Parameter optimization system, parameter optimization method, and storage medium

Abstract

In a case where a transformed image generated using a learning model for generating the transformed image is used for measurement, parameters of the learning model are optimized, so that dimensional measurement can be improved. Therefore, with reference to Figure 1, the present disclosure proposes a parameter optimization system that optimizes parameters for generating a transformed image from an input image, the parameter optimization system comprising: a storage device that holds a parameter optimization program; and a processor that reads in and executes a parameter optimization program from the storage device, the processor performing the processes of: calculating a transformation error between the transformed image and a training image input at the time of learning; calculating a dimensional measure of a pattern included in the training image based on the transformed image and the training image; calculating a composite error based on the transform error and the dimensional metric; and learning a machine learning model for generating the transformed image and optimizing parameters of the machine learning model such that the composite error is optimized.

Description

Technical Field

[0001] This disclosure relates to a parameter optimization system, a parameter optimization method, and a storage medium carrying a parameter optimization program for optimizing the parameters of a learning model that performs image transformation. Background Technology

[0002] In recent years, attempts have been made to utilize AI (Artificial Intelligence) technology to efficiently perform image measurements. For example, Patent Document 1 discloses a technique for generating high-quality images using a learning model based on an input image generated by an image generation tool to generate transformed images. Specifically, this learning model performs different learning on specific regions than on other regions in order to improve the generation quality of specific parts of the transformed image, such as edges of objects that are the objects being measured or inspected.

[0003] Existing technical documents

[0004] Patent documents

[0005] Patent Document 1: International Publication No. 2021 / 140620 Summary of the Invention

[0006] The problem that the invention aims to solve

[0007] However, in Patent Document 1, when learning a learning model for outputting a transformed image that becomes the object of measurement, only the transformation error of the transformed image and the training image input during learning is used. Therefore, size metric is not utilized. Here, size metric refers to an index value that includes the measurement accuracy calculated based on size values ​​and the throughput from generating the input image using an image generation tool to calculating the measurement accuracy of the transformed image. Therefore, even using the technique disclosed in Patent Document 1, it is sometimes impossible to generate a transformed image that sufficiently improves the size metric.

[0008] On the other hand, deep learning, which utilizes neural networks and the like, presupposes that the errors used in the learning process can be differentiated. Therefore, size metrics, which correspond to non-differentiable operations like size values, cannot be used as errors. Consequently, it is difficult to achieve the target value of the size metric desired by the user, making it challenging to generate a learning model suitable for image transformations in practical applications.

[0009] In view of this situation, this disclosure proposes a technique to optimize the parameters of the learning model in order to improve size measurement when the transformed image generated by the learning model for transform image generation is flexibly applied to measurement.

[0010] Methods for solving problems

[0011] To address the aforementioned issues, this disclosure proposes a parameter optimization system for generating transformed images from input images. This system comprises: a storage device holding a parameter optimization program; and a processor that reads from the storage device and executes the parameter optimization program, the processor performing the following processes: calculating a transformation error between the transformed image and a training image input during learning; calculating a size metric of a pattern contained in the training image based on the transformed image and the training image; calculating a composite error based on the transformation error and the size metric; learning a machine learning model for generating the transformed image; and optimizing the parameters of the machine learning model such that the composite error is optimized.

[0012] Further features relating to this disclosure will become clear from the description and drawings herein. Furthermore, the manner of this disclosure is achieved and implemented through elements and combinations thereof, as well as the detailed description thereafter and the appended scope of patent protection.

[0013] The descriptions in this specification are merely typical examples and do not limit the scope of patent protection or application of this disclosure in any sense.

[0014] Invention Effects

[0015] According to the technology disclosed herein, a learning model can be generated that produces a transformed image best suited to the size metric. Therefore, it is easy to generate a transformed image that meets the user's required size metric. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating a schematic structural example of the parameter optimization system 10 of the first embodiment.

[0017] Figure 2 It is a flowchart illustrating the details of parameter optimization processing (optimizing the generation of the learned model and the generation of size metrics).

[0018] Figure 3 It is used to illustrate parameter optimization processing ( Figure 2 A flowchart detailing the transformation parameter optimization process (step S204) in the process.

[0019] Figure 4 This is a diagram illustrating a structural example of an optimization condition setting GUI (Graphical User Interface) 301 that is displayed on the display screen of the output device 102 and used for inputting optimization conditions via the input device 101.

[0020] Figure 5 This is a diagram illustrating a structural example of the optimized result output GUI401 displayed on the output device 102.

[0021] Figure 6 This is a diagram illustrating the structure of GUI301', which represents the optimization condition setting of variant example 1.

[0022] Figure 7 This is a diagram illustrating the structure of the parameter optimization system 10 in variant example 1.

[0023] Figure 8 This is a diagram illustrating a schematic example of the structure of the parameter optimization system 10 of the learning model in the second embodiment.

[0024] Figure 9 This is a diagram illustrating a schematic structural example of the parameter optimization system 10 of the third embodiment.

[0025] Figure 10 This is a flowchart illustrating the details of the parameter optimization process in the third embodiment. Detailed Implementation

[0026] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. In the drawings, functionally identical elements are sometimes shown with the same numbers. Furthermore, the drawings illustrate specific embodiments that follow the principles of the present disclosure, but they are for the purpose of understanding the present disclosure and are in no way intended to limit the interpretation of the present disclosure.

[0027] In this embodiment, the description has been provided in sufficient detail by those skilled in the art for the purpose of implementing this disclosure. However, other installations and methods may also be used. It should be understood that structural and constructional changes and various element substitutions can be made without departing from the scope and spirit of the technical concept of this disclosure. Therefore, the following description is not intended to be limited to this.

[0028] Furthermore, in this embodiment, the measurement of circuit patterns in semiconductor images is described, focusing on the measurement of the dimensions of the circuit patterns. However, the technology disclosed herein is not limited to the measurement of circuit patterns in semiconductors and can be widely and universally applied to image measurement.

[0029] (1) First implementation method

[0030] The first embodiment relates to a technique for improving throughput by generating a higher quality image than the input image through image transformation of the input image.

[0031] For example, in order to perform dimensional measurements using a scanning electron microscope, signals obtained from multiple scans of the same field of view are sometimes averaged to generate a high-quality image with a high signal-to-noise ratio (S / N ratio) for measurement. However, performing multiple scans requires a long acquisition time, thus reducing throughput. Therefore, in this embodiment, an image transformation is performed on a low-quality image generated from fewer scans than when generating a high-quality image (an image with quality sufficient for dimensional measurement with a predetermined accuracy), generating an image comparable to high-quality. Thus, it is unnecessary to generate a high-quality image acquired at a low speed, thereby improving throughput.

[0032] <Structure Example of a Parameter Optimization System>

[0033] Figure 1 The following is a schematic structural example of the parameter optimization system 10 according to the first embodiment. The parameter optimization system 10 includes, for example: a computing device 100, which computes parameters; an input device 101, which is used when a user (operator) inputs a target value for a size measurement, which will be described later; an output device 102, which outputs (displays) the optimization result; and a storage device 201, which stores a high-quality image (e.g., a SEM image) with image quality that can be measured with a predetermined accuracy and a low-quality image (e.g., a SEM image, which may not be unmeasurable) with image quality worse than the high-quality image.

