Machine learning device, machine learning method, and machine learning program

Abstract

To improve the accuracy of predicting material properties of metals.SOLUTION: A machine learning apparatus includes: a division unit 15B which divides an original image obtained by imaging a metal through a microscope into a plurality of partial images; and a parameter update unit 15E which updates a parameter of the machine learning model on the basis of a label for material properties which is output by a machine learning model to which the partial images have been input and a ground truth label added to the original image.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a machine learning device, a machine learning method, and a machine learning program. [Background technology]

[0002] A material property estimation device has been proposed as one technology for estimating the material properties of metals. For example, the material property estimation device reads a microstructure image of a test specimen taken with a microscope and estimates the material property value category to which the material belongs using a neural network model. For example, if the tensile strength of all test specimens is within the range of 560 MPa to 699 MPa, 11 categories, from category 0 to category 10, are assigned in increments of 10 MPa starting from 560 MPa. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-12037 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in conventional technologies, including the material property estimation device described above, the microstructure image is compressed to a size corresponding to the prediction model when the prediction model is created, which results in the loss of information essential for predicting material properties, resulting in a decrease in the accuracy of predictions of material properties.

[0005] In one aspect, the present invention aims to provide a machine learning device, a machine learning method, and a machine learning program that can improve the prediction accuracy of metal material properties. [Means for solving the problem]

[0006] One aspect of the machine learning device includes a division unit that divides an original image of a metal captured by a microscope into multiple partial images, and a parameter update unit that updates the parameters of the machine learning model based on labels of material properties output by a machine learning model to which the partial images are input and on correct labels assigned to the original image. [Effects of the Invention]

[0007] This can improve the accuracy of predicting metal material properties. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram illustrating an example of the functional configuration of a server device. [Figure 2] FIG. 2 is a schematic diagram (1) showing an example of resizing an original image. [Figure 3] FIG. 3 is a schematic diagram (2) showing an example of resizing an original image. [Figure 4] FIG. 4 is a diagram showing an example of a correct label. [Figure 5] FIG. 5 is a schematic diagram showing an example of a method for dividing an original image. [Figure 6] FIG. 6 is a schematic diagram showing an example of a method for calculating loss. [Figure 7] FIG. 7 is a flowchart showing the procedure of the machine learning process. [Figure 8] FIG. 8 is a diagram illustrating an example of a hardware configuration. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, embodiments of a machine learning device, a machine learning method, and a machine learning program according to the present application will be described with reference to the accompanying drawings. Each embodiment merely illustrates one example or aspect, and such examples do not limit the range of values, functions, or usage scenarios. Furthermore, each embodiment can be appropriately combined within the scope of not causing contradictions in the processing content.

[0010] Fig. 1 is a block diagram showing an example of the functional configuration of a server device 10. The server device 10 shown in Fig. 1 provides a machine learning function for learning a machine learning model that outputs information related to the material properties of metals.

[0011] Below, we will use the example of metal hardness as just one example of material properties predicted by a machine learning model. However, we would like to point out that other material properties, such as tensile strength, fatigue resistance, and creep resistance, can also be predicted by a machine learning model.

[0012] The server device 10 is an example of a computer that provides the above-described machine learning function. In one embodiment, the server device 10 can provide the above-described machine learning function by executing software that realizes the above-described machine learning function. For example, the server device 10 can be realized as a server that provides the above-described machine learning function on-premise. Alternatively, the server device 10 can be realized as a PaaS (Platform as a Service) or SaaS (Software as a Service) application, thereby providing the above-described machine learning function as a cloud service.

[0013] As shown in Fig. 1, the server device 10 can be communicatively connected to a client terminal 30 via a network NW. For example, the network NW may be any type of communication network, whether wired or wireless, such as the Internet or a local area network (LAN). Note that Fig. 1 shows an example in which one client terminal 30 is connected to one server device 10, but any number of client terminals 30 may be connected.

