[0073] The training unit 13 is a training means that trains the machine learning of the inference model using the noise-added learning image generated by the noise adding unit 12. The training unit 13 may train the machine learning of the inference model using the contrast-reduced analysis image for learning as an image that does not contain random noise and the noise-added image for learning as an image that contains random noise.

[0074] The training unit 13 performs machine learning training of the inference model, for example, as follows: The training unit 13 receives a combination of a contrast-reduced analysis image for training and a noise-added image for training. The training unit 13 uses this combination to train the inference model.

[0075] The training unit 13 trains the inference model for machine learning by using a contrast-reduced analysis image for learning that does not contain random noise and a noise-added image for learning that contains random noise. For example, if the inference model receives an analysis image containing random noise and outputs an image after flicker caused by the random noise has been reduced, the training unit 13 trains the inference model so that the noise-added image for learning is received and the inference model outputs an analysis image for learning that has reduced contrast. For example, if the analysis image is an luminescent image, as shown in FIG. 5 , when a composite luminescent image (noise-added image for learning) with altered luminescent intensity, which is an luminescent image with reduced contrast plus zero-mean noise, is input to the inference model, the inference model is trained so that the inference model outputs an luminescent image with altered luminescent intensity (analysis image for learning with reduced contrast). In other words, the inference model is trained so that the random noise component added to the analysis image for learning with reduced contrast is eliminated. The training unit 13 trains the inference model using a conventional method (e.g., Noise2Clean).

[0076] In this case, for example, the training unit 13 inputs a noise-added image for training into the inference model, and obtains an image in which flicker caused by random noise has been reduced by the inference model at that time as an output from the inference model. The training unit 13 compares the output from the inference model with the analysis image for training with reduced contrast, and updates the parameters of the inference model by backpropagation based on a loss obtained from the comparison.

[0077] Furthermore, if the inference model is configured to input an analysis image containing random noise and output an image of random noise, the training unit 13 trains the inference model so that it inputs a noise-added image for learning and outputs a noise image for learning corresponding to the input image.

[0078] The training unit 13 repeatedly trains the inference model using multiple combinations of contrast-reduced analysis images for learning and noise-added images for learning. The training of the inference model is repeated, similar to conventional machine learning training, for example, a preset number of times or until the generation of the inference model converges based on preset conditions. When the repetition of the training of the inference model is completed, the training unit 13 outputs the generated inference model to the noise reduction system 20. The above are the functions of the learning system 10.

[0079] As shown in FIG. 1, the noise reduction system 20 includes a noise reduction acquisition unit 21 and an image generation unit 22 .

[0080] The reduction acquisition unit 21 is a reduction acquisition means for acquiring an analysis image for reduction that includes a target signal related to a semiconductor device and also includes random noise. The reduction acquisition unit 21 may also acquire a pattern image of the semiconductor device that corresponds to the analysis image for reduction.

[0081] The reduction acquisition unit 21 acquires an image, for example, as follows: The reduction acquisition unit 21 acquires the reduction analysis image by receiving it from an imaging device that captured the reduction analysis image, or by accepting a user's operation to input an image to the noise reduction system 20. If the size of the reduction analysis image to be input to the inference model is smaller than the original image that is cut out from the original image obtained by capturing an image with an imaging device, similar to the above-described learning analysis image, the reduction acquisition unit 21 may acquire the original image, divide the original image into pieces the size of the analysis image to be input to the inference model, and use the divided image as the reduction analysis image.

[0082] The reduction acquisition unit 21 may acquire the analysis image for reduction by a method other than the above. The reduction acquisition unit 21 outputs the acquired analysis image for reduction to the image generation unit 22.

[0083] The reduction acquisition unit 21 may also acquire a pattern image of the semiconductor device corresponding to the reduction analysis image. The pattern image is an image obtained by capturing a pattern of a semiconductor device. The pattern image is an image that has conventionally been used in failure analysis of semiconductor devices together with an emission image. The reduction acquisition unit 21 acquires, for example, the reduction analysis image or a pattern image corresponding to the original image (an image in which pixels at the same positions indicate the same position on the semiconductor device). The reduction acquisition unit 21 may acquire the pattern image in the same manner as the reduction analysis image. The reduction acquisition unit 21 outputs the acquired pattern image of the semiconductor device to the image generation unit 22.