[0034] As a means of generating input images (high-quality and low-quality images stored in storage device 201) for optimizing the parameters of the learning model, a scanning electron microscope or the like can be used as the image generation tool. Furthermore, the input device 101 can be configured as a keyboard, mouse, or the like. Additionally, the output device 102 can be configured as a display or the like. It is configured to accept target values ​​(refer to...) of various dimensions input by the user from the input device 101. Figure 4 The optimized result is output to the output device 102.

[0035] Although Figure 1 Although not shown, the computing device 100 can be composed of a computer system, which includes one or more CPUs (Central Processing Units) and GPUs (Graphics Processing Units), as well as various memories (storage devices) composed of RAM (Random Access Memory), ROM (Read Only Memory), etc., and holding various programs, parameters, etc.

[0036] The arithmetic unit 100 uses a processor (CPU, etc.) to read the parameter optimization program stored in memory, expands the program in the processor's internal memory, and thereby realizes the parameter optimization function.

[0037] like Figure 1 As shown, the computing device 100 includes, for example, an image transformation unit 1001, a transformation error calculation unit 1002, a size measurement calculation unit 1003, a composite error calculation unit 1004, a transformation parameter update unit 1005, and a learning parameter update unit 1006, to perform the above-mentioned parameter optimization function.

[0038] The image transformation unit 1001 takes an image generated using a scanning electron microscope or an image (low-resolution image) stored in the storage device (storage medium) 201 as input image, and generates a transformed image based on the input image. The image transformation unit 1001 includes a learning model that adjusts the transformation parameters (weights, biases) and hyperparameters, and performs image transformation using this learning model. The learning model can be, for example, a neural network, a convolutional neural network, or a neural network with other structures, which are types of machine learning models.

[0039] The transformation parameter update unit 1005 uses the hyperparameters obtained from the learning parameter update unit 1006 to adjust the transformation parameters, and provides the adjusted transformation parameters to the image transformation unit 1001. Here, hyperparameters refer to parameters that determine the learning behavior of the learning model. As various parameters used in the operation (learning operation) used to determine the coefficients of the learning model, hyperparameters include, for example, the learning rate, batch size, construction of the learning model, and number of layers, which determine the update amount of the transformation parameters.

[0040] The learning parameter update unit 1006 adjusts the hyperparameters based on the composite error (described later) and provides the adjusted hyperparameters to the transformation parameter update unit 1005. Furthermore, as described later, the transformation parameters and hyperparameters are adjusted by repeatedly updating each parameter in a way that minimizes the index values ​​of errors (transformation error, composite error), so that the learning model outputs the desired result.

[0041] The transformation error calculation unit 1002 calculates the transformation error between the transformed image generated from the low-quality image by the image transformation unit 1001 and the high-quality image used as the training image. More specifically, the transformation error calculation unit 1002 calculates the average absolute error or mean square error, etc., calculated based on the pixel values ​​of the transformed image output by the learning model and the corresponding pixel values ​​of the high-quality image, as the transformation error. The average absolute error can be obtained by the operation formula shown in equation (1) below. Let x_i,j represent the pixel value at coordinate (i,j) of the transformed image, and let y_i,j represent the pixel value at coordinate (i,j) of the high-quality image. N represents the number of pixels in the transformed image or the high-quality image. The mean square error can be obtained by changing the absolute value of equation (1) to the square.

[0042] [Formula 1]

[0043]

[0044] The size measurement calculation unit 1003 calculates the size measurement of the transformed image generated by the image transformation unit 1001 from the low-quality image. The size measurement includes, for example, the measurement accuracy obtained by performing arbitrary calculations based on the size values ​​obtained from measuring the dimensions of the semiconductor circuit pattern, and the throughput from inputting the input image generated by a scanning electron microscope to calculating the measurement accuracy of the transformed image. Here, the size values ​​are obtained by calculating the distance between edges, roughness values, etc., using two or more edge coordinates (x_edge, y_edge) obtained based on the brightness gradient on the image using a predetermined measurement algorithm. The measurement algorithm includes, for example, measurement parameters that control the measurement method, such as smoothing processing and edge detection thresholds. On the other hand, since the processing time of the neural network varies depending on the structure and number of layers of the learning model adjusted by hyperparameters, the throughput is affected.

[0045] Measurement accuracy can be calculated using a statistical metric that represents the similarity between size values ​​obtained from the transformed image and the high-quality image, respectively. For example, this includes the average error, sensitivity, coefficient of determination, and reproducibility between the size values ​​obtained from the transformed image and the high-quality image. For instance, throughput is expressed as the number of images that can be processed per unit time from the input image (or the input image generated by an image generation tool such as a scanning electron microscope) to the calculation of the size value. Size metrics calculated through statistical processing are generally not related to the output of the neural network, just like size values, and therefore cannot be differentiated using the gradient descent method described later with respect to the transformation parameters. Therefore, it cannot be applied to deep learning systems that use neural networks. On the other hand, algorithms for optimizing hyperparameters do not have restrictions on whether they can be differentiated. Therefore, in this disclosure, size metrics are included only in the error (composite error) when optimizing hyperparameters.

[0046] Furthermore, generally speaking, the aforementioned error is a transformation error using the brightness value of the transformed image output by the learning model. However, when applying the generated transformed image to measurement, as exemplified by Patent Document 1, size measurement has not been used as an indicator in the past. Therefore, it is impossible to generate a transformed image that adequately improves the size measurement. In addition, for deep learning using neural networks or the like, there is a prerequisite that the error used in the learning process can be differentiated, making it difficult to use size measurement, which corresponds to operations that cannot be differentiated, as an error. Therefore, it is impossible to achieve the target value of the size measurement desired by the user, and it is difficult to generate a learning model suitable for image transformation in practical applications. Therefore, in this disclosure, by using the composite error described later for hyperparameter adjustment (optimization), it is possible to generate a learning model that generates a transformed image best suited to the size measurement.

[0047] The composite error calculation unit 1004 calculates the composite error by integrating the transformation error L_image calculated by the transformation error calculation unit 1002 and the size measurement L_measure calculated by the size measurement calculation unit 1003 through a weighted sum. For example, the composite error can be calculated based on an expression such as equation (2).

[0048]

[0049] Here, α and β are weighting coefficients represented by arbitrary real numbers. For example, when the weighting coefficient β of the size metric L_measure is set to a value larger than the weighting coefficient α of the transformation error L_image, the size metric becomes a larger error than the transformation error, which can prompt updates to hyperparameters that place greater emphasis on minimizing the size metric.

