JP7760427B2 - Particle size class determination device, image size determination support device, method and program - Google Patents

Description

The present invention relates to a particle size class determination device that determines the particle size class of an object to be measured when particle sizes are divided into multiple particle size classes in multiple different ranges, and to an image size determination support device, image size determination support method, and image size determination support program that support the determination of the size of an image to be used for particle size class determination.

In the production of steel using blast furnaces, not only mined iron ore is used, but also sintered ore. Sintered ore particles are pseudo-particles consisting of core powders, such as iron ore, limestone, and coke, and finer particles of these particles attached to the surface of the core powder. For example, fine ore with a wide particle size distribution, with a maximum diameter of approximately 15 mm, is used as raw material. Fine ore particles of 1 mm or less are attached to fine ore particles of 1 mm or more, resulting in sintered ore particles averaging approximately 3 to 5 mm through the so-called sintering process. In this sintering process, heat is transferred to a stack of raw materials stacked in layers by heating them from above and drawing air below. Therefore, gaps are required within the stack to allow air to circulate. Sintered ore particles smaller than 2 mm typically impair air permeability and reduce sinter productivity. Therefore, there is a need to measure particle size and particle size distribution to avoid the generation of particles that impair air permeability during the sintering process.

A technology for measuring this particle size distribution is disclosed, for example, in Patent Document 1. The soil particle size distribution estimation method disclosed in Patent Document 1 estimates the particle size distribution of sampled soil using a trained model obtained through machine learning by artificial intelligence installed on a server, using numerous images of each single-particle size soil, which has been sieved from raw soil and classified into multiple particle sizes, as training data. The images of each single-particle size soil that serve as training data are taken with a camera at a predetermined shooting distance of the surface of the single-particle size soil, which has been spread evenly to a thickness that leaves no voids in a planar view, and are linked to the proportions of each particle size into the classified areas. The particle size distribution of the sampled soil is estimated by inputting images taken with the camera at the predetermined shooting distance of the surface of the sampled soil, which has been spread evenly to a thickness that leaves no voids in a planar view, into the trained model and outputting the proportions of each particle size into which the sampled soil has been classified. Paragraph [0029] of the same document states, "Furthermore, soil properties also include moisture content and particle composition, but these properties can be excluded from the scope of the present invention. In this embodiment, naturally dried soil generated on-site is used as the raw soil, and this raw soil is sieved to obtain single-grain size soils of each classified grain size that serve as training data. In addition, naturally dried raw soil generated on-site that is in a state where soils of various grain sizes remain mixed and is not sieved is used as the collected soil (generated on-site soil) for which the grain size distribution is to be estimated, and the grain size distribution can be estimated."

Japanese Patent Application Laid-Open No. 2021-117625

As described in paragraph [0029] of Patent Document 1, the soil particle size distribution estimation method disclosed in Patent Document 1 can estimate particle size distribution when the measurement object is in a naturally dried state. For this reason, the soil particle size distribution estimation method disclosed in Patent Document 1 is considered unsuitable for measurement objects that are not in a naturally dried state. In particular, in the sintering process, water is added to the raw materials, the mixture is kneaded, and then baked. Therefore, the soil particle size distribution estimation method disclosed in Patent Document 1 is considered unsuitable for measuring the particle size distribution of this mixture. Furthermore, because the sintered ore particles produced in the sintering process are pseudo-particles, it is necessary to determine the particle size of not only the powder but also the fine powder attached to the powder. However, the soil particle size distribution estimation method disclosed in Patent Document 1 may estimate the particle size distribution based only on the powder or the fine powder.

The present invention has been made in consideration of the above-mentioned circumstances, and its purpose is to provide a particle size class determination device that can determine the particle size class of a wider range of objects to be determined, as well as an image size determination support device, image size determination support method, and image size determination support program that support the determination of the size of the image to be used for particle size class determination in these devices.

After extensive investigation, the inventors have found that the above-mentioned object can be achieved by the present invention. Specifically, a particle size class determination device according to one aspect of the present invention includes an image acquisition unit that acquires images of multiple particles, including particles of different particle sizes, and a particle size class processing unit that receives an input image and classifies the input image into a plurality of particle size classes based on particle sizes in a plurality of different ranges. The particle size class processing unit calculates the probability of each of the multiple particle size classes for the image acquired by the image acquisition unit using a machine learning model (first machine learning model) that has been machine-learned and outputs, for each of the multiple particle size classes, the probability that the input image will be classified into that particle size class. One of the boundary values in the multiple ranges is a target particle size value for determination, and the size of the image input to the machine learning model is a size obtained by multiplying the target particle size value by a predetermined constant. Preferably, in the above-mentioned particle size class determination, the target particle size value is set to the upper limit of the range corresponding to the particle size class with the smallest particle size. Preferably, in the above-mentioned particle size class determination, the particle size class processing unit directly uses the calculated probabilities as the probability of each particle size class for the input image.

According to the inventor's findings, when particle size classes are set to three - a large particle size class for large particle sizes, a small particle size class for small particle sizes, and a medium particle size class between these - if the image size is too small, an image capturing large particle sizes will not capture the entire particle, making it difficult to extract them properly. In particular, if the particle is a pseudo-particle, there is a risk that fine powder (an example of a second particle, described below) attached to the core powder (an example of a first particle, described below) will be extracted, making it difficult to extract the actual pseudo-particle. On the other hand, if the image size is too large and both large and small particles are captured in the image, there is a risk that the outline information of the small particle will be hidden, making it difficult to extract them properly. For this reason, it is believed that there is an optimal image size for the target particle size value.

The particle size class determination device uses a machine learning model to determine the particle size class of images of multiple particles. By performing machine learning on a variety of images of particles to be determined, it is possible to determine a wider range of particle size classes. During this determination, the size of the image input to the machine learning model is a predetermined constant multiplied by the particle size value of the determination target, allowing the particle size class determination device to appropriately determine the particle size class. The particle size class determination device is particularly suitable when the particle size to be determined is a pseudo particle.

In another aspect, in the particle size class determination device described above, the predetermined constant is greater than 7 and less than 12. Preferably, in the particle size class determination device described above, the predetermined constant is greater than 8 and less than 11. Preferably, in the particle size class determination device described above, the predetermined constant is 10.

This makes it possible to provide a particle size class determination device in which the predetermined constant is greater than 7 and less than 12.

In another aspect, in the above-mentioned particle size class determination device, the particles are pseudo-particles comprising a first particle that serves as a nucleus and a second particle that is smaller than the first particle and attached to the surface of the first particle. Preferably, in the above-mentioned particle size class determination, the pseudo-particles are particles of sintered ore used in blast furnaces.