[0084] The image generation unit 22 is an image generation means that uses an inference model generated in advance by machine learning training to generate an inference image in which flicker due to random noise is reduced from the reduction analysis image acquired by the reduction acquisition unit 21. The inference model used by the image generation unit 22 may be an inference model generated by the learning system 10. The image generation unit 22 may generate a composite image by superimposing the generated inference image and the pattern image acquired by the reduction acquisition unit 21.

[0085] The image generation unit 22 generates an image, for example, as follows: The image generation unit 22 inputs and stores the inference model generated by the learning system 10 and uses it to generate an inference image in which flickering due to random noise is reduced. The image generation unit 22 may also store an inference model other than that generated by the learning system 10 and use it to generate an inference image in which flickering due to random noise is reduced.

[0086] The image generation unit 22 inputs the reduction analysis image from the reduction acquisition unit 21. The image generation unit 22 inputs the input reduction analysis image to an inference model to generate an inference image after flickering due to random noise has been reduced. For example, if the inference model outputs an inference image after flickering due to random noise has been reduced, the image generation unit 22 acquires the output from the inference model as the inference image after flickering due to random noise has been reduced. Alternatively, if the inference model outputs an image of random noise, the image generation unit 22 removes the random noise indicated by the random noise image obtained using the inference model from the reduction analysis image to obtain the inference image after flickering due to random noise has been reduced. Furthermore, the image generation unit 22 may generate an inference image after flickering due to random noise has been reduced from the reduction analysis image by a method other than the above, as long as it uses an inference model.

[0087] When a pattern image of the semiconductor device is acquired by the reduction acquisition unit 21, the image generation unit 22 inputs the pattern image of the semiconductor device from the reduction acquisition unit 21. The image generation unit 22 generates a composite image (superimposed image) by superimposing the inference image after flicker reduction due to the generated random noise on the pattern image. The generation of the composite image may be performed in the same manner as conventional.

[0088] The image generated by the image generation unit 22 (the inference image or the composite image after the flicker caused by random noise has been reduced) is used for the failure analysis of the semiconductor device. The image generation unit 22 outputs the generated image. The generated image may be output in the same manner as in the conventional method depending on the failure analysis of the semiconductor device being performed. In this case, it may be possible to selectively display the analysis image for reduction (the analysis image before the flicker caused by random noise has been reduced) and the inference image after the flicker caused by random noise has been reduced. The above are the functions of the noise reduction system 20.

[0089] Next, the processing executed by the learning system 10 and the noise reduction system 20 according to this embodiment (the operation method performed by the learning system 10 and the noise reduction system 20) will be described using the flowcharts of FIGS.

[0090] First, a learning method, which is a process executed by the learning system 10 according to this embodiment, will be described using the flowchart of FIG. In this process, the learning acquisition unit 11 acquires a learning analysis image that includes a target signal related to a semiconductor device and also includes random noise, and a learning noise image that does not include the target signal and also includes random noise (S01, learning acquisition step). Next, the noise addition unit 12 reduces the contrast of the learning analysis image to generate a contrast-reduced learning analysis image (S02, noise addition step). Next, the noise addition unit 12 adds random noise from the learning noise image to the contrast-reduced learning analysis image to generate a noise-added learning image (S03, noise addition step).

[0091] Next, the training unit 13 uses the noise-added images for learning to train the machine learning of the inference model (S04, training step). Then, the inference model generated by the training is output from the learning system 10 to the noise reduction system 20 (S05). The noise reduction system 20 stores the inference model and uses it in the following noise reduction process. The above is the learning method, which is the process executed by the learning system 10 according to this embodiment.

[0092] Next, a noise reduction method, which is a process executed by the noise reduction system 20 according to this embodiment, will be described using the flowchart of FIG. 7 . In this process, the reduction acquisition unit 21 acquires a reduction analysis image that includes a target signal related to a semiconductor device and also includes random noise (S11, reduction acquisition step). Next, the image generation unit 22 uses an inference model to generate an inference image in which flicker due to random noise has been reduced from the reduction analysis image (S12, image generation step). The inference image in which flicker due to random noise has been reduced is output from the image generation unit 22 to a predetermined output destination (S13). When the reduction acquisition unit 21 acquires a pattern image of a semiconductor device corresponding to the reduction analysis image, the image generation unit 22 generates and outputs a composite image by superimposing the inference image and the pattern image. This concludes the noise reduction method, which is a process executed by the noise reduction system 20 according to this embodiment.