[0050] As an example, the measurement indicators included in the size metric include the average error between the size values ​​obtained from the transformed image and the size values ​​obtained from the high-quality image, sensitivity, coefficient of determination, reproducibility, etc. Furthermore, multiple of these optimization indicators can be included to constitute the size metric. Below, the calculation methods for each optimization indicator are shown for n sets of transformed images and high-quality images. Regarding the average error, it can be obtained by averaging the absolute error between the size value X1 obtained from the transformed image and the size value X2 obtained from the high-quality image across n sets. Regarding sensitivity, the n sets of X1 and X2 can be plotted on the x-axis and y-axis respectively, and represented by the slope of the approximate straight line. The coefficient of determination represents the degree of deviation of the n sets of X1 and X2 from the approximate straight line used to calculate the sensitivity, and can be obtained from the sum of squared deviations of the n sets of X1 and X2. Regarding reproducibility, it can be determined by setting it to three times (3σ) the standard deviation of the size values ​​when the action of repeatedly taking multiple images of the same field of view using a scanning electron microscope and averaging them to generate an image (images within the range of 3σ are considered reproducible).

[0051] When a size metric contains multiple optimization metrics, for example, the optimization metrics can be weighted and summed as in equation (3) to form a size metric.

[0052]

[0053] Here, the coefficients γ, δ, ε, and ζ are weighting coefficients represented by arbitrary real numbers. For example, by setting the sensitivity-related coefficient δ to a value larger than that of the coefficients γ, ε, and ζ, the error of the sensitivity-related term can be strongly reflected in the optimization index, and the hyperparameters can be updated to place greater emphasis on sensitivity improvement.

[0054] Furthermore, in equation (3) above, a specific optimization index that causes the target value to be input through the input device 101 can also be configured to satisfy the target value in terms of size. For example, as in equation (4), it can be configured such that when the specific optimization index is lower than the target value, the residual and size measure with respect to the target value are added together.

[0055]

[0056] Here, f, g, h, and i represent arbitrary functions. Functions f and g are configured, for example, to output 0 when the value of their first independent variable (mean error, |1-sensitivity|) is below the target value, and to output the residual calculated by subtracting the target value from the first independent variable when the value is above the target value. Functions h and i are configured, for example, to output 0 when the value of their first independent variable (coefficient of determination, throughput) exceeds the target value, and to output the residual calculated by subtracting the first independent variable from the target value when the value is below the target value. Alternatively, one can raise the power or root of (first independent variable) - (target value) or (target value) - (first independent variable); or take a certain margin from the target value to determine whether it exceeds or falls below; or use constrained optimization, discarding (not using) the learning model and hyperparameters that become size measures above or below the target value.

[0057] These weighting coefficients γ, δ, ε, and ζ are parameters that can be adjusted (set) according to the user's desired priority of size metrics. The user can configure these parameters by inputting them into GUI301 (see reference). Figure 4 The priority coefficient input area 3013 contains the weighting coefficients, which allow the composite error to reflect the desired priority of improvement for the dimensional measure. For example, if the user wants to improve the coefficient of determination more strongly than other dimensional measures, by entering a larger value in the text box related to the coefficient of determination in the priority coefficient input area 3013 than in the text boxes related to other dimensional measures, the values ​​of the terms related to the coefficient of determination in equations (3) and (4) become relatively larger compared to the terms of other dimensional measures, and the parameters are optimized, resulting in a stronger improvement in the coefficient of determination. Thus, the user can control the improvement of each dimensional measure according to the desired priority.

[0058] The predetermined values ​​for each weighting coefficient can be set to 1, for example, and can be adjusted appropriately by the user. Alternatively, the predetermined values ​​can be determined based on preliminary experiments using pre-defined sample images, in a way that does not deviate from the user's perception (in a way that is not empirically uncommon). In addition, the size measurement may include not only the above-mentioned optimization indicators, but also other optimization indicators.

[0059] The learning parameter update unit 1006 adjusts (optimizes) the hyperparameters based on the composite error calculated by the composite error calculation unit 1004 to suppress the composite error, and provides this to the transformation parameter update unit 1005. For example, the learning parameter update unit 1006 optimizes the hyperparameters by using hyperparameters arbitrarily selected by the user from a set of possible hyperparameter values. Alternatively, the learning parameter update unit 1006 may use parameter search algorithms such as grid search, random search, or Bayesian optimization to successively select hyperparameters, searching for combinations of hyperparameters that minimize or maximize a predetermined optimization index, thereby optimizing the hyperparameters. The hyperparameters are set based on the methods described above. When learning ends and the termination condition for parameter optimization is not met, the hyperparameters are successively updated and set as parameters that determine the learning behavior when the learning model is learned again. By adjusting (optimizing) the hyperparameters, the performance of image transformation and the computation time of the image transformation model change. Therefore, by setting appropriate parameters for learning, size metrics can be improved. For example, when adjusting the number of layers in the learning model as a hyperparameter, a smaller number of layers results in shorter processing time but lower image quality. Conversely, a larger number of layers enables the generation of images faithful to the training images, but increases processing time. The transformation parameter update unit 1005 adjusts (optimizes) the transformation parameters based on the transformation error calculated by the transformation error calculation unit 1003 to suppress the transformation error, and provides the transformation parameters to the image transformation unit 1001. The transformation parameters are optimized, for example, by using a method called gradient descent to successively update the transformation parameters with respect to the assigned error. The transformation parameter update unit 1005 uses the hyperparameters obtained by the learning parameter update unit 1006 in the above calculation.

[0060] <Details of parameter optimization processing>

[0061] Figure 2 This is a flowchart illustrating the details of parameter optimization processing (optimizing the generation of the learned model and the generation of size metrics). The main actions in each step are the respective processing units (image transformation unit 1001, transformation parameter update unit 1005, learning parameter update unit 1006, etc.). The functions of these processing units are implemented by expanding the program in the processor's internal memory, so the actions of each step can be replaced by actions mainly based on the computing device 100.

[0062] (i) Step S201

[0063] The image conversion unit 1001 acquires a pair of low-quality images and high-quality images stored in the storage device (storage medium) 201. Alternatively, the image conversion unit 1001 may also acquire a pair of low-quality images and high-quality images captured using a scanning electron microscope.

[0064] (ii) Step S202

[0065] The learning parameter update unit 1006 receives user input from the input device 101. Figure 4 The target value for the size measurement input in the GUI screen shown.

[0066] (iii) Step S203

[0067] The learning parameter update unit 1006 sets the hyperparameters of the learning model using a predetermined method. As described above, methods for setting hyperparameters include, for example, allowing the user to arbitrarily select from a set of possible hyperparameter values, using a grid search, random search, or Bayesian optimization, or other parameter search algorithms.

[0068] (iv) Step S204

[0069] The transformation parameter update unit 1005 uses the hyperparameters set in step S203 to perform learning calculations, optimizing the transformation parameters of the learning model (generating a learned model). (See below for further details.) Figure 3 Describe the details of step S204.

[0070] (v) Step S205

[0071] The image transformation unit 1001 generates a transformed image by applying a low-quality image to the learning model learned in step S204.

[0072] (vi) Step S206

[0073] The transformation error calculation unit 1002 calculates the transformation error using a high-quality image paired with the low-quality image and the transformation image generated in step S205. Additionally, the size measurement calculation unit 1003 calculates the size measurement. The size measurement can be calculated, for example, by performing image processing (e.g., edge extraction processing, etc.) on the transformation image generated in step S205, using measurements of predetermined patterns (e.g., the width of a circuit pattern) in the image (multiple measurements obtained from multiple transformation images), and performing statistical processing on the throughput. Then, the composite error calculation unit 1004 calculates the composite error based on the transformation error and the size measurement (refer to equation (2)).