[0014] The client terminal 30 corresponds to an example of a computer that receives the above-described machine learning function. For example, the client terminal 30 may be realized by a desktop or laptop personal computer. This is merely an example, and the client terminal 30 may be any computer, such as a mobile terminal device or a wearable terminal.

[0015] Note that while Figure 1 shows an example in which the above machine learning function is provided in a client-server system, this is merely an example, and the above machine learning function may also be provided as a standalone function.

[0016] <One aspect of the issue> As explained in the background art above, in the conventional techniques, the microstructure image is compressed to an input size corresponding to the prediction model at the stage when the prediction model is created.

[0017] Figure 2 is a schematic diagram (1) showing an example of resizing an original image. Figure 2 shows how the image size of an original image 21, which is 4 pixels wide by 4 pixels high, is resized to the input size of a machine learning model, which is 2 pixels wide by 2 pixels high.

[0018] 2, an original image 21 is compressed to a compressed image 22 of 2 pixels wide and 2 pixels high by thinning out 12 pixels out of 16 pixels. When resizing is achieved by compressing the image size in this way, 12 pixels are thinned out during compression, so the detailed information observable in the original image 21 is lost from the compressed image 22.

[0019] Here, Fig. 2 shows a schematic example, but in reality, more information will be lost than in the example shown in Fig. 2. For example, in the example described in the above-mentioned prior art, if an image has a size of 1252 pixels wide by 990 pixels high, it will be compressed to 120 pixels wide by 94 pixels high, so as many as 1,228,200 pieces of information will be lost.

[0020] When the amount of information lost increases when the image is resized to fit the input size of a machine learning model, information essential for predicting material properties may also be lost. For example, when capturing an image of a tempered martensite structure, such as that found in spring steel, information about fine iron carbides that can be observed in the original image may be lost in the resized, compressed image.

[0021] These fine iron carbides have a significant impact on the material properties of metals. Therefore, even if a machine learning model is trained using compressed images with fine details such as fine iron carbides lost as training data, it is difficult for the trained machine learning model to capture the material properties of metals, which can result in a decrease in the accuracy of predictions of material properties.

[0022] <One aspect of the problem-solving approach> Therefore, the machine learning function according to this embodiment differs from the conventional techniques in that it takes a different approach to resizing an original image of a metal structure captured by a microscope such as a scanning electron microscope to the input size of the machine learning model. That is, the machine learning function according to this embodiment achieves resizing to the input size of the machine learning model by dividing the original image.

[0023] Figure 3 is a schematic diagram (2) showing an example of resizing an original image. Figure 3 also shows a schematic diagram of how the image size of an original image 21, which is 4 pixels wide by 4 pixels high, is resized to the input size of a machine learning model, which is 2 pixels wide by 2 pixels high.

[0024] As shown in FIG. 3, the machine learning function according to this embodiment divides an original image 21 into four partial images 21A to 21D, and resizes them to the input size of a machine learning model of 2 pixels wide by 2 pixels high.

[0025] In this way, the partial images 21A to 21D obtained by dividing the original image 21 contain all of the information contained in the original image 21 without any loss of information due to thinning or interpolation, and therefore, loss of detailed information such as fine iron carbides can be prevented.

[0026] Therefore, the machine learning model can be trained using partial images that do not lose detailed information such as fine iron carbides as training data, which increases the likelihood that the trained machine learning model will capture the material properties of the metal.

[0027] Therefore, the machine learning function according to this embodiment can improve the accuracy of predicting the material properties of metals. By realizing the prediction of material properties using a machine learning model in this way, it is possible to reduce the number of material evaluation tests that need to be performed and to support the construction of a materials database in which metallographic images and material properties are associated with each other.

[0028] Furthermore, the machine learning function according to this embodiment divides an original image, so that multiple partial images can be acquired from a single original image as training data. This is particularly significant in the field of machine learning, particularly deep learning, because achieving a certain level of performance in deep learning requires a huge amount of data. In addition, the correct label assigned to the original image can be assigned to partial images divided from the same original image. This reduction in labeling work is also of great technical significance.