[0093] In the learning system 10 according to this embodiment, random noise from a noise image for learning is added to an analysis image for learning with reduced contrast, thereby generating a noise-added image for learning. An inference model generated using such a noise-added image for learning can appropriately reduce flickering due to random noise in the analysis image. Therefore, the learning system 10 according to this embodiment can appropriately reduce flickering due to random noise in the analysis image. Furthermore, in this embodiment, there is no need to prepare an analysis image that does not contain random noise in advance.

[0094] As in the present embodiment, machine learning training of the inference model may be performed using a contrast-reduced analysis image for learning that does not contain random noise and a noise-added image for learning that contains random noise. This configuration allows for more appropriate training of the machine learning of the inference model, thereby generating a more appropriate inference model. As a result, flicker caused by random noise can be more appropriately reduced from the analysis image. However, the machine learning training of the inference model does not necessarily need to be performed using the contrast-reduced analysis image for learning and the noise-added image for learning in the above manner; it is sufficient to use the noise-added image for learning.

[0095] As in the present embodiment, a contrast-reduced analysis image for learning may be generated by subtracting a pixel value corresponding to the noise image for learning from the pixel value of each pixel in the analysis image for learning, multiplying the pixel value of each pixel after subtraction by a preset coefficient (brightness multiplier) less than 1, and adding the subtracted pixel value to the pixel value of each pixel after multiplication. This configuration allows for the generation of a more appropriate analysis image for learning with reduced contrast. As a result, flicker due to random noise can be more appropriately reduced from the analysis image. However, the generation of the analysis image for learning with reduced contrast does not necessarily have to be performed as described above, and any method may be used.

[0096] As in the present embodiment, multiple contrast-reduced analysis images for learning may be generated using multiple coefficients (brightness multipliers) that are different from each other, thereby generating multiple noise-added images for learning. With this configuration, multiple contrast-reduced analysis images for learning can be generated from a single analysis image for learning, thereby increasing the variety of images used for machine learning training of an inference model. This allows for more appropriate machine learning training of the inference model and the generation of a more appropriate inference model. As a result, flickering due to random noise can be more appropriately reduced from the analysis image. However, it is not necessary to use multiple coefficients (brightness multipliers) that are different from each other when generating noise-added images for learning.

[0097] As in the present embodiment, an original image larger than the analysis image for learning may be acquired, including a portion that can serve as the analysis image for learning. A portion of the acquired original image may be cropped, and a value based on the pixel values ​​of the cropped image may be compared with a value corresponding to the pixel values ​​of the noise image for learning. The cropped image may be selected based on the comparison result. Specifically, a value based on the pixel values ​​of the cropped image may be compared with the average or variance value of the pixel values ​​of the noise image for learning. The image may be selected based on the comparison result. These configurations allow for easy and appropriate acquisition of the analysis image for learning. As a result, flickering due to random noise can be more appropriately reduced from the analysis image. However, acquisition of the analysis image for learning does not have to be performed as described above.

[0098] According to the noise reduction system 20 of this embodiment, flicker caused by random noise can be appropriately reduced from the analysis image.

[0099] As in the present embodiment, a pattern image of the semiconductor device corresponding to the analysis image for reduction may also be acquired, and a composite image may be generated by superimposing the generated inference image on the pattern image. With this configuration, a composite image can be obtained by superimposing the inference image with the pattern image, in which flicker due to random noise has been appropriately reduced.

[0100] Next, an example of this embodiment will be shown. FIG. 8 shows an example of flicker reduction due to random noise performed by this embodiment. FIG. 8(a) is an analysis image of a target for flicker reduction due to random noise, and FIG. 8(b) is an inference image of the analysis image in which flicker reduction due to random noise has been performed by this embodiment. The analysis image of the target for flicker reduction due to random noise shown in FIG. 8(a) was captured under a lens magnification of 5x. Also shown is the standard deviation (S.D.) of pixel values ​​in a portion R1 of each image in FIG. 8. The portion R1 is a background portion that does not correspond to a target signal related to the semiconductor device. Therefore, the smaller the standard deviation of pixel values ​​in the portion R1, the less flicker due to random noise there is.