This provides a particle size class determination device for determining pseudo particles. Because the particle size class determination device is used to determine the particle size class of pseudo particles, the particle size class of the pseudo particles can be determined appropriately.

In another aspect, the above-mentioned particle size class determination device further includes an illumination unit that illuminates the plurality of particles, and the image acquisition unit is an imaging unit that generates an image.

This type of particle size class determination device further includes an illumination unit and uses an imaging unit for the image acquisition unit, so the particle size class determination device can be used, for example, on a production line within a factory, allowing particle size classes to be determined during production.

An image size determination assistance device according to another aspect of the present invention is a device for assisting in determining the size of an image to be used to determine a particle size class when particle sizes are divided into a plurality of different ranges as particle size classes, the device comprising: a first data acquisition unit that acquires a predetermined training dataset; a second data acquisition unit that acquires a predetermined size determination dataset; a machine learning unit that uses the predetermined training dataset acquired by the first data acquisition unit to train a second machine learning model to which an image is input and that outputs, for each of the plurality of particle size classes, a probability that the input image is classified into that particle size class; and an output unit that inputs the predetermined size determination dataset acquired by the second data acquisition unit into the second machine learning model trained by the machine learning unit and outputs each probability for each of the plurality of particle size classes output from the second machine learning model trained by the machine learning unit, the predetermined training dataset and the predetermined size determination dataset each comprising a plurality of data groups each comprising image data of a plurality of images, the plurality of data groups having different image sizes, and the plurality of images in each of the plurality of data groups in the predetermined training dataset being classified into the plurality of particle sizes. the plurality of images including images of each class, and each of the plurality of images in the data group is linked to the particle size class of the image; in each of the plurality of data groups in the specified size determination dataset, the plurality of images in the data group are multiple images including images with different content rates of particles belonging to the particle size class with the smallest particle size, and each of the plurality of images in the data group is linked to the content rate of the image; the number of the unmachine-learned second machine learning models is the same as the number of the plurality of data groups and is provided corresponding to each of the plurality of data groups; the machine learning unit trains, for each of the plurality of data groups in the specified training dataset, an unmachine-learned second machine learning model corresponding to the data group using the plurality of images in the data group; and the output unit inputs, for each of the plurality of data groups in the specified size determination dataset, the plurality of images in the data group into the second machine learning model trained by the machine learning unit using a data group in the specified training dataset that has the same image size as the image in the data group, and outputs the probability of each of the plurality of particle size classes output from the second machine learning model trained by the machine learning unit.

This type of image size determination assistance device outputs the probability for each of the multiple particle size classes obtained by inputting a predetermined size determination dataset into a machine-learned second learning model. The user can then determine the optimal image size from among multiple image sizes by referencing the output probability for each of the multiple particle size classes and the content associated with the image. Therefore, the image size determination assistance device can effectively assist the user in determining the image size. This provides an image size determination assistance device that assists in determining the image size to be used to determine the particle size class.

In another aspect, the above-mentioned particle size class determination device further includes the above-mentioned image size determination support device, and the machine-learned machine learning model uses a second machine learning model trained by the machine learning unit.

This makes it possible to provide a particle size class determination device with a function to assist in determining image size. Because the particle size class determination device is integrated with the image size determination assistance device, a single device can determine image size and particle size class.

A particle size class determination method according to another aspect of the present invention includes an image acquisition step of acquiring an image of a plurality of particles, including particles of different particle sizes, and a particle size class processing step of inputting an image and dividing the particle sizes into a plurality of different ranges into a plurality of particle size classes, using a machine learning model that has been machine-learned to output, for each of the plurality of particle size classes, the probability that the input image will be classified into that particle size class. One of the boundary values in the plurality of ranges is a particle size value for determination, and the size of the image input to the machine learning model is the size obtained by multiplying the particle size value for determination by a predetermined constant.

In another aspect of the present invention, a particle size class determination program is a computer-executable program that includes an image acquisition step for acquiring images of multiple particles, including particles of different particle sizes, and a particle size class processing step for inputting an image and dividing the particle sizes into multiple different ranges into multiple particle size classes, using a machine learning model that has been machine-learned to output, for each of the multiple particle size classes, the probability that the input image will be classified into that particle size class. One of the boundary values in the multiple ranges is a target particle size value for determination, and the size of the image input to the machine learning model is the target particle size value multiplied by a predetermined constant.

This particle size class determination method and particle size class determination program uses a machine learning model to determine the particle size class of images of multiple particles, so by performing machine learning on a variety of images of the particle to be determined, it is possible to determine a wider range of particle size classes for the particle to be determined. Because the size of the image input to the machine learning model during this determination is a predetermined constant multiplied by the particle size value of the determination target, the particle size class determination method and particle size class determination program can appropriately determine the particle size class. The particle size class determination method and particle size class determination program are particularly suitable when the particle to be determined is a pseudoparticle.

An image size determination assistance method according to another aspect of the present invention is a method for assisting in determining the size of an image to be used to determine a particle size class when particle sizes are divided into a plurality of different ranges as particle size classes, the method comprising: a first data acquisition process for acquiring a predetermined training dataset; a second data acquisition process for acquiring a predetermined size determination dataset; a machine learning process for inputting an image and using the predetermined training dataset acquired in the first data acquisition process to train a second machine learning model that is not yet machine-learned and outputs, for each of the plurality of particle size classes, a probability that the input image will be classified into that particle size class; and an output process for inputting the predetermined size determination dataset acquired in the second data acquisition process into the second machine learning model trained in the machine learning process, thereby outputting the probability for each of the plurality of particle size classes output from the second machine learning model trained in the machine learning process.The predetermined training dataset and the predetermined size determination dataset each comprise a plurality of data groups each comprising image data of a plurality of images, the plurality of data groups having different image sizes, and the plurality of images in each of the plurality of data groups in the predetermined training dataset are classified into the plurality of data groups. the plurality of images including images of each particle size class, and the particle size class of each of the plurality of images in the data group is linked; the plurality of images in each of the plurality of data groups in the predetermined size determination dataset are a plurality of images including images with different content rates of particles belonging to the particle size class with the smallest particle size, and the content rate of each of the plurality of images in the data group is linked; the number of the unmachine-learned second machine learning models is the same as the number of the plurality of data groups and is provided corresponding to each of the plurality of data groups; the machine learning process, for each of the plurality of data groups in the predetermined training dataset, machine-learns an unmachine-learned second machine learning model corresponding to the data group using the plurality of images in the data group; and the output process, for each of the plurality of data groups in the predetermined size determination dataset, inputs the plurality of images in the data group into the second machine learning model machine-learned in the machine learning process using a data group in the predetermined training dataset that has the same image size as the image in the data group, thereby outputting the probability of each of the plurality of particle size classes output from the second machine learning model machine-learned in the machine learning process.