[0074] (vii) Step S207

[0075] The learning parameter update unit 1006 determines whether the optimization termination condition is met. Here, whether the optimization termination condition is met can be determined, for example, by whether the size metric (each index value obtained by statistical processing) calculated in step S206 meets the target value set by the user in step S202, or by whether a predetermined number of optimization attempts has been reached.

[0076] If the optimization conditions are not met in step S207 (for example, if the size metric does not meet the target value), the process proceeds to step S208. That is, the hyperparameter values ​​and size metric are fed back as the selection criteria for the hyperparameters. On the other hand, if the optimization termination conditions are met (for example, if the size metric does not meet the target value in step S208), the process proceeds to step S209.

[0077] (viii) Step S208

[0078] The parameter update unit 1006 updates the hyperparameters based on the composite error. After updating the hyperparameters, the process returns to step S203.

[0079] The learning parameter update unit 1006 can, in the same manner as described in step S203, use the hyperparameter arbitrarily selected by the user from the set of possible hyperparameter values ​​as the update parameter, or it can use a parameter search algorithm such as grid search, random search, or Bayesian optimization, or other search algorithms to determine the update parameter.

[0080] (ix) Step S209

[0081] The arithmetic unit 100 uses Figure 5 The GUI shown outputs the optimized learned model and size measurement to the output device 102.

[0082] Through the above-described processing, the parameters of the learning model can be optimized. On the other hand, even after the optimization processing of the parameter optimization system 10, sometimes the target value of the size measurement input by the user through the input device 101 is not met. In this case, for example, the user can change the generation conditions of the image generation tool (filtering, etc.) and perform the above-described parameter optimization processing again using an input image with better image quality (as a result, slightly sacrificing throughput). For example, in the case of an image generation tool that is a scanning electron microscope, the generation conditions such as the number of summed average images, filtering, and autofocus processing of low-quality images can be changed. As a result, by generating an input image with better image quality, a transformed image that is closer to a high-quality image can be generated, increasing the possibility of improving the size measurement. Furthermore, when the image quality of the input image is changed in this way (changed to high quality), when the measurement processing is actually performed by generating a transformed image using the learned model, a low-quality image with the changed image quality level is input to the image transformation unit 1001 to generate a transformed image of the measurement object.

[0083] In addition, in the first embodiment, the size measurement is obtained based on the size value obtained by measuring the size of the circuit pattern of the semiconductor. As the size value, for example, the offset of the center of gravity between the layers of the multilayer semiconductor, i.e., the superimposed measurement value, or the size value obtained by measuring any shape in other images can be used.

[0084] <Details of the transformation parameter optimization process (step S204)>

[0085] Figure 3 It is used to illustrate parameter optimization processing ( Figure 2 A flowchart detailing the transformation parameter optimization process (step S204) in the process.

[0086] (i) Step S2041

[0087] The transformation parameter update unit 1005 obtains the transformation parameters (initial values) set by the user. When setting the initial values ​​of the transformation parameters of the learning model, the user can, for example, set the initial value of He, the initial value of Xavier, or the initial value obtained by other initialization methods.

[0088] (ii) Step S2042

[0089] The image transformation unit 1001 inputs a low-quality image for learning into the learning model to generate a transformed image.

[0090] (iii) Step S2043

[0091] The transformation error calculation unit 1002 uses a high-quality image paired with the low-quality image used for learning and the transformed image generated in step S2042 to calculate the transformation error. Examples of transformation errors include mean absolute error (refer to formula (1)) and mean square error.

[0092] (iv) Step S2044

[0093] The transformation parameter update unit 1005 updates (adjusts) the transformation parameters to suppress the transformation error based on the transformation error calculated in step S2043 and the hyperparameters set in step S203. As described above, the transformation parameter update can, for example, use the gradient descent method.

[0094] (v) Step S2045

[0095] The transformation parameter update unit 1005 determines whether the learning termination condition is met. The learning termination condition can be set by whether the transformation error calculated in step S2043 has reached a predetermined number of learning iterations, or it can be determined by whether the change in transformation error (learning curve) during learning has updated to a minimum value within a predetermined number of iterations.

[0096] If the learning termination condition is met ("Yes" in step S2045), the learning process in S204 ends. On the other hand, if the learning termination condition is not met ("No" in step S2045), the process returns to step S2042.

[0097] <Example of GUI for setting optimization conditions>

[0098] Figure 4 This describes an example of the structure of the optimization condition setting GUI (Graphical User Interface) 301, which is displayed on the display screen of the output device 102 and used for inputting optimization conditions via the input device 101. Through this GUI screen, optimization conditions that serve as the optimization benchmark in the computing device 100 of the parameter optimization system 10 can be set.

[0099] The optimization condition setting GUI 301 consists of an optimization index selection area 3011, an optimization index input area 3012, and a priority coefficient input area 3013. Users can select the desired optimization index by checking the checkboxes for the desired size metric within the optimization index selection area 3011.

[0100] Furthermore, users can set the target value of the optimization index by inputting the desired target value into the optimization index input area 3012. Moreover, users can adjust the improvement intensity of the size metric by inputting a priority coefficient (for example, the larger the value, the higher the priority) in the priority coefficient input area 3013, which represents the priority of the size metric to be improved.

[0101] By sending the selected optimization index, the target values ​​of each optimization index, and the priority coefficients to the computing device 100, the computing device 100 can calculate the size measurement. Furthermore, Figure 4 The dimensional measurement shown is just one example and is not limited to... Figure 4 The indicators shown are (mean error, sensitivity, coefficient of determination). Additional parameters can also be added. Figure 4 Add external size measurements and corresponding checkboxes and text boxes.

[0102] <Example of GUI for outputting optimization results>

[0103] Figure 5 This is an example of the structure of the GUI401, which displays the optimization results in the output device 102. Through this GUI, the user can be prompted with the optimization results of the parameter optimization processing of the computing unit 100 of the parameter optimization system 10.

[0104] The optimization result output GUI 401 consists of a target achievement display area 4011, an optimal parameter display area 4012, and a parameter contribution rate display area 4013. The target achievement display area 4011 displays the target achievement degree of the optimal size metric output in the parameter optimization system 10, based on the optimization index and its target value input in the optimization index input area 3012. For example, a size metric indicating that the target value has been achieved is a target achievement rate of 100%. A size metric indicating that the target value has not been achieved represents the target achievement rate at that time. In the case of not achieving the target value, it means that even with parameter optimization, it is difficult to achieve the desired image quality improvement in the provided input image. In this case, the user can adjust the imaging parameters of the scanning electron microscope to obtain a lower-quality image with better image quality, and then perform parameter optimization again to try to achieve the target value.

[0105] The optimal parameter display area 4012 displays the hyperparameters optimized by the parameter optimization system 10. Additionally, the parameter contribution rate display area 4013 shows how much the parameters of the optimization index object input in the optimization index input area 3012 contribute to the improvement of the composite error. Thus, users can identify parameters that contribute to improving dimensional measurements, making it easier to determine whether to add or exclude parameters that are part of the parameter optimization object.