[0029] <Configuration of Server Device 10> Next, an example of the functional configuration of the server device 10 according to this embodiment will be described. FIG. 1 shows a block diagram related to the machine learning function of the server device 10. As shown in FIG. 1, the server device 10 has a communication control unit 11, a storage unit 13, and a control unit 15. Note that FIG. 1 only shows a selection of functional units related to the above-mentioned machine learning function, and the server device 10 may also be provided with functional units other than those shown.

[0030] The communication control unit 11 is a functional unit that controls communication with other devices such as the client terminal 30. As just one example, the communication control unit 11 is realized by a network interface card such as a LAN card. In one aspect, the communication control unit 11 receives a request for training a machine learning model from the client terminal 30. In another aspect, the communication control unit 11 outputs to the client terminal 30 the results of training the machine learning model, such as the trained machine learning model, as well as the accuracy and loss for the training data, the accuracy and loss for the validation data, and the like.

[0031] The storage unit 13 is a functional unit that stores various types of data. As just one example, the storage unit 13 is realized by internal, external, or auxiliary storage of the server device 10. For example, the storage unit 13 stores a training dataset 13D, first model data 13M1, and second model data 13M2. Data other than these, such as settings for the shooting magnification of the original image, the division size of the original image, and data used for filtering partial images, may also be stored in the storage unit 13.

[0032] The training data set 13D is a collection of data used in machine learning, that is, training data. As just one example, the training data may be image data with a correct answer label.

[0033] As just one example, the image data may be a scanning electron microscope (SEM) image of a metal structure photographed by a microscope such as a scanning electron microscope (SEM).

[0034] In the following, in order to distinguish between the labels of the SEM image captured by the microscope and the partial images into which the SEM image is divided, the former may be referred to as the "original image" and the latter as the "partial image."

[0035] For example, when photographing a metal structure, the magnification of the SEM can be set to a value within the range of 5,000x to 50,000x. Setting the magnification within this range has the advantage of preventing situations where information on fine iron carbides, etc., cannot be observed due to photographing at a magnification lower than the above-mentioned magnification. Furthermore, it has the advantage of preventing bias in the amount of iron carbides appearing in the photographed area due to photographing at a magnification higher than the above-mentioned magnification.

[0036] Such SEM images are assigned a correct label of Vickers hardness. FIG. 4 is a diagram showing an example of the correct label. Although four SEM images 13d1 to 13d4 are selected and shown in FIG. 4 as an example only, the number of training data items is not limited to this example. As shown in FIG. 4, a correct label of the Vickers hardness class is assigned to each of the four SEM images 13d1 to 13d4 as a correct label.

[0037] In the example shown in Figure 4, four types of correct labels, "500HV," "600HV," "700HV," or "800HV," may be assigned to the 500HV range (500HV or more but less than 600HV), as well as the 600HV range, 700HV range, and 800HV range classes.

[0038] As an example only, the correct label for Vickers hardness is assigned based on the measurement result of Vickers hardness obtained in a material evaluation test carried out for each test piece corresponding to each of the four SEM images 13d1 to 13d4.

[0039] For example, training data 13D1 is obtained by assigning a correct label "500HV" to SEM image 13d1. Similarly, training data 13D2 is obtained by assigning a correct label "600HV" to SEM image 13d2. Furthermore, training data 13D3 is obtained by assigning a correct label "700HV" to SEM image 13d3. Furthermore, training data 13D4 is obtained by assigning a correct label "800HV" to SEM image 13d4.

[0040] The training data set 13D thus obtained may be divided according to a validation algorithm such as cross-validation. For example, the training data set 13D can be divided into training data 13TR used to train parameters such as weights and biases of the machine learning model, validation data used to verify overfitting during the training process, and test data used to test generalization performance.

[0041] Among the data stored in the storage unit 13, the first model data 13M1 and the second model data 13M2 will be described together when the first model data 13M1 and the second model data 13M2 are referenced or registered.