An image size determination assistance program according to another aspect of the present invention is a program that is executed by a computer and that assists in determining the size of an image to be used to determine a particle size class when particle sizes are divided into a plurality of different ranges as particle size classes. The program includes: a first data acquisition process for acquiring a predetermined training dataset; a second data acquisition process for acquiring a predetermined size determination dataset; a machine learning process for inputting an image and using the predetermined training dataset acquired in the first data acquisition process to create a second machine learning model that is not yet machine-learned and outputs, for each of the plurality of particle size classes, the probability that the input image will be classified into that particle size class; and an output process for inputting the predetermined size determination dataset acquired in the second data acquisition process into the second machine learning model that was machine-learned in the machine learning process, and outputting each probability for each of the plurality of particle size classes output from the second machine learning model that was machine-learned in the machine learning process. The predetermined training dataset and the predetermined size determination dataset each include a plurality of data groups each comprising image data for a plurality of images, the plurality of data groups having different image sizes, and each of the plurality of data groups in the predetermined training dataset includes multiple data groups each comprising image data for a plurality of images. the plurality of images are a plurality of images including an image of each of the plurality of particle size classes, and each of the plurality of images in the data group is linked to the particle size class of the image; in each of the plurality of data groups in the predetermined size determination dataset, the plurality of images in the data group are a plurality of images including images with different content rates of particles belonging to the particle size class with the smallest particle size, and each of the plurality of images in the data group is linked to the content rate of the image; the number of the unmachine-learned second machine learning models is the same as the number of the plurality of data groups and is provided corresponding to each of the plurality of data groups; the machine learning process, for each of the plurality of data groups in the predetermined training dataset, machine-learns an unmachine-learned second machine learning model corresponding to the data group using the plurality of images in the data group; and the output process, for each of the plurality of data groups in the predetermined size determination dataset, inputs the plurality of images in the data group into a second machine learning model machine-learned in the machine learning process using a data group in the predetermined training dataset having the same image size as the image in the data group, thereby outputting the probability of each of the plurality of particle size classes output from the second machine learning model machine-learned in the machine learning process.

Such image size determination support methods and programs output the probabilities for each of the multiple particle size classes obtained by inputting a predetermined size determination dataset into a machine-learned second learning model. A user can then determine the optimal image size from among multiple image sizes by referencing the output probabilities for each of the multiple particle size classes and the content associated with the image. Therefore, the image size determination support methods and programs described above can effectively assist users in determining image sizes. Therefore, these provide an image size determination support method and program that assists users in determining the size of an image to be used to determine particle size classes.

The particle size class determination device according to the present invention can determine the particle size class of a wider range of objects to be determined. According to the present invention, an image size determination support device, an image size determination support method, and an image size determination support program can be provided that support the determination of the size of an image used for particle size class determination.

1 is a block diagram showing the configuration of a particle size class determination device with an image size determination support function according to an embodiment. 10 is a schematic diagram for explaining a situation in which an image acquisition unit and an illumination unit in the particle size class determination device are arranged in a production line. FIG. FIG. 1 is a diagram illustrating a training dataset and a validation dataset. FIG. 3 is a diagram for explaining a machine learning model in the particle size class determination device. FIG. 10 is a diagram for explaining a size determination dataset. 10 is a flowchart showing the operation of the particle size class determination device regarding machine learning. 10 is a flowchart showing the operation of the particle size class determination device in relation to image size determination support. 10 is a graph showing the correlation between the powder rate determined using a sieve and the powder rate estimated using the particle size class determination device. 4 is a flowchart showing the operation of the particle size class determination device regarding particle size class determination. One example is an image in which a mark is used to indicate the particle size class corresponding to the particle size value of the determination target in the image to be determined.

One or more embodiments of the present invention will be described below with reference to the drawings. However, the scope of the invention is not limited to the disclosed embodiments. Components with the same reference numerals in each drawing are identical components, and their description will be omitted where appropriate. In this specification, generic reference numerals are used without subscripts to refer to components, and subscripted reference numerals are used to refer to individual components.

The particle size class determination device in one embodiment includes an image acquisition unit that acquires images of multiple particles, including particles of different particle sizes, and a particle size class processing unit that, when an image is input and particle sizes are divided into multiple different ranges as multiple particle size classes, calculates the probability of each of the multiple particle size classes for the image acquired by the image acquisition unit using a machine learning model (first machine learning model) that has been machine-learned and outputs, for each of the multiple particle size classes, the probability that the input image will be classified into that particle size class. One of the boundary values in the multiple ranges is a target particle size value for determination, and the size of the image input to the first machine learning model is the target particle size value multiplied by a predetermined constant. The image size determination support device in one embodiment is a device that supports the determination of the size of an image to be used for particle size class determination. This image size determination assistance device includes: a first data acquisition unit that acquires a predetermined training dataset; a second data acquisition unit that acquires a predetermined size determination dataset; a machine learning unit that uses the predetermined training dataset acquired by the first data acquisition unit to train a second machine learning model that, when an image is input and particle sizes are divided into a plurality of particle size classes with different particle sizes, outputs, for each of the plurality of particle size classes, a probability that the input image will be classified into that particle size class; and an output unit that inputs the predetermined size determination dataset acquired by the first data acquisition unit to the second machine learning model trained by the machine learning unit, and outputs each probability for each of the plurality of particle size classes output from the second machine learning model trained by the machine learning unit. The predetermined training dataset and the predetermined validation dataset each include a plurality of data groups each containing image data of a plurality of images, and the plurality of data groups have different image sizes. In each of the plurality of data groups in the predetermined training dataset, the plurality of images in the data group include a plurality of images each containing an image of the plurality of particle size classes, and each of the plurality of images in the data group is associated with a particle size class for the image. In each of the plurality of data groups in the predetermined size determination dataset, the plurality of images in the data group are a plurality of images including images with different content rates of particles belonging to a particle size class with a smallest particle size, and the content rate of each of the plurality of images in the data group is linked to the plurality of images in the data group. The number of the unmachine-learned second machine-learning models is the same as the number of the plurality of data groups, and the number is provided corresponding to each of the plurality of data groups. The machine learning unit trains, for each of the plurality of data groups in the predetermined training dataset, an unmachine-learned second machine-learning model corresponding to the data group using the plurality of images in the data group. The output unit inputs, for each of the plurality of data groups in the predetermined size determination dataset, the plurality of images in the data group to the second machine-learning model trained by the machine learning unit using a data group in the predetermined training dataset having the same image size as the image in the data group, and outputs the probability of each of the plurality of particle size classes output from the second machine-learning model trained by the machine learning unit. Below, we will explain in more detail the particle size class determination device, the particle size class determination method implemented therein, and the particle size class determination program implemented therein, as well as the image size determination support device, the image size determination support method implemented therein, and the image size determination support program implemented therein, using a particle size class determination device with image size determination support functions that integrates the particle size class determination device and image size determination support device.