[0106] <Variation Example>

[0107] (Variation Example 1)

[0108] The parameter optimization system 10 of the first embodiment, in addition to the variation 1, Figure 1 In addition to the components shown, it also includes a loop for optimizing the measurement parameters. According to Variation 1, the user can optimize the parameters of the learning model without manually adjusting the measurement parameters so that they do not hinder the improvement of dimensional measurement.

[0109] (i) A variation of the GUI for setting optimization conditions

[0110] Figure 6 This illustrates a structural example of the optimization condition setting GUI301' for variant example 1. The optimization condition GUI301' is configured to allow selection of whether or not to include measurement parameters. For... Figure 4 The same constituent elements, in Figure 6 The same reference numerals are used in the figures, therefore repeated descriptions are omitted below.

[0111] The optimization condition setting GUI301' includes a measurement parameter selection area 5001 as an additional structural item. In the measurement parameter selection area 5001, the user can select the measurement parameters included in the updated object (measurement) by checking the checkboxes for the measurement parameters they want to set as the optimization object. Figure 6 In the optimization object, if the "Smoothing" checkbox is selected, the parameters for smoothing the transformed image during pattern measurement are included. Similarly, if the "Edge Threshold" checkbox is selected, the threshold used for edge detection in the transformed image during pattern measurement is included. Figure 6 The measurement parameters shown are only one example and are not limited to... Figure 6 The measurement parameters shown are (smoothing, edge threshold). Additional parameters can also be added. Figure 6 Other measurement parameters, and add corresponding checkboxes.

[0112] (ii) Structural example of parameter optimization system 10

[0113] Figure 7 This describes a structural example of the parameter optimization system 10 in this modified example 1. The parameter optimization system 10 in this modified example has a path for feeding back measurement parameters from the learning parameter update unit 1006 to the dimension measurement calculation unit 1003.

[0114] In addition to a portion of the hyperparameters, the parameter update unit 1006 also includes the measurement parameters used to calculate the size values ​​in the update objective (optimization objective).

[0115] In situations where the measurement algorithm cannot calculate the dimension value (or the dimension value is inappropriate) even when the measurement parameters are not included in the optimization object, including the measurement parameters in the optimization object (by checking the checkbox for the measurement parameters) makes hard-to-see parts clear (e.g., emphasizing edges), thus enabling the measurement of the dimension value. Including the measurement parameters in the optimization object is effective when the measurement algorithm is sometimes the main obstacle to improving dimension measurement. For example, to measure on an image where edges are hard to see due to noise, the image is sometimes smoothed during measurement. Because the value of the measurement parameter that determines the degree of smoothing is small, the edges obtained from the image are biased, sometimes making it impossible to calculate an appropriate dimension value. Due to the edge bias, the dimension value becomes large or cannot be measured, thus degrading measurement accuracy. Therefore, by optimizing the parameters to improve measurement accuracy, measurement precision can be improved, thereby enhancing dimension measurement.

[0116] (Variation Example 2)

[0117] According to Modification Example 2, the transformation parameter update unit 1005 (refer to...) Figure 1 Data expansion can also be performed on any one or more of the low-resolution and high-resolution images. This data expansion refers to image processing such as adding Gaussian noise, smoothing, or scaling up / down using random parameters, which improves the robustness of learning between the low-resolution and high-resolution images by increasing the variation between them. In this case, the learning parameter update unit 1006 can adjust at least one of the parameters related to this data expansion—namely, the type of image processing performed, the number of image processing operations, and the maximum strength of the random parameters of the image processing (e.g., the maximum standard deviation of the added noise in the case of Gaussian noise)—as part of the hyperparameters.

[0118] According to Variation Example 2, the dimensional measure shown in Equation (3) or Equation (4) above can be easily improved. Therefore, users can easily obtain a learning model that achieves the target value of the desired dimensional measure.

[0119] (Variation Example 3)

[0120] According to Modification Example 3, when the dimensional measurement is differentiable, the composite error can be used to adjust (optimize) the transformation parameters. Therefore, the transformation parameter update unit 1005 can use the composite error output by the composite error calculation unit 1004 to adjust the transformation parameters. At this time, in Figure 1 In this modified example, the parameter optimization system 10 replaces the path of sending the transformation error from the transformation error calculation unit 1002 to the transformation parameter update unit 1005 with the path of sending the composite error from the composite error calculation unit 1004 to the transformation parameter update unit 1005. Furthermore, in Figure 3In the process, the parameter optimization process (step S204) is derived from the parameter optimization process in the first embodiment (refer to...). Figure 2 The steps between steps S2043 and S2044 include the following steps S2046 and S2047.

[0121] (i) Step S2046

[0122] The size measurement calculation unit 1003 calculates the size measurement using a high-quality image paired with the low-quality image used for learning and the transformed image generated in step S2042. The size measurement calculation method can be the method described in step S206.

[0123] (ii) Step S2047

[0124] The composite error calculation unit 1004 uses the transformation error calculated in step S2043 and the size measurement calculated in step S2046 to calculate the composite error.

[0125] According to Variation 3, the transformation parameters can be adjusted to minimize the size metric without relying on hyperparameter tuning (optimization). Therefore, by improving the size metric shown in Equation (3) or Equation (4) above, the user can obtain a learning model that achieves the target value of the desired size metric. In Variation 3, hyperparameter tuning (optimization) can also be omitted, but by having hyperparameter tuning (optimization), superior hyperparameters can be used to optimize the transformation parameters, making it easier to achieve the target value of the desired size metric compared to optimizing only the transformation parameters.

[0126] <Technical Effects of the First Embodiment>

[0127] According to Patent Document 1, as described above, a method is presented for generating a transformed image aimed at improving the image quality of a measurement area or for directly predicting size values ​​through a learning model. On the other hand, according to the first embodiment, a transformed image can be generated based on a composite error comprising the image's transformation error and a user-desired size metric (through statistical processing). Furthermore, by constructing the size metric from user-desired elements, the user-desired size metric can be improved, and a transformed image with improved size metric can be generated.

[0128] (2) Second implementation method

[0129] In addition to the features of the first embodiment, the second embodiment also involves a process of including image generation parameters in an optimization object and feeding back the optimized image generation parameters to an image generation unit (e.g., a SEM device).

[0130] <Structure Example of a Parameter Optimization System>

[0131] Figure 8 This is a diagram illustrating a schematic structural example of the parameter optimization system 10 of the learning model in the second embodiment. For... Figure 1 The same constituent elements, in Figure 8 The same reference numerals are used in the figures, therefore repeated descriptions are omitted below.

[0132] In this parameter optimization system 10, the storage device 201 used to hold the SEM image is replaced (see reference). Figure 1 The system includes an image generation unit 7001 that generates images input to the learning model using a scanning electron microscope. Furthermore, the parameter optimization system 10 is configured to include the image generation parameters of the scanning electron microscope (image generation unit 7001) as the object of parameter optimization, and feeds back the image generation parameters from the learning parameter update unit 1006 to the image generation unit 7001. Alternatively, the image generation unit 7001 may be a storage device that pre-stores images generated using the scanning electron microscope with various image generation parameters and those image generation parameters, and outputs the stored images.