[0042] The control unit 15 is a functional unit that performs overall control of the server device 10. For example, the control unit 15 can be realized by a hardware processor. As shown in FIG. 1, the control unit 15 has a reception unit 15A, a division unit 15B, a filtering unit 15C, a model execution unit 15D, and a parameter update unit 15E. Note that the control unit 15 may also be realized by hardwired logic.

[0043] The reception unit 15A is a processing unit that receives various requests from the client terminal 30. As just one aspect, the reception unit 15A can receive a training request requesting the execution of training of a machine learning model from the client terminal 30. For example, when receiving a training request, the reception unit 15A can also receive a specification of a learning dataset to be used in training the machine learning model, as well as a specification of hyperparameters of the machine learning model.

[0044] The dividing unit 15B is a processing unit that divides an original image. In one embodiment, the dividing unit 15B divides the original image included in each piece of training data 13TR of the learning data set 13D stored in the storage unit 13.

[0045] Fig. 5 is a schematic diagram showing an example of a method for dividing an original image. As shown in Fig. 5, the dividing unit 15B divides the original image 13TR into a mesh. The image size of the grid cells during such mesh division can be set to a width and height that are equal to or smaller than the size of an input layer of a machine learning model used to predict the material properties of metal, for example, a convolved neural network (CNN).

[0046] For example, if the image size of the grid cells is made too small, in other words, if the mesh density is made too fine, the iron carbides that appear within the range of the grid cells may be unevenly distributed, which may adversely affect the prediction accuracy.

[0047] Such uneven distribution of iron carbides is more frequently observed when the width and height of the image size of the grid cell are 2 μm or less, so it is preferable to set the width and height of the image size of the grid cell to values ​​exceeding 2 μm.

[0048] By the above mesh division, 15 partial images 13TR1 to 13TR15 are obtained from the original image 13TR. These partial images 13TR1 to 13TR15 can be assigned a common correct label of the same class as the correct label assigned to the original image 13TR.

[0049] By performing this mesh division, 15 training samples can be obtained from one training sample, the original image 13TR. Furthermore, by assigning a common correct label, the labeling work can be reduced.

[0050] The filtering unit 15C is a processing unit that filters partial images. The term "filtering" here refers to selection, such as extracting partial images that satisfy certain conditions listed below from the partial images divided from the original image, or excluding partial images that satisfy certain conditions.

[0051] In one aspect, the filtering unit 15C extracts, from the M partial images into which the original image is divided by the division unit 15B, partial images that contain detailed information that contributes to predicting material properties, for example, partial images in which the number of pixels in the frequency band corresponding to fine iron carbides is greater than or equal to a threshold value.

[0052] More specifically, the filtering unit 15C applies a band-pass filter, the center frequency of which corresponds to iron carbide, to each of the M partial images. Based on the result of applying this band-pass filter, the filtering unit 15C calculates the number of pixels corresponding to iron carbide in the partial image. At this time, if the number of pixels corresponding to iron carbide in the partial image is equal to or greater than a threshold value Th1, the filtering unit 15C extracts the partial image as a training sample. On the other hand, if the number of pixels corresponding to iron carbide in the partial image is less than the threshold value Th1, the filtering unit 15C prohibits the extraction of the partial image as a training sample. This enables filtering to extract partial images that contribute to predicting material properties as training samples.

[0053] In another aspect, the filtering unit 15C excludes, from the M partial images into which the original image is divided by the division unit 15B, partial images that contain information that becomes noise when predicting material properties, such as peculiar parts that are similar to contamination such as organic matter.

[0054] More specifically, the filtering unit 15C applies the following template matching to each of the M partial images. For example, in template matching, images of one or more patterns of unique parts observed in an SEM image are used as templates. These templates are then compared with the input partial image by enlarging, reducing, and rotating them. This allows for searching for areas where the similarity between the unique part template and the partial image is equal to or greater than threshold Th2, or areas where the distance between the unique part template and the partial image is equal to or less than threshold Th3. If the search results in an area similar to the unique part template, the filtering unit 15C excludes the partial image from the training sample. On the other hand, if no area similar to the unique part template is found, the filtering unit 15C prohibits the partial image from being excluded from the training sample. This enables filtering to exclude partial images containing unique parts such as contamination from the training sample.