Figure 1 is a block diagram showing the configuration of a particle size class determination device with an image size determination support function according to an embodiment. Figure 2 is a schematic diagram illustrating a situation in which the image acquisition unit and illumination unit of the particle size class determination device are arranged on a production line. Figure 3 is a diagram illustrating a training dataset and a validation dataset. Figure 3A shows, as an example, an image of multiple particles with a particle size of 2 mm or less. Figure 3B shows, as an example, an image of multiple particles with a particle size of more than 2 mm but not more than 2.8 mm. Figure 3C shows, as an example, an image of multiple particles with a particle size of more than 2.8 mm but not more than 4.75 mm. Figure 3D shows how the image shown in Figure 3A is divided into multiple parts. Figure 3E shows how the image shown in Figure 3B is divided into multiple parts. Figure 3F shows how the image shown in Figure 3C is divided into multiple parts. FIG. 3G shows, as an example, a group of images of multiple data groups divided by a first division size; FIG. 3H shows, as an example, a group of images of multiple data groups divided by a second division size; FIG. 3I shows, as an example, a group of images of multiple data groups divided by a third division size; and FIG. 3J shows, as an example, a group of images of multiple data groups divided by a fourth division size. FIG. 4 is a diagram illustrating a machine learning model in the particle size class determination device. FIG. 5 is a diagram illustrating a size determination dataset. FIG. 5A shows, as an example, multiple images of a data group with a 20% powder content; FIG. 5B shows, as an example, multiple images of a data group with a 30% powder content; FIG. 5C shows, as an example, multiple images of a data group with a 40% powder content; and FIG. 5D shows, as an example, multiple images of a data group with a 50% powder content.

In this embodiment, the particle size class determination device D includes, for example, an image acquisition unit 1, an illumination unit 2, a control processing unit 3, an input unit 4, an output unit 5, an interface unit (IF unit) 6, and a memory unit 7, as shown in FIG. 1.

The image acquisition unit 1 is connected to the control processing unit 3 and, under the control of the control processing unit 3, acquires images of multiple particles, including particles of different particle sizes. The particles may be any type of particle, but in this embodiment, the particles are pseudo-particles, for example, comprising a first particle serving as a nucleus and one or more second particles smaller than the first particle attached to the surface of the first particle. More specifically, the particles are sintered ore particles (sintered ore particles) comprising a nucleus powder (an example of a first particle) such as iron ore, limestone, or coke, and fine particles smaller than the powder (an example of a second particle) attached to the surface of the powder. Such sintered ore is preferably used in blast furnaces. In this embodiment, the image acquisition unit 1 is, for example, an imaging unit 1 that generates images during the sintering process to acquire images of the sintered ore particles after the sintering process. As shown in FIG. 2 , the imaging unit 1 is disposed above the belt conveyor BC that transports the sintered ore particles so as to obtain a bird's-eye view of the sintered ore particles on the belt conveyor BC. More specifically, the imaging unit 1 is disposed directly above the belt conveyor BC, with its optical axis aligned with the normal direction of the conveying surface (the surface on which the sintered ore particles are placed) of the belt conveyor BC. The imaging unit 1 preferably has an angle of view that allows it to capture the entire width direction (direction perpendicular to the conveying direction) of the belt conveyor BC. By capturing images of the sintered ore particles after the sintering process and determining their particle size class, the sintering conditions of the sintering process can be adjusted according to the determination results, optimizing the sintering process and efficiently performing it. Such an imaging unit 1 is, for example, a digital camera equipped with an imaging optical system that forms an optical image of the imaging subject on a predetermined imaging plane, an area image sensor that is positioned so that its light-receiving surface is aligned with the imaging plane and converts the optical image of the imaging subject into an electrical signal, and an image processing unit that processes the output of the area image sensor to generate image data that represents the image of the imaging subject.

Note that the image acquisition unit 1 is not limited to the imaging unit 1 and may be another device. For example, the image acquisition unit 1 is an interface circuit that inputs and outputs data to and from an external device. The external device is a storage medium, such as a USB (Universal Serial Bus) memory or an SD card (registered trademark), that stores images of multiple particles, including particles of different particle sizes. Alternatively, the external device may be a drive device that reads data from a recording medium on which the image is recorded, such as a CD-ROM (Compact Disc Read Only Memory), a CD-R (Compact Disc Recordable), a DVD-ROM (Digital Versatile Disc Read Only Memory), or a DVD-R (Digital Versatile Disc Recordable). The interface circuit serving as the image acquisition unit 1 may be connected to the external device by wire or wirelessly. Alternatively, the image acquisition unit 1 may be, for example, a communication interface circuit that sends and receives communication signals to and from an external device, and the external device may be a server device that manages the images and is connected to the communication interface circuit via a network (WAN (Wide Area Network, including a public communication network)) or LAN (Local Area Network). With this type of image acquisition unit 1, the particle size class of the image can be determined even when production is not in progress, and past sintering processes can be verified. Here, if the image acquisition unit 1 is an interface circuit or a communication interface circuit, the image acquisition unit 1 may also serve as the IF unit 6 (i.e., the IF unit 6 may be used as the image acquisition unit 1).

The illumination unit 2 is connected to the control processing unit 3 and illuminates the plurality of particles to be evaluated under the control of the control processing unit 3. In this embodiment, as shown in FIG. 2, the illumination unit 2 includes four illumination units, first through fourth, 2-1 through 2-4, which surround the imaging unit 1 on all four sides and are arranged so as to illuminate each sintered ore particle on the belt conveyor BC. More specifically, if a virtual square is defined on a virtual plane parallel to the conveying surface of the belt conveyor BC, the imaging unit 1 is arranged so that its optical axis passes through the center of the square (the intersection of two diagonals), and the first through fourth illumination units 2-1 through 2-4 are arranged, one at each of the four corners of the square, so that their optical axes pass through the intersections where the optical axis of the imaging unit 1 intersects with the conveying surface of the belt conveyor BC. The illumination unit 2 may be turned on and off manually.