[0133] Furthermore, the parameter optimization system 10 in the second embodiment is a system that, in order to improve the composite error, controls the image generation parameters of the input image, controlling the image quality and image generation time of the low-quality image input to the learning model. This improves the image quality of the transformed image and expands the control range of the size metric. For example, when the image quality of the low-quality image input to the learning model is improved, it is easier to generate a transformed image close to a high-quality image, and improvements in image quality and measurement metrics can be expected. Additionally, the generation time of the input image also affects the throughput in order to control the image generation parameters. Therefore, by shortening the generation time of the input image, an increase in throughput can be expected.

[0134] <Examples of image generation parameters that become the object of optimization>

[0135] Examples of image generation parameters that can be optimized in the image generation unit 7001 include the sum-average number of frames for low-quality images, scanning speed (high-speed scanning, low-speed scanning), filtering processing (smoothing filtering, edge filtering), and autofocus processing (shooting based on a predetermined value without autofocus adjustment). For example, increasing the sum-average number of frames, setting the scanning speed to low, or applying filtering or autofocus processing can generate low-quality images with better image quality, but the image generation time becomes longer, thus reducing throughput. On the other hand, decreasing the sum-average number of frames, setting the scanning speed to high, or not applying filtering or autofocus processing can generate low-quality images with higher throughput, but the image quality of the low-quality images deteriorates, and the image quality and size of the transformed images also deteriorate.

[0136] In this way, the measurement metrics (measured metrics such as mean error, sensitivity, coefficient of determination, reproducibility, etc.), image quality, and throughput are in a trade-off relationship. When the target values ​​for both the measurement metrics and throughput are input through the input device 101, the computing device 100 in the parameter optimization system 10 performs optimization processing to output image generation parameters, transformation parameters, hyperparameters, and measurement parameters that are appropriately balanced so that each size metric (both measurement metrics and throughput) satisfies the target values.

[0137] Typically, there are a vast number of combinations of image generation parameters, transform parameters, hyperparameters, and measurement metrics to achieve the desired measurement indicators and throughput. Therefore, it is very difficult for users to pre-define their desired combinations (appropriate combinations). However, by performing parameter optimization processing using the parameter optimization system 10 of this embodiment, appropriate combinations of various parameters can be derived.

[0138] <Processing after updating image generation parameters>

[0139] When the image generation unit 7001 is composed of a scanning electron microscope, the scanning electron microscope acquires image generation parameters fed back from the learning parameter update unit 1006, and after changing the shooting conditions according to the image generation parameters, it takes another image. However, it is not limited to this; a computer (not shown) may pre-store a group of images taken under multiple shooting conditions in a storage device (not shown), and select an image group suitable for the fed-back image generation parameters to input into the learning model. Alternatively, the computer may instruct the scanning electron microscope to take another image if no image group suitable for the fed-back image generation parameters exists.

[0140] <Technical Effects of the Second Embodiment>

[0141] According to the second embodiment, users can optimize the parameters of the learning model without manually adjusting image generation parameters such as those used to improve size metrics. Furthermore, in the second embodiment, image generation parameters are also included in the parameter optimization objects, and therefore, these parameters can also be displayed in the parameter contribution rate display area 4013. Thus, as an additional effect of the parameter contribution rate display area 4013, users can see the image generation parameters that contribute to improving size metrics, making it easier to adjust these parameters.

[0142] In the second embodiment, a scanning electron microscope was described as an image generation tool, but it is not limited to this. For example, an optical microscope or other image generation tool that can control the image quality of the captured image through image generation parameters can also be used.

[0143] (3) Third implementation method

[0144] The third implementation involves processing the preprocessing / postprocessing parameters of a learning model that generates transformed images from low-resolution images using composite errors instead of hyperparameters.

[0145] <Structure Example of a Parameter Optimization System>

[0146] Figure 9 This is a diagram illustrating a schematic structural example of the parameter optimization system 10 according to the third embodiment. This parameter optimization system 10 is a system for optimizing preprocessing or post-processing parameters when inferring transformed images using a learned model. Figure 9 In the middle, to and Figure 8 The same structural elements are labeled with the same reference numerals in the drawings; therefore, repeated descriptions are omitted below. For example... Figure 9 As shown, in the parameter optimization system 10 of the third embodiment, instead of Figure 8 The transformation parameter update unit 100 and the learning parameter update unit 1006 shown are equipped with an inference parameter update unit 9001, which is used to update the post-processing and pre-processing parameters applied to the low-quality image or the inferred transformation image input to the learning model during inference.

[0147] <Post-processing and pre-processing content>

[0148] (i) Post-processing refers to image processing performed on the transformed image (processing performed after image transformation). Examples of such image processing include Gaussian noise addition processing, smoothing processing, and brightness normalization processing, as well as synthesis processing of a second image different from the transformed image, but are not limited to these. Furthermore, examples of a second image different from the transformed image include high-quality images obtained through simulation (e.g., images generated by a predetermined simulation based on a semiconductor circuit pattern design), noisy design diagrams, or low-quality images, but are not limited to these. Additionally, the second image can be input from the input device 101 to the image transformation unit 1001, or the image transformation unit 1001 can retrieve it from a storage device not shown.

[0149] The parameter update unit 9001 infers parameters related to the synthesis of the transformed image and a second image different from the transformed image. For example, in the case of synthesis by blending two images, the parameter of the blending ratio (image blending ratio) of the two images. This post-processing has the effect of changing the image quality of the transformed image (the amount of noise in the transformed image, the degree of smoothness of the transformed image, etc.). By changing the size value obtained from the transformed image through this post-processing, the size measurement obtained from the size value can be improved by optimizing the post-processing parameters.

[0150] (ii) Preprocessing refers to image processing performed on the low-quality image (processing performed before image transformation). Specifically, image processing performed on the low-quality image, similar to post-processing, can include operations such as Gaussian noise addition, smoothing, and brightness normalization, as well as compositing with a second image different from the low-quality image. This preprocessing has the effect of changing the image quality of the low-quality image. By performing this preprocessing, the image quality of the transformed image obtained from the preprocessed low-quality image changes, and the size value also changes. Therefore, by optimizing the preprocessing parameters, the size measurement obtained from the size value can be improved.

[0151] <Details of parameter optimization processing>

[0152] Figure 10 This is a flowchart illustrating the details of the parameter optimization process in the third embodiment. In the third embodiment, a learned model is used to infer the transformed image. Furthermore, for... Figure 2 The same process, in Figure 10 The same reference numerals are used in the figures, therefore repeated descriptions are omitted below.

[0153] In the third embodiment, the computing device 100 optimizes the parameters related to the image processing described above, namely the type and quantity of image processing performed, and the parameters required for image processing (in the case of image synthesis with a noisy image, the synthesis rate of the noisy image) as hyperparameters to suppress the composite error shown by the above equation (3) or equation (4). This parameter optimization processing, in addition to the parameter optimization processing of the first embodiment (refer to...), Figure 2 Except for step S204, the process consists of the same steps. However, step S203 "setting hyperparameters" is replaced by step S1001 "setting preprocessing / postprocessing parameters". Additionally, step 208 "updating hyperparameters" is replaced by step S1002 "updating preprocessing / postprocessing parameters". Hyperparameters are parameters that determine the learning behavior as described above, but preprocessing / postprocessing parameters are parameters used when inferring transformed images using a learning model that performs image transformation.