[0055] The filtering unit 15C can perform one or both of filtering to extract partial images that contribute to predicting material properties and filtering to remove partial images that include peculiar parts such as contamination. Although an example in which partial image filtering is performed has been given here, all partial images may be used as training samples, or some partial images may be randomly extracted or excluded.

[0056] The model execution unit 15D is a processing unit that executes a machine learning model. Hereinafter, as an example of a machine learning model, a machine learning model that executes a classification task of inputting an image and outputting a Vickers hardness class will be described. This is merely an example, and the machine learning model may output a class related to a material property other than Vickers hardness, or may output a confidence level for each class.

[0057] In one aspect, at the start of training when the receiving unit 15A receives a training request, the model executing unit 15D references the first model data 13M1 stored in the storage unit 13 and deploys a machine learning model on a work area (not shown). For example, the first model data 13M1 defines parameters corresponding to objective functions, such as weights and biases for each layer, as well as hyperparameters related to the layer structure of the machine learning model, such as neurons and synapses in the input layer, hidden layer, and output layer that form the CNN. Note that the first model data 13M1 referenced at the start of training may define parameters of the machine learning model that are initially set using random numbers or the like.

[0058] In another aspect, the model execution unit 15D inputs each partial image extracted as a training sample by the filtering unit 15C to a machine learning model, thereby enabling the model execution unit 15D to obtain a class related to material properties output by the machine learning model to which the partial image is input.

[0059] The parameter update unit 15E is a processing unit that updates the parameters of the machine learning model. As one aspect, the parameter update unit 15E calculates, for each partial image extracted as a training sample by the filtering unit 15C, a loss of the output of the machine learning model with respect to the correct label as an example of the difference between the output of the machine learning model and the correct label. FIG. 6 is a schematic diagram showing an example of a method for calculating the loss. FIG. 6 shows an example in which partial image 13TR1, of partial images 13TR1 to 15TR15 shown in FIG. 5, is input to the machine learning model 13M1. In the example shown in FIG. 6, when partial image 13TR1 is input to the machine learning model 13M1, the machine learning model 13M1 outputs a label of the class "800HV." Then, the parameter update unit 15E calculates the error of the label of the class "800HV" with respect to the correct label assigned to the partial image 13TR1 as the loss.

[0060] In another aspect, when the loss for the number of data items corresponding to a predetermined batch size is calculated, the parameter update unit 15E updates the parameters of the machine learning model based on the loss calculated for each partial image in that batch size. As just one example, the parameters can be updated by optimizing an objective function such as log-likelihood. This parameter update is repeated a number of times corresponding to the number of batches, completing training of one epoch of the machine learning model.

[0061] The training of the machine learning model is then repeated until an arbitrary termination condition is met. As just one example, the training of the machine learning model can be terminated when the number of epochs for which the training of the machine learning model has been performed reaches a threshold. As another example, the training of the machine learning model can be continued until overfitting is detected due to an increase in the difference between the loss of the training data and the loss of the validation data, or the difference between the accuracy rate of the training data and the accuracy rate of the validation data. If overfitting is detected, the hyperparameters of the machine learning model can be tuned based on the difference between the loss of the training data and the loss of the validation data, or the difference between the accuracy rate of the training data and the accuracy rate of the validation data.

[0062] The model information of the trained machine learning model obtained in this manner, such as parameters and hyperparameters, is stored as second model data 13M2 in the storage unit 13. In addition, the trained machine learning model can be output as the training result of the machine learning model to any output destination, such as an external device such as the client terminal 30, or software or a service that uses the machine learning model. When providing a trained machine learning model in this manner, it can also be provided as a module or library that incorporates a function for dividing an input image into image sizes that are the same as or equivalent to the image sizes into which the original image is divided when training the machine learning model.

[0063] <Processing flow> 7 is a flowchart showing the steps of the machine learning process. This process may be executed, by way of example only, when a training request for training a machine learning model is received from the client terminal 30.