The input unit 4 is connected to the control processing unit 3 and is a device that inputs various commands, such as commands to start assistance and commands to start judgment, as well as various data required to operate the particle size class judgment device D with image size determination assistance function, such as the image size determined as a result of the assistance, to the particle size class judgment device D. This device may be, for example, multiple input switches assigned with specified functions, a keyboard, or a mouse. The output unit 5 is connected to the control processing unit 3 and is a device that outputs the commands and data input from the input unit 4 and the particle size class of the judgment results under the control of the control processing unit 3. This device may be, for example, a display device such as a CRT display, LCD (liquid crystal display), or organic EL display, or a printing device such as a printer.

The input unit 4 and output unit 5 may be configured as a touch panel. In this touch panel configuration, the input unit 4 is a position input device, such as a resistive or capacitive type, that detects and inputs an operation position, and the output unit 5 is a display device. In this touch panel, a position input device is provided on the display surface of the display device, and one or more input content candidates that can be input to the display device are displayed. When a user touches the display position showing the input content they want to input, the position is detected by the position input device, and the display content displayed at the detected position is input to the particle size class determination device D as the user's operation input content. Such a touch panel makes it easy for users to intuitively understand input operations, providing a particle size class determination device D that is easy for users to use.

The IF unit 6 is connected to the control processing unit 3 and is a circuit that inputs and outputs data to and from external devices under the control of the control processing unit 3. For example, it may be an interface circuit for the RS-232C serial communication method, an interface circuit using the Bluetooth (registered trademark) standard, or an interface circuit using the USB standard. The IF unit 6 may also be a communications interface circuit that transmits and receives communications signals to and from external devices, such as a data communications card or a communications interface circuit conforming to the IEEE 802.11 standard.

The input unit 4 accepts and acquires input of a predetermined training dataset and a validation dataset, which will be described later, and accepts and acquires input of a predetermined sizing dataset, which will be described later. The input unit 4 corresponds to an example of a first data acquisition unit that acquires a predetermined training dataset, and corresponds to an example of a second data acquisition unit that acquires a predetermined sizing dataset. Alternatively, the IF unit 6 acquires the predetermined training dataset from a storage medium that stores the predetermined training dataset, a recording medium that records the predetermined training dataset, or a server device that manages the predetermined training dataset. The IF unit 6 acquires the predetermined sizing dataset from a storage medium that stores the predetermined sizing dataset, a recording medium that records the predetermined sizing dataset, or a server device that manages the predetermined sizing dataset. The IF unit 6 corresponds to another example of the first data acquisition unit and another example of the second data acquisition unit.

The memory unit 7 is connected to the control processing unit 3 and is a circuit that stores various specified programs and various specified data in accordance with the control of the control processing unit 3. The various predetermined programs include, for example, a control processing program, which includes, for example, a control program that controls each of the units 1, 2, 4 to 7 of the particle size class determination device D according to the function of each unit; a machine learning program that uses a predetermined learning dataset acquired by the first data acquisition unit (input unit 4 and IF unit 6 in this embodiment) to train a second machine learning model that is not yet machine-learned and that outputs, for each of the plurality of particle size classes, the probability that the input image will be classified into that particle size class when an image is input and particle sizes are divided into a plurality of different ranges; a particle size class processing program that uses a machine-learned machine learning model (first machine learning model) that is input and outputs, for each of the plurality of particle size classes, the probability that the input image will be classified into that particle size class when an image is input and particle sizes are divided into a plurality of different ranges; and an image size adjustment program that adjusts the image acquired by the image acquisition unit 1 to the size of the image accepted by the input unit 4. In this embodiment, the first machine-learned model is a second machine-learned model trained by the machine learning program. The various types of predetermined data include, for example, images acquired by the image acquisition unit 1, a training dataset, a validation dataset and a size determination dataset (described below), and other data required to execute each program. The storage unit 7 includes, for example, a non-volatile storage element such as a ROM (Read Only Memory) or a rewritable non-volatile storage element such as an EEPROM (Electrically Erasable Programmable Read Only Memory). The storage unit 7 also includes a RAM (Random Access Memory), which serves as the working memory of the control processing unit 3 and stores data generated during execution of the predetermined program. The storage unit 7 may also be configured with a hard disk drive with a relatively large storage capacity.

The control processing unit 3 is a circuit that controls each of the units 1, 2, 4-7 of the particle size class determination device D according to the function of each unit, determines the particle size class of a particle based on an image of multiple particles including particles of different particle sizes, and assists in determining the size of the image used to determine the particle size class. The control processing unit 3 is configured, for example, with a CPU (Central Processing Unit) and its peripheral circuits. By executing a control processing program, the control processing unit 3 functionally comprises a control unit 31, a machine learning unit 32, a particle size class processing unit 33, and an image size adjustment unit 34.

The control unit 11 controls each of the units 1, 2, 4-7 of the particle size class determination device D according to the function of each unit, and is responsible for the overall control of the particle size class determination device D.

The machine learning unit 32 uses a predetermined learning dataset acquired by the first data acquisition unit (in this embodiment, the input unit 4 or IF unit 6) to train a second machine learning model that, when an image is input and particle sizes are divided into multiple particle size classes with different ranges, outputs the probability that the input image will be classified into each of the multiple particle size classes.

The predetermined training dataset comprises multiple data groups each containing image data of multiple images. The multiple data groups have different image sizes. In each of the multiple data groups in the predetermined training dataset, the multiple images in the data group are multiple images including images of each of the multiple particle size classes, and each of the multiple images in the data group is linked to the particle size class of the image. Therefore, the machine learning unit 32 performs supervised machine learning of the untrained second machine learning model as a classification problem.

The number of the untrained second machine learning models is the same as the number of the multiple data groups, and is provided for each of the multiple data groups. For each of the multiple data groups in the specified training dataset, the machine learning unit 32 trains an untrained second machine learning model corresponding to that data group using multiple images in that data group.

In this embodiment, as described above, the particles are sintered ore particles after the sintering process, and the particle size class determination device D determines the particle size class based on images taken while the particles are being transported on the belt conveyor BC. According to the inventor's knowledge, multiple sintered ore particles on this belt conveyor BC exist locally with approximately the same particle size. In this embodiment, each sintered ore particle that is locally present with approximately the same particle size is classified into one of three particle size classes: large, medium, and small. In this embodiment, the objective is to detect sintered ore particles that do not exceed 2 mm in size in order to operate the sintering process efficiently. For this reason, in this embodiment, the three particle size classes are determined by taking into account the particle sizes typically observed in the sintered ore particles: a particle size class (small particle size class) of particles with a particle size of 2 mm or less (i.e., sintered ore particles not exceeding 2 mm, small particle size particles) that can be separated using a sieve according to JIS standards (JIS Z8801-1); a particle size class (medium particle size class) of particles with a particle size of more than 2 mm and not more than 2.8 mm (medium particle size class); and a particle size class (large particle size class) of particles with a particle size of more than 2.8 mm and not more than 4.75 mm (large particle size particles).