[0101] FIG. 9 shows an example of random noise flicker reduction performed by this embodiment. FIG. 9(a) is an analysis image of the target of random noise flicker reduction, and FIG. 9(b) is an inference image of the analysis image after random noise flicker reduction has been performed on the analysis image using this embodiment. The analysis image of the target of random noise flicker reduction shown in FIG. 9(a) was captured using a lens magnification of 20x. Also shown is the standard deviation (S.D.) of pixel values ​​in portion R2 of each image in FIG. 9. The portion R2 is a background portion that does not correspond to the target signal related to the semiconductor device. Therefore, the smaller the standard deviation of pixel values ​​in portion R2, the less flicker caused by random noise. In both the examples of FIG. 8 and FIG. 9, the flicker caused by random noise in the analysis image was appropriately reduced.

[0102] 10 shows an example of a composite image (superimposed image) generated by this embodiment. FIG. 10(a) is a composite image obtained by superimposing an analysis image of a target for reducing flicker due to random noise and a pattern image. FIG. 10(b) is a composite image obtained by superimposing an inference image of the analysis image in which flicker due to random noise has been reduced by this embodiment and a pattern image. The analysis image of a target for reducing flicker due to random noise used in the composite image of FIG. 10(a) was captured at a semiconductor device drive frequency of 1.2 MHz.

[0103] 10(c) and 10(d) show enlarged images of a portion R3 in each of the images in FIGS. 10(a) and 10(b). The portion R2 corresponds to the target signal related to the semiconductor device. In the example of FIG. 10, flicker due to random noise in the analysis image was also appropriately reduced. In the example of FIG. 10, when flicker due to random noise is reduced (denoised) (FIGS. 10(b) and 10(d)), signals with weak luminescence intensity can be confirmed compared to when flicker due to random noise is not reduced (denoised) (FIGS. 10(a) and 10(c)).

[0104] Next, we will explain a learning program and a noise reduction program for executing the above-mentioned series of processes by the learning system 10 and the noise reduction system 20. As shown in Fig. 11(a), the learning program 100 is stored in a program storage area 111 formed on a computer-readable recording medium 110 that is inserted into a computer and accessed, or that is provided in the computer. The recording medium 110 may be a non-transitory recording medium.

[0105] The learning program 100 is configured to include a learning acquisition module 101, a noise addition module 102, and a training module 103. Functions realized by executing the learning acquisition module 101, the noise addition module 102, and the training module 103 are similar to the functions of the learning acquisition unit 11, the noise addition unit 12, and the training unit 13 of the learning system 10 described above, respectively.

[0106] 11(b), the noise reduction program 200 is inserted into a computer and accessed, or is stored in a program storage area 211 formed on a computer-readable recording medium 210 provided in the computer. The recording medium 210 may be a non-transitory recording medium. Note that when the learning program 100 and the noise reduction program 200 are executed on the same computer, the recording medium 210 may be the same as the recording medium 110.

[0107] The noise reduction program 200 is configured to include a noise reduction acquisition module 201 and an image generation module 202. Functions realized by executing the noise reduction acquisition module 201 and the image generation module 202 are similar to the functions of the noise reduction acquisition unit 21 and the image generation unit 22 of the noise reduction system 20 described above, respectively.

[0108] The learning program 100 and the noise reduction program 200 may be configured so that part or all of them are transmitted via a transmission medium such as a communication line, and are received and recorded (including installed) by another device. Furthermore, each module of the learning program 100 and the noise reduction program 200 may be installed not on one computer but on one of multiple computers. In this case, the above-described series of processes are performed by a computer system consisting of the multiple computers.