[0064] 7, loop process 1 is executed to repeat the processes from step S101 to step S105 described below until an arbitrary termination condition is met. Loop process 1 includes loop process 2 that repeats the processes from step S101 to step S105 described below a number of times corresponding to the number of batches equivalent to one epoch, in other words, the total number K of training data. Furthermore, loop process 2 further includes loop process 3 that repeats the processes from step S101 to step S104 described below a number of times corresponding to the number L of data items in the batch size.

[0065] For example, the dividing unit 15B divides an original image included in the training data 13TR of the learning data set 13D into M partial images (step S101). Subsequently, the filtering unit 15C performs filtering such as extracting partial images that satisfy a specific condition from the partial images obtained by the division in step S101, or excluding partial images that satisfy a specific condition (step S102).

[0066] Thereafter, loop processing 4 is executed in which the processing of the following step S103 and step S104 is repeated the number of times corresponding to the M partial images extracted as training samples by filtering in step S102.

[0067] That is, the model executing unit 15D inputs the partial images extracted as training samples by filtering in step S102 to the machine learning model, thereby obtaining labels of the material property classes output by the machine learning model (step S103).

[0068] Then, the parameter update unit 15E calculates the difference between the label acquired as the output of the machine learning model in step S103 and the correct label assigned to the original image from which the image was divided as the loss of the output of the machine learning model for the correct label (step S104).

[0069] By repeating this loop process 4, a loss is obtained for each of the M partial images extracted as training samples by filtering in step S102. Furthermore, by repeating loop process 3, a loss for each of the M partial images is obtained for each number of data items corresponding to a predetermined batch size.

[0070] Then, the parameter update unit 15E updates the parameters of the machine learning model based on the loss in the number of data items corresponding to the batch size (step S105).

[0071] After that, training of the machine learning model for one epoch is completed by repeating loop process 2. Furthermore, training of the machine learning model is completed by repeating loop process 1. As a result, a trained machine learning model can be generated.

[0072] <One aspect of the effect> As described above, the server device 10 according to this embodiment resizes the original image to the input size of the machine learning model by dividing the original image. Therefore, the server device 10 according to this embodiment can improve the prediction accuracy of the material properties of metals. By realizing prediction of material properties using a machine learning model in this way, it is possible to reduce the number of material evaluation tests performed and to support the construction of a materials database in which metallographic images and material properties are associated with each other.

[0073] Furthermore, according to the server device 10 of this embodiment, since the original image is divided, multiple partial images can be obtained from one original image as training data. This is particularly technically significant in the field of machine learning, particularly deep learning, because a huge amount of data is required to achieve a certain level of performance in deep learning. In addition, the correct label assigned to the original image can be commonly assigned to partial images divided from the same original image. This reduction in labeling work is also of great technical significance.

[0074] Additionally, the server device 10 according to this embodiment can split original images inline during training of a machine learning model. This allows images stored as training data to remain unchanged as original images, thereby preventing the management of training data and, ultimately, learning datasets from becoming complicated.

[0075] <Modification> Furthermore, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method. In addition, the information including the processing procedures, specific names, various data, and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified. For example, the various information shown in each drawing is not limited to the information shown in the drawings.

[0076] Furthermore, the components of each device shown in the figure are conceptual functional components and do not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or part of them can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc.

[0077] Furthermore, the effects of each embodiment described in this specification are merely examples and are not intended to be limiting, and other effects may also be obtained.

[0078] <Hardware configuration> Furthermore, the various processes described in the above embodiments can be realized by executing a prepared program on a computer such as a personal computer or a workstation. Therefore, an example of a computer that executes a machine learning program having the same functions as those in the above embodiments will be described below with reference to FIG. 8.

[0079] Fig. 8 is a diagram showing an example of a hardware configuration. As shown in Fig. 8, a computer 100 has an operation unit 110a, a speaker 110b, a camera 110c, a display 120, and a communication unit 130. The computer 100 also has a CPU 150, a ROM 160, an HDD 170, and a RAM 180. These units 110 to 180 are connected via a bus 140.