According to the inventor's findings, when particle size classes are set to three particle size classes (large, medium, and small), if the image size is too small, an image capturing large particle sizes will not capture the entire particle, resulting in a loss of particle outline, making it difficult to properly extract the particles. In particular, if the particles are pseudo-particles of sintered ore, there is a risk that fine powder adhering to the core powder will be extracted, making it difficult to extract the actual pseudo-particles of sintered ore. On the other hand, if the image size is too large, when large and small particles are captured in the image, the outline information of the small particles may be hidden, making proper extraction difficult. For this reason, it is believed that there is an optimal solution for the image size (image size) relative to the target particle size value. In this embodiment, the objective is to detect sintered ore particles that do not exceed 2 mm, so the target particle size value is 2 mm.

For this reason, the image sizes (image sizes) of the multiple data groups in the predetermined training dataset are set to be different from each other by a constant multiple of the target particle size value. Specifically, the multiple data groups in the predetermined training dataset are four data groups, first to fourth, as a means of optimal solution search, and the image size (first image size) of the first data group is set to 12 times the target particle size value, the image size (second image size) of the second data group is set to 10 times the target particle size value, the image size (third image size) of the third data group is set to 7 times the target particle size value, and the image size (fourth image size) of the fourth data group is set to 5 times the target particle size value.

More specifically, the predetermined training dataset was created as follows. In this embodiment, a predetermined validation dataset was also created along with the predetermined training dataset. The predetermined validation dataset has the same configuration as the training dataset, and is data used to verify whether the machine learning model has been trained appropriately (i.e., whether it can produce correct answers).

First, the sintered ore was sieved through a JIS standard sieve to prepare aggregates of small-sized sintered ore particles (small-sized sintered ore aggregates), aggregates of medium-sized sintered ore particles (medium-sized sintered ore aggregates), and aggregates of large-sized sintered ore particles (large-sized sintered ore aggregates).

Next, the small-particle sinter aggregates were spread out on a flat surface and photographed from directly above with a digital camera, generating a 2456 x 2054 pixel image (small-particle image), as shown in Figure 3A. The same process was repeated multiple times to generate multiple small-particle images. The same process was performed on both the medium-particle sinter aggregates and the large-particle sinter aggregates, generating multiple images of the medium-particle sinter aggregates (medium-particle images) and multiple images of the large-particle sinter aggregates (large-particle images). Figures 3B and 3C show the medium-particle image and the large-particle image, respectively.

Next, to make the small grain size images approximately 12 times larger (in this embodiment, 1 pixel is 0.06 mm, so for example, as shown in Figure 3D, the images were divided into 411 x 375 pixel sizes), and 2,842 images (small grain size divided images of the first image size) were generated from the multiple small grain size images. Of these 2,842 small grain size divided images of the first image size, 2,085 were designated as image data belonging to the first data group in the training dataset, and each of these image data was associated with a small grain size class. Of the 2,842 small grain size divided images of the first image size, 784 were designated as image data belonging to the first data group in the validation dataset, and each of these image data was associated with a small grain size class. For example, as shown in Figures 3E and 3F, similar operations were performed on each of the medium-sized images and the large-sized images, and 1,813 images (medium-sized divided images of the first image size) were generated by dividing the medium-sized images from the plurality of medium-sized images, and 1,960 images (large-sized divided images of the first image size) were generated by dividing the large-sized images from the plurality of large-sized images. Of these 1,813 medium-sized divided images of the first image size, 1,372 were designated as image data belonging to the first data group in the training dataset, and the remaining 441 were designated as image data belonging to the first data group in the verification dataset, and each of these image data was associated with a medium-sized class. Of these 1,960 large-sized divided images of the first image size, 1,372 were designated as image data belonging to the first data group in the training dataset, and the remaining 588 were designated as image data belonging to the first data group in the verification dataset, and each of these image data was associated with a large-sized class. As a result, for example, the first data group in the training dataset shown in Figure 3G was generated, and the first data group in the validation dataset (not shown) was generated.

Next, the same process was performed on each of the multiple small-grain images, the multiple medium-grain images, and the multiple large-grain images, but at a size of 319 x 290 pixels, resulting in approximately ten times the size. As a result, for example, the second data group in the training dataset shown in Figure 3H was generated, and the second data group in the validation dataset (not shown) was generated. The number of small-grain divided images of the second image size (images obtained by dividing a small-grain image at the second image size), medium-grain divided images of the second image size (images obtained by dividing a medium-grain image at the second image size), and large-grain divided images of the second image size (images obtained by dividing a large-grain image at the second image size) in the second data group of the training dataset was 2,058, 1,127, and 1,372, respectively. The number of small-grain divided images of the second image size, medium-grain divided images of the second image size, and large-grain divided images of the second image size in the second data group of the validation dataset was 784, 411, and 588, respectively.

Next, the same process was performed on each of the multiple small-grain images, the multiple medium-grain images, and the multiple large-grain images, but at a size of 237 x 216 pixels, resulting in approximately seven times the size. As a result, for example, the third data group in the training dataset shown in Figure 3I was generated, and the third data group in the validation dataset (not shown) was generated. The number of small-grain divided images of the third image size (images obtained by dividing a small-grain image at the third image size), medium-grain divided images of the third image size (images obtained by dividing a medium-grain image at the third image size), and large-grain divided images of the third image size (images obtained by dividing a large-grain image at the third image size) in the third data group of the training dataset was 3,402, 1,863, and 2,268, respectively. The number of small-grain divided images of the third image size, medium-grain divided images of the third image size, and large-grain divided images of the third image size in the third data group of the validation dataset was 1,296, 972, and 729, respectively.

Next, the same process was performed on each of the multiple small-grain images, the multiple medium-grain images, and the multiple large-grain images, but at a size of 184 x 167 ixel, approximately five times larger. This resulted in the fourth data group in the training dataset shown in Figure 3J, for example, and the fourth data group in the validation dataset (not shown). The number of small-grain divided images with the fourth image size (images obtained by dividing a small-grain image by the fourth image size), medium-grain divided images with the fourth image size (images obtained by dividing a medium-grain image by the fourth image size), and large-grain divided images with the fourth image size (images obtained by dividing a large-grain image by the fourth image size) in the fourth data group of the training dataset was 6,048, 3,312, and 4,032, respectively. The number of small-grain divided images with the fourth image size, medium-grain divided images with the fourth image size, and large-grain divided images with the fourth image size in the fourth data group of the validation dataset was 2,304, 1,296, and 1,728, respectively.