[0109] The learning method, inference model, noise reduction method, noise reduction system, and noise reduction program disclosed herein have the following configuration: [1] A learning method for generating an inference model used to reduce flicker caused by random noise from an analysis image used in failure analysis of a semiconductor device, the learning method including: a learning acquisition step for acquiring a learning analysis image that includes a target signal related to the semiconductor device and also includes random noise, and a learning noise image that does not include the target signal and also includes random noise, a noise addition step for reducing the contrast of the learning analysis image acquired in the learning acquisition step and adding random noise from the learning noise image to the contrast-reduced learning analysis image to generate a noise-added learning image, and a training step for training machine learning of the inference model using the noise-added learning image generated in the noise addition step. [2] The learning method described in [1], wherein in the training step, the contrast-reduced learning analysis image is used as an image that does not include random noise, and the noise-added learning image is used as an image that includes random noise, and the noise-added learning image is used as an image that includes random noise. [3] The learning method according to [1] or [2], wherein in the noise addition step, a pixel value corresponding to the learning noise image is subtracted from the pixel value of each pixel of the learning analysis image, the pixel value of each pixel after subtraction is multiplied by a preset coefficient less than 1, and the subtracted pixel value is added to the pixel value of each pixel after multiplication to generate a learning analysis image with reduced contrast. [4] The learning method according to [3], wherein in the noise addition step, a plurality of learning analysis images with reduced contrast are generated using a plurality of the coefficients that are different from each other to generate a plurality of learning noise-added images. [5] The learning method according to any of [1] to [4], wherein in the learning acquisition step, an original image that is larger than the learning analysis image and includes a portion that can become the learning analysis image is acquired, a portion of the acquired original image is cut out, a value based on the pixel values ​​of the cut-out image is compared with a value based on the pixel values ​​of the learning noise image, and the cut-out image is used as the learning analysis image depending on the comparison result.[6] The learning method according to [5], in which, in the acquisition step for learning, a value based on pixel values ​​of the extracted image is compared with an average or variation value of pixel values ​​of the noise image for learning, and the extracted image is used as the analysis image for learning according to the comparison result. [7] An inference model for causing a computer to input information based on an analysis image used in failure analysis of a semiconductor device, perform calculations according to the input, and output information, the inference model being generated by the learning method according to any of [1] to [6]. [8] A noise reduction method for reducing flicker due to random noise from an analysis image used in failure analysis of a semiconductor device, the noise reduction method comprising: an acquisition step for reduction of an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise; and an image generation step for generating an inference image in which flicker due to random noise has been reduced from the analysis image for reduction acquired in the acquisition step for reduction, using an inference model generated in advance by machine learning training. [9] The noise reduction method according to [8], in which the inference model is an inference model generated by the learning method according to any of [1] to [6].

[10] The noise reduction method according to [8] or [9], wherein in the reduction acquisition step, a pattern image of the semiconductor device corresponding to the reduction analysis image is also acquired, and in the image generation step, a composite image is generated by superimposing the generated inference image and the pattern image acquired in the reduction acquisition step.

[11] A noise reduction system for reducing flickering due to random noise from an analysis image used in failure analysis of a semiconductor device, comprising: a reduction acquisition means for acquiring a reduction analysis image that includes a target signal related to the semiconductor device and also includes random noise; and an image generation means for generating an inference image in which flickering due to random noise has been reduced from the reduction analysis image acquired by the reduction acquisition means, using an inference model generated in advance by machine learning training.

[12] The noise reduction system according to

[11] , wherein the inference model is an inference model generated by the learning method according to any of [1] to [6].

[13] A noise reduction program for reducing flickering due to random noise from an analysis image used in failure analysis of a semiconductor device, the noise reduction program causing a computer to execute the following steps: a reduction acquisition process for acquiring an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise; and an image generation process for generating an inference image in which flickering due to random noise has been reduced from the analysis image for reduction acquired by the reduction acquisition process, using an inference model generated in advance by machine learning training.

[14] The noise reduction program according to

[13] , wherein the inference model is an inference model generated by the learning method according to any one of [1] to [6].

[0110] 10...Learning system, 11...Learning acquisition unit, 12...Noise addition unit, 13...Training unit, 20...Noise reduction system, 21...Reduction acquisition unit, 22...Image generation unit, 100...Learning program, 101...Learning acquisition module, 102...Noise addition module, 103...Training module, 110...Recording medium, 111...Program storage area, 200...Noise reduction program, 201...Reduction acquisition module, 202...Image generation module, 210...Recording medium, 211...Program storage area.