[0080] 8, the HDD 170 stores a machine learning program 170a that performs the same functions as the receiving unit 15A, dividing unit 15B, filtering unit 15C, model executing unit 15D, and parameter updating unit 15E shown in the first embodiment. This machine learning program 170a may be integrated or separated, similar to the components of the receiving unit 15A, dividing unit 15B, filtering unit 15C, model executing unit 15D, and parameter updating unit 15E shown in FIG. 1. In other words, the HDD 170 does not necessarily have to store all of the data shown in the first embodiment, as long as the data used for processing is stored in the HDD 170.

[0081] In this environment, CPU 150 reads machine learning program 170a from HDD 170 and loads it into RAM 180. As a result, machine learning program 170a functions as machine learning process 180a, as shown in FIG. 8. This machine learning process 180a loads various data read from HDD 170 in an area of ​​RAM 180's storage area allocated to machine learning process 180a, and executes various processes using the loaded data. For example, examples of processes executed by machine learning process 180a may include the processes shown in FIG. 7. Note that CPU 150 does not necessarily need to operate all of the processing units shown in the above embodiments; it is sufficient that processing units corresponding to the processes to be executed are virtually implemented.

[0082] The above-described machine learning program 170a does not necessarily have to be stored in the HDD 170 or the ROM 160 from the beginning. For example, the machine learning program 170a may be stored on a "portable physical medium" such as a flexible disk, a so-called FD, a CD-ROM, a DVD disk, a magneto-optical disk, or an IC card, which is inserted into the computer 100. The computer 100 may then acquire and execute the machine learning program 170a from such a portable physical medium. Alternatively, the machine learning program 170a may be stored in another computer or server device connected to the computer 100 via a public line, the Internet, a LAN, a WAN, or the like. The machine learning program 170a thus stored may be downloaded to the computer 100 and then executed. [Explanation of symbols]

[0083] 10 Server device 11 Communication control section 13 Storage section 13D training dataset 13TR training data 13M1 First model data 13M2 Second model data 15 Control Unit 15A Reception 15B Split section 15C Filtering section 15D Model Execution Department 15E Parameter Update Section 30 client terminals

Claims

1. a division unit that divides an original image of a metal imaged under a microscope at a magnification corresponding to observation of iron carbides into a plurality of partial images; a parameter update unit that updates parameters of the machine learning model based on a label of material properties output by the machine learning model to which the partial image is input and a correct answer label assigned to the original image; and The machine learning device is characterized in that the division unit divides the original image into image sizes equal to the size of the input layer of the machine learning model.

2. a filtering unit configured to extract, from the plurality of partial images, a partial image in which the number of pixels in a frequency band corresponding to a specific object is equal to or greater than a threshold value; The machine learning device according to claim 1 , wherein the parameter update unit uses the partial image extracted by the filtering unit to update parameters of the machine learning model.

3. The method further includes a filtering unit that excludes partial images including a peculiar part similar to a contamination from among the plurality of partial images, The machine learning device according to claim 1 , wherein the parameter update unit prohibits the partial image excluded by the filtering unit from being used to update the parameters of the machine learning model.

4. 4. The machine learning device according to claim 1, wherein the plurality of partial images are commonly assigned the same correct label as that assigned to the original image.

5. Dividing an original image of a metal taken under a microscope at a magnification corresponding to the observation of iron carbides into a plurality of partial images; updating parameters of the machine learning model based on the label of the material property output by the machine learning model to which the partial image is input and the correct label assigned to the original image; The computer executes the processing, A machine learning method characterized in that the division process divides the original image into image sizes equal to the size of the input layer of the machine learning model.

6. Dividing an original image of a metal taken under a microscope at a magnification corresponding to the observation of iron carbides into a plurality of partial images; updating parameters of the machine learning model based on the label of the material property output by the machine learning model to which the partial image is input and the correct label assigned to the original image; Have the computer execute the process, The machine learning program is characterized in that the division process divides the original image into image sizes equal to the size of the input layer of the machine learning model.

Images