In this embodiment, the second unmachined machine learning model comprises four models: an unmachined second machine learning model according to the first data group (an unmachined second machine learning model for the first image size), an unmachined second machine learning model according to the second data group (an unmachined second machine learning model for the second image size), an unmachined second machine learning model according to the third data group (an unmachined second machine learning model for the third image size), and an unmachined second machine learning model according to the fourth data group (an unmachined second machine learning model for the fourth image size). The machine learning unit 32 performs supervised machine learning of an unmachine-learned second machine learning model for the first image size using a plurality of images of a first data group in the training dataset (a plurality of small particle size divided images of the first image size, a plurality of medium particle size divided images of the first image size, and a plurality of large particle size divided images of the first image size); performs supervised machine learning of an unmachine-learned second machine learning model for the second image size using a second data group in the training dataset (a plurality of small particle size divided images of the second image size, a plurality of medium particle size divided images of the second image size, and a plurality of large particle size divided images of the second image size); performs supervised machine learning of an unmachine-learned second machine learning model for the third image size using a third data group in the training dataset (a plurality of small particle size divided images of the third image size, a plurality of medium particle size divided images of the third image size, and a plurality of large particle size divided images of the third image size); and performs supervised machine learning of an unmachine-learned second machine learning model for the fourth image size using a fourth data group in the training dataset (a plurality of small particle size divided images of the fourth image size, a plurality of medium particle size divided images of the fourth image size, and a plurality of large particle size divided images of the fourth image size). This generates a second machine learning model that has been machine-trained for the first data group (first image size), a second machine learning model that has been machine-trained for the second data group (second image size), a second machine learning model that has been machine-trained for the third data group (third image size), and a second machine learning model that has been machine-trained for the fourth data group (fourth image size).

In this embodiment, the machine learning model used is a well-known deep learning convolutional neural network (CNN).

This CNN 8 is a CNN with the function of avoiding overfitting, as shown in Figure 4, and includes multiple layers of pre-processing units 81 (81-1 to 81-6) and a post-processing unit 82.

In this embodiment, the CNN 8 has six layers of first to sixth preprocessing units 81-1 to 81-6. These first to sixth preprocessing units 81-1 to 81-6 do not necessarily have the same number of input channels and output channels, but they have the same structure and include convolutional layers 811 (811-1 to 811-6), pooling layers 812 (812-1 to 812-6), and dropout layers 813 (813-1 to 813-6). Each convolutional layer 811 (811-1 to 811-6) performs a convolution operation on the input image of that layer and extracts features of the input image. Each pooling layer 812 (812-1 to 812-6) is connected to a convolutional layer 811 (811-1 to 811-6) and reduces the size of the output (feature map) of the convolutional layer 811 (811-1 to 811-6) through a predetermined process. This predetermined process typically uses max pooling. For example, when reducing a 12x12 feature map to a 3x3 size, max pooling divides the 12x12 feature map into nine 4x4 regions, extracts the maximum value in each region, and sets it as each value of the 3x3 size. Note that, instead of the maximum value, average pooling, which calculates the average value of the region, may be used. Each dropout layer 813 (813-1 to 813-6) is connected to a pooling layer 812 (812-1 to 812-6) and prevents overfitting. The dropout layer 813 includes multiple nodes, and each time the machine learning is updated, one of the multiple nodes is randomly selected with a certain probability and deactivated (disabled). For example, the first preprocessing unit 81-1 has 16 output channels and is connected to the second preprocessing unit 81-2 via the 16 channels. The dropout layer 813 includes 16 nodes, and these 16 nodes are randomly deactivated each time the machine learning is updated. In the example shown in FIG. 4, a dropout layer 813 (813-1 to 813-6) is provided in each of the preprocessing units 81-1 to 81-6, but some preprocessing units 81 may not include a dropout layer 813. The first preprocessing unit 81-1 has one input channel and 16 output channels. The second preprocessing unit 81-2 has 16 input channels and 32 output channels. The third preprocessing unit 81-3 has 32 input channels and 64 output channels. The fourth preprocessing unit 81-4 has 64 input channels and 64 output channels. The fifth preprocessing unit 81-5 has 64 input channels and 128 output channels. The sixth preprocessing unit 81-6 has 128 input channels and 64 output channels.

The post-processing unit 82 is a neural network comprising a fully connected layer 821 and an output layer 822, and uses the neural network to recognize the input image input to the pre-processing unit 81 based on the processing results of the pre-processing unit 81. In this embodiment, an image of multiple particles is input as the input image, and the probabilities of the small particle size class, medium particle size class, and large particle size class for the image (the input image) are output.

Returning to FIG. 1 , the particle size class processing unit 33 receives an input image and divides particle sizes into multiple particle size classes into multiple different ranges. The particle size class processing unit 33 calculates the probability of each of the multiple particle size classes for the image acquired by the image acquisition unit 1 using a machine-learned first machine learning model that outputs the probability of the input image being classified into that particle size class for each of the multiple particle size classes. The control unit 31 outputs the probability of each of the multiple classes for the image calculated by the particle size class processing unit 33 to the output unit 5. The particle size class determination device D in this embodiment is equipped with an image size determination assistance function as described above, and the particle size class processing unit 33 inputs a predetermined size determination dataset acquired by the second data acquisition unit into a machine-learned second machine learning model trained by the machine learning unit 32, thereby calculating the probability of each of the multiple classes for each image in the predetermined size determination dataset. The control unit 31 outputs the probability of each of the multiple classes for each image in the predetermined size determination dataset calculated by the particle size class processing unit 33 to the output unit 5. The second machine learning model that has been machine-learned by the machine learning unit 32 is used as the first machine learning model.

In this embodiment, as described above, the probability of the small particle size class (small particle size probability), the probability of the medium particle size class (medium particle size probability), and the probability of the large particle size class (large particle size probability) for the image are calculated and output.

The size determination dataset includes multiple data groups each containing image data for multiple images. The multiple data groups have different image sizes. In each of the multiple data groups in the specified size determination dataset, the multiple images in the data group include images with different content rates of particles belonging to the smallest particle size class, and the content rate of each image is linked to each of the multiple images in the data group.

The particle size class processing unit 33 then inputs, for each of the multiple data groups in the predetermined size determination dataset, multiple images from the data group into a second machine learning model trained by the machine learning unit 32 using a data group in the predetermined training dataset that has the same image size as the image in the data group, and thereby calculates the probability of each of the multiple particle size classes output from the second machine learning model trained by the machine learning unit 32. The control unit 31 outputs to the output unit 5 the probability of each of the multiple classes for each image in the predetermined size determination dataset that has been calculated for each of the multiple data groups by the particle size class processing unit 33.