Claims

1. A learning method for generating an inference model used to reduce flicker caused by random noise from an analysis image used in semiconductor device failure analysis, the learning method comprising: a learning acquisition step for acquiring a learning analysis image that includes a target signal related to the semiconductor device and also includes random noise, and a learning noise image that does not include the target signal and also includes random noise; a noise addition step for reducing the contrast of the learning analysis image acquired in the learning acquisition step, and adding random noise from the learning noise image to the contrast-reduced learning analysis image, to generate a noise-added learning image; and a training step for training the machine learning of the inference model using the noise-added learning image generated in the noise addition step.

2. A learning method as described in claim 1, wherein in the training step, the contrast-reduced analysis image for learning is an image that does not contain random noise, and the noise-added image for learning is an image that contains random noise, and machine learning of the inference model is trained.

3. A learning method according to claim 1 or 2, wherein in the noise addition step, a pixel value corresponding to the noise image for learning is subtracted from the pixel value of each pixel of the analysis image for learning, the pixel value of each pixel after subtraction is multiplied by a preset coefficient less than 1, and the subtracted pixel value is added to the pixel value of each pixel after multiplication, thereby generating an analysis image for learning with reduced contrast.

4. A learning method as described in claim 3, wherein in the noise addition step, a plurality of analysis images for learning with reduced contrast are generated using a plurality of coefficients that are different from each other, thereby generating a plurality of noise-added images for learning.

5. A learning method as described in claim 1 or 2, wherein in the learning acquisition step, an original image larger in size than the analysis image for learning is acquired, the original image includes a portion that can become the analysis image for learning, a portion of the acquired original image is cut out, a value based on the pixel values ​​of the cut-out image is compared with a value corresponding to the pixel values ​​of the noise image for learning, and the image cut out according to the comparison result is used as the analysis image for learning.

6. A learning method as described in claim 5, wherein in the learning acquisition step, a value based on the pixel values ​​of the extracted image is compared with the average or variation value of the pixel values ​​of the noise image for learning, and the extracted image is used as the analysis image for learning according to the comparison result.

7. An inference model for inputting information based on an analysis image used in the failure analysis of semiconductor devices, causing a computer to perform calculations according to the input and output information, said inference model being generated by the learning method described in claim 1 or 2.

8. A noise reduction method for reducing flickering due to random noise from an analysis image used in failure analysis of a semiconductor device, comprising: a reduction acquisition step for acquiring an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise; and an image generation step for generating an inference image in which flickering due to random noise has been reduced from the analysis image for reduction acquired in the reduction acquisition step, using an inference model generated in advance by machine learning training.

9. A noise reduction method according to claim 8, wherein the inference model is an inference model generated by the learning method according to claim 1.

10. A noise reduction method as described in claim 8, wherein in the reduction acquisition step, a pattern image of the semiconductor device corresponding to the reduction analysis image is also acquired, and in the image generation step, a composite image is generated by superimposing the generated inference image and the pattern image acquired in the reduction acquisition step.

11. A noise reduction system that reduces flicker caused by random noise from an analysis image used in failure analysis of a semiconductor device, comprising: a reduction acquisition means that acquires an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise; and an image generation means that uses an inference model that is generated in advance by machine learning training to generate an inference image in which flicker caused by random noise has been reduced from the analysis image for reduction acquired by the reduction acquisition means.

12. A noise reduction system as described in claim 11, wherein the inference model is an inference model generated by the learning method described in claim 1.

13. A noise reduction program that reduces flicker caused by random noise from an analysis image used in semiconductor device failure analysis, the noise reduction program causing a computer to execute the following steps: a reduction acquisition process that acquires an analysis image for reduction that includes a target signal related to the semiconductor device and also includes random noise; and an image generation process that generates an inference image in which flicker caused by random noise has been reduced from the analysis image for reduction acquired by the reduction acquisition process, using an inference model that is generated in advance by machine learning training.

14. A noise reduction program as described in claim 13, wherein the inference model is an inference model generated by the learning method described in claim 1.

Images