[0154] (i) Step S1001

[0155] The inference parameter update unit 9001 sets the preprocessing / postprocessing parameters (e.g., the image blending ratio parameter) of the learning model using a predetermined method. Methods for setting the preprocessing / postprocessing parameters include, for example, as described above, methods that allow the user to arbitrarily select from a set of acceptable preprocessing / postprocessing parameter values, methods that use grid search, random search, or parameter search algorithms such as Bayesian optimization, or other search algorithms.

[0156] (ii) Step S1002

[0157] The parameter update unit 9001 updates the preprocessing / postprocessing parameters (the parameters of the image blending ratio) based on the composite error (used as an optimization index). After updating the preprocessing / postprocessing parameters, the process returns to step S1001.

[0158] Similar to the method described in step S1001, the parameter update unit 9001 can set the parameter arbitrarily selected by the user from the set of possible preprocessing / postprocessing parameter values ​​as the update parameter, or it can use a grid search, random search, or Bayesian optimization parameter search algorithm or other search algorithm to determine the update parameter.

[0159] <Technical Effects of the Third Embodiment>

[0160] According to the third embodiment, the preprocessing / postprocessing parameters used in transformed image inference can be optimized. Furthermore, it is possible to obtain a transformed image with further improved size metric of the transformed image inferred through the learning model. Therefore, users can easily achieve the desired target value for size metric.

[0161] (4) Summary

[0162] (i) In this embodiment (first to third embodiments), the computing device (processor or computer) 100 executes a parameter optimization program to perform the following processes: processing to calculate the transformation error between the transformed image generated by applying the input image (low-quality image) to the machine learning model and the high-quality image (training image) input during the learning of the learning model; processing to calculate the size metric of the pattern contained in the training image based on the transformed image and the training image; processing to calculate the composite error based on the transformation error and the size metric; learning the machine learning model for generating the transformed image to optimize the composite error; and processing to optimize the parameters of the machine learning model. In the first embodiment, the hyperparameters defining the behavior of the machine learning and the transformation parameters of the machine learning model performing image transformation are the objects of parameter optimization processing. In the second embodiment, in addition to the hyperparameters defining the behavior of the machine learning and the transformation parameters of the machine learning model performing image transformation, the image generation parameters in the image generation unit (image generation tool) 7001 are the objects of parameter optimization processing. Furthermore, in the third embodiment, instead of the transformation parameters and hyperparameters of the first embodiment, the parameters of the preprocessing and postprocessing performed when inferring the transformed image are the objects of parameter optimization processing. This allows for the improvement of the size measurement desired by the user, and the generation of transformed images that improve that size measurement.

[0163] (ii) In this embodiment, the size metric (the size metric of the transformed image) includes: measurement accuracy calculated using a measurement index that statistically represents the similarity between size values ​​obtained from the transformed image and the training image, respectively; and throughput from the input of the input image (low-resolution image) to the calculation of the size value using the transformed image. The size metric is represented by a function of measurement accuracy and throughput as variables.

[0164] (iii) In the above parameter optimization process, the computing device 100 determines that the parameters have been optimized when the size metric (represented by equations (3) and (4)) satisfies the optimization termination condition based on the target value (user-set value) (for example, when the size metric reaches the target value, or when the predetermined number of optimization processes is reached). By repeatedly performing optimization processes, the size metric can be optimized with a high probability. In addition, if all the set indicators have not reached the target value even after repeatedly performing optimization processes a predetermined number of times, an input image with better image quality than the original image (low-quality image) can be re-input to perform optimization processes again. Furthermore, the computing device 100 can also output the result of the optimization process, indicating whether the size metric has reached the set target value, to the output device 102.

[0165] Furthermore, the computing device 100 alternately and repeatedly learns the machine learning model and optimizes the parameters (optimization of hyperparameters and transformation parameters, optimization of hyperparameters, transformation parameters and image generation parameters, or optimization of preprocessing / postprocessing parameters) until the composite error (refer to Equation (2)) is optimized. In addition, the computing device 100 uses a predetermined parameter search algorithm to adjust the hyperparameters to optimize the composite error. As shown in Equation (2), the composite error is represented by multiplying the transformation error and the size metric by weighting coefficients (α and β) respectively and then adding them together. By setting the weighting coefficient (β) of the size metric to be larger than the weighting coefficient (α) of the transformation error, the transformation parameters, hyperparameters, image generation parameters, and preprocessing / postprocessing parameters that prioritize minimizing the size metric can be updated.

[0166] (iv) In this embodiment, the type, number, or maximum intensity of image processing performed on at least one of the input image (low-resolution image) and training image (high-resolution image) can be included in the hyperparameters used to define the behavior of the machine learning. This allows for the optimization of these metrics, resulting in a significant improvement in size metrics. Consequently, the user can obtain a machine learning model that achieves the desired target value for size metrics.

[0167] (v) According to the second embodiment, the parameter optimization system 10 may include an image generation tool (e.g., a scanning electron microscope) that generates the input image (low-resolution image) instead of the storage device 201 that stores the image (SEM image). In this case, the computing device 100 optimizes the image generation parameters when the image generation tool generates the input image by optimizing the aforementioned composite error. Thus, a machine learning model that generates an appropriate transformed image while simultaneously generating a low-resolution image using the image generation tool in real time can be constructed.

[0168] (vi) According to the third embodiment, as described above, the computing device 100 optimizes parameters related to the inference of the transformed image using the machine learning model. Specifically, the computing device 100 optimizes parameters of at least one of preprocessing before generating the transformed image or postprocessing after generating the transformed image, as parameters related to the inference of the transformed image. Specifically, the preprocessing is image processing performed on the input image (low-quality image), including any one of Gaussian noise addition processing, smoothing processing, brightness normalization processing, and compositing the input image with an image different from the input image (a simulated image based on the design drawing). Furthermore, the postprocessing is image processing performed on the transformed image, including any one of Gaussian noise addition processing, smoothing processing, brightness normalization processing, and compositing the transformed image with an image different from the transformed image (a simulated image based on the design drawing). As a result, the size measurement of the transformed image inferred by the machine learning model is improved, thus allowing the user to easily achieve the desired target value for the size measurement.

[0170] (vii) The functions of each embodiment can also be implemented through software program code. In this case, a storage medium storing program code is provided to the system or device, and the computer (or CPU, MPU) of the system or device reads the program code stored in the storage medium. In this case, the program code read from the storage medium itself implements the functions of the above-described embodiments, and the program code itself and the storage medium storing the program code constitute this disclosure. Examples of storage media for supplying such program code include floppy disks, CD-ROMs, DVD-ROMs, hard disks, optical disks, optical discs, CD-Rs, magnetic tapes, non-volatile memory cards, ROMs, etc.

[0171] Alternatively, based on the instructions in the program code, the OS (operating system) running on the computer may perform some or all of the actual processing to achieve the functions of the above-described embodiments. Furthermore, after the program code read from the storage medium is written into the computer's memory, the computer's CPU (processor) or similar device may perform some or all of the actual processing based on the instructions in the program code to achieve the functions of the above-described embodiments.

[0172] Furthermore, the program code of the software used to implement the functions of each implementation can be distributed via a network and stored in a storage unit such as a hard disk or memory of the system or device, or a storage medium such as a CD-RW or CD-R. When in use, the computer (or CPU, MPU) of the system or device reads and executes the program code stored in the storage unit or storage medium.

[0173] The processes and techniques described herein are not inherently associated with any specific device and can be implemented through combinations of components. Furthermore, various types of general-purpose devices can be added. Dedicated devices can also be constructed to perform the functions of each embodiment. Additionally, various functions can be formed by appropriately combining the multiple constituent elements disclosed in each embodiment. For example, several constituent elements can be deleted from all the constituent elements shown in each embodiment, or constituent elements from different embodiments can be appropriately combined.

[0174] Specific embodiments are described in this disclosure, but these embodiments are for illustrative purposes only and not for limiting the understanding of the technology disclosed herein. Those skilled in the art will recognize that many combinations of hardware, software, and firmware are available to implement the technology disclosed herein. For example, the described software can be installed using a wide range of program or scripting languages ​​such as assembler, C / C++, Perl, Shell, PHP, and Java (registered trademark).

[0175] Furthermore, in the above embodiments, control lines and information lines refer to the lines deemed necessary for the description, but may not represent all control lines and information lines on the product. All structures can also be interconnected.

[0176] Furthermore, anyone with ordinary knowledge in this art can identify other embodiments of this disclosure based on an examination of each embodiment. The specification and specific examples are merely typical examples, and the technical scope and spirit of this disclosure are shown within the scope of patent protection.

[0177] Explanation of reference numerals in the attached figures

[0178] 10-parameter optimization system

[0179] 100 computing devices

[0180] 101 Input Device

[0181] 102 Output Device

[0182] 201 storage device

[0183] 1001 Image Transformation Unit

[0184] 1002 Transformation Error Calculation Unit

[0185] 1003 Dimensional Measurement Calculation Department

[0186] 1004 Composite Error Calculation Department

[0187] 1005 Transformation Parameter Update Department

[0188] 1006 Learning Parameter Update Department

[0189] 7001 Image Generation Unit

[0190] 9001 Inference Parameter Update Unit.

Claims

1. A parameter optimization system for optimizing parameters used to generate a transformed image from an input image, characterized in that, The parameter optimization system has the following features: Storage devices, their parameter optimization procedures; and The processor reads from the storage device and executes the parameter optimization program. The processor performs the following processing: Calculate the transformation error between the transformed image and the training image input during learning; Based on the transformed image and the training image, calculate the size metric of the pattern contained in the training image; The composite error is calculated based on the transformation error and the size metric. as well as Learn a machine learning model for generating the transformed image and optimize the parameters of the machine learning model to optimize the composite error.

2. The parameter optimization system according to claim 1, characterized in that, The processor includes the measurement accuracy and throughput calculated using measurement metrics in the size metric, which statistically represents the similarity between size values ​​obtained from the transformed image and the training image, respectively, and the throughput is the throughput from the input of the input image to the calculation of the size value using the transformed image.

3. The parameter optimization system according to claim 1, characterized in that, If the size metric satisfies the optimization termination condition based on the set target value, the processor determines that the parameter has been optimized.

4. The parameter optimization system according to claim 1, characterized in that, The processor alternately and repeatedly learns the machine learning model and optimizes the parameters until the composite error is optimized.

5. The parameter optimization system according to claim 4, characterized in that, The processor uses a predetermined parameter search algorithm to adjust the hyperparameters that define the learning behavior of the machine learning model, thereby optimizing the composite error.

6. The parameter optimization system according to claim 1, characterized in that, The processor uses gradient descent to learn the machine learning model.

7. The parameter optimization system according to claim 1, characterized in that, The machine learning model includes at least one model that uses a neural network.

8. The parameter optimization system according to claim 1, characterized in that, The processor will specify the hyperparameters of the learning behavior of the machine learning model and optimize the parameters of the size metric.

9. The parameter optimization system according to claim 8, characterized in that, The processor includes the type, number, or maximum intensity of image processing performed on at least one of the input image and the training image in the hyperparameters and optimizes the hyperparameters.

10. The parameter optimization system according to claim 1, characterized in that, The composite error is expressed by the following formula: Composite error = α × L_image + β × L_measure Where L_image represents the transformation error, L_measure represents the size metric, and α and β represent the weighting coefficients.

11. The parameter optimization system according to claim 2, characterized in that, The dimensional measure is expressed by the following formula: Size measurement = γ × L_bias + δ × |1-S| + ε × (1-R) ​​- ζ × T Where L_bias represents the average error of the size value, S represents sensitivity, R represents the coefficient of determination, T represents throughput, and γ, δ, ε and ζ represent weighting coefficients.

12. The parameter optimization system according to claim 3, characterized in that, The processor output indicates the result of the optimized processing and whether the size metric has achieved the target value.

13. The parameter optimization system according to claim 1, characterized in that, The processor will specify the optimization of the parameters of the size metric.

14. The parameter optimization system according to claim 1, characterized in that, The parameter optimization system also includes an image generation tool for generating the input image. The processor optimizes the image generation parameters when the image generation tool generates the input image.

15. The parameter optimization system according to claim 1, characterized in that, The parameter optimization system also includes an image generation tool for generating the input image. The throughput of the processor from generating the input image from the image generation tool to calculating the measurement accuracy for the transformed image is included in the size metric.

16. The parameter optimization system according to claim 1, characterized in that, The processor optimizes the parameters related to the inference of the transformed image using the machine learning model.

17. The parameter optimization system according to claim 16, characterized in that, The processor sets parameters of at least one of the preprocessing before generating the transformed image and the postprocessing after generating the transformed image as the optimization target, thereby serving as parameters related to the inference of the transformed image.

18. The parameter optimization processing according to claim 17, characterized in that, The preprocessing refers to image processing performed on the input image, including any one of the following: Gaussian noise addition processing, smoothing processing, brightness normalization processing, and compositing the input image with an image different from the input image. The post-processing is image processing performed on the transformed image, including any one of the following: Gaussian noise addition processing, smoothing processing, brightness normalization processing, and processing of combining the transformed image with an image different from the transformed image.

19. A parameter optimization method for generating a transformed image from an input image, characterized in that, The parameter optimization method includes: The processor reads from the storage device and executes the parameter optimization program to calculate the transformation error between the transformed image and the training image input during learning; The processor calculates the size metric of the pattern contained in the training image based on the transformed image and the training image; The processor calculates the composite error based on the transformation error and the size metric; as well as The processor learns a machine learning model for generating the transformed image and optimizes the parameters of the machine learning model to optimize the composite error.

20. A storage medium carrying a program for causing a computer to perform parameter optimization processing, said parameter optimization processing being used to generate a transformed image based on an input image. Its features are, The program causes the computer to perform the following processes: Calculate the transformation error between the transformed image and the training image input during learning; Based on the transformed image and the training image, calculate the size metric of the pattern contained in the training image; The composite error is calculated based on the transformation error and the size metric. as well as Learn a machine learning model for generating the transformed image and optimize the parameters of the machine learning model to optimize the composite error.

Images