In this embodiment, the particle size classes are the small particle size class, medium particle size class, and large particle size class, as described above, and the particle size class with the smallest particle size is the small particle size class. Particles belonging to the smallest particle size class have a particle size of 2 mm or less when sieved through a JIS standard sieve, and the content of particles belonging to the smallest particle size class is the content of particles with a particle size of 2 mm or less. In this embodiment, particles with a particle size of 2 mm or less are defined as "powder," and the content of "powder" in the measurement object (the ratio of powder to the entire measurement object) is defined as the "powder content."

The image sizes of the multiple data groups in the predetermined size determination dataset are the same as the image sizes of the multiple data groups in the predetermined training dataset. In the above example, the multiple data groups in the predetermined size determination dataset are four data groups, first to fourth, and the image size of the first data group (first image size) is set to 12 times the target particle size value, the image size of the second data group (second image size) is set to 10 times the target particle size value, the image size of the third data group (third image size) is set to 7 times the target particle size value, and the image size of the fourth data group (fourth image size) is set to 5 times the target particle size value.

To determine the image size, different dust ratios were set to 20%, 30%, 40%, and 50%.

More specifically, the specified size determination dataset was created as follows:

First, by using a JIS standard sieve to sieve the sintered ore, two aggregates each of sintered ore particle aggregates with a fineness of 20% (20% fineness sintered ore aggregate), sintered ore particle aggregates with a fineness of 30% (30% fineness sintered ore aggregate), sintered ore particle aggregates with a fineness of 40% (40% fineness sintered ore aggregate), and sintered ore particle aggregates with a fineness of 50% (50% fineness sintered ore aggregate) were prepared.

Next, for each of the two 20% fine sinter aggregates, a process similar to the process for generating small particle size images for the specified training dataset described above was performed to generate an image of sintered ore with a 20% fineness (20% fineness image). Then, a process similar to the process for generating small particle size divided images for the specified training dataset described above was performed to generate images obtained by dividing the 20% fineness image into a first image size (first image size 20% fineness divided image), images obtained by dividing the 20% fineness image into a second image size (second image size 20% fineness divided image), images obtained by dividing the 20% fineness image into a third image size (third image size 20% fineness divided image), and images obtained by dividing the 20% fineness image into a fourth image size (fourth image size 20% fineness divided image). The actual fineness of each of the two 20% fineness sinter aggregates prepared above was separately measured using a JIS standard sieve test. For one sintered ore aggregate with a fineness of 20%, 50 divided images of each of the first to fourth image sizes with a fineness of 20% were generated. An example is shown in Figure 5A.

Next, the same procedure as above was performed on each of the two 30% fineness sinter aggregates, and 50 images each were generated: a first image size 30% fineness split image, a second image size 30% fineness split image, a third image size 30% fineness split image, and a fourth image size 30% fineness split image. The fineness of each of the two 30% fineness sinter aggregates was then measured. An example is shown in Figure 5B.

Next, the same procedure as above was performed on each of the two 40% fineness sinter aggregates, and 50 images each were generated: a first image size 40% fineness split image, a second image size 40% fineness split image, a third image size 40% fineness split image, and a fourth image size 40% fineness split image, and the fineness of each of the two 40% fineness sinter aggregates was measured. An example is shown in Figure 5C.

Next, the same procedure as above was performed on each of the two 50% fineness sinter aggregates, and 50 images each were generated: a first image size 50% fineness split image, a second image size 50% fineness split image, a third image size 50% fineness split image, and a fourth image size 50% fineness split image, and the fineness of each of the two 50% fineness sinter aggregates was measured. An example is shown in Figure 5D.

Next, the first image size divided image with a 20% powder rate, the first image size divided image with a 30% powder rate, the first image size divided image with a 40% powder rate, and the first image size divided image with a 50% powder rate are considered to be image data belonging to the first data group in the size determination dataset, and the second image size divided image with a 20% powder rate, the second image size divided image with a 30% powder rate, the second image size divided image with a 40% powder rate, and the second image size divided image with a 50% powder rate are considered to be image data belonging to the second data group in the size determination dataset. The third image size 20% fineness split image, the third image size 30% fineness split image, the third image size 40% fineness split image, and the third image size 50% fineness split image were designated as image data belonging to the third data group in the size determination dataset, and the fourth image size 20% fineness split image, the fourth image size 30% fineness split image, the fourth image size 40% fineness split image, and the fourth image size 50% fineness split image were designated as image data belonging to the fourth data group in the size determination dataset. In this embodiment, two each of the 20% fineness sintered ore aggregates, the 30% fineness sintered ore aggregates, the 40% fineness sintered ore aggregates, and the 50% fineness sintered ore aggregates were prepared, and accordingly, two size determination datasets were also generated.

The particle size class processing unit 33 inputs a plurality of images in the first data group in the predetermined size determination dataset into a machine-learned second machine-learning model for a first image size, thereby determining the probability of each of the plurality of particle size classes output from the machine-learned machine-learning model for the first image size, and the control unit 31 outputs each of these probabilities to the output unit 5. The particle size class processing unit 33 inputs a plurality of images in the second data group in the predetermined size determination dataset into a machine-learned second machine-learning model for a second image size, thereby determining the probability of each of the plurality of particle size classes output from the machine-learned machine-learning model for the second image size, and the control unit 31 outputs each of these probabilities to the output unit 5. The particle size class processing unit 33 inputs a plurality of images in the third data group in the predetermined size determination dataset into a machine-learned second machine-learning model for a third image size, thereby determining the probability of each of the plurality of particle size classes output from the machine-learned machine-learning model for the third image size, and the control unit 31 outputs each of these probabilities to the output unit 5. The particle size class processing unit 33 inputs multiple images in the fourth data group in the predetermined size determination dataset into a second machine-learned machine learning model for the fourth image size, thereby determining the probability of each of the multiple particle size classes output from the machine-learned machine learning model for the fourth image size, and the control unit 31 outputs these probabilities to the output unit 5.

The image size adjustment unit 34 adjusts the image acquired by the image acquisition unit 1 to the image size accepted by the input unit 4. The image size adjustment unit 34 may, for example, generate a single image of the image size accepted by the input unit 4 by cutting out an image area of the image size accepted by the input unit 4 (e.g., an image area in the center of the image) from a single image acquired by the image acquisition unit 1. Alternatively, the image size adjustment unit 34 may, for example, generate multiple images of the image size accepted by the input unit 4 by cutting out multiple discontinuous image areas of the image size accepted by the input unit 4 from a single image acquired by the image acquisition unit 1. In this embodiment, for example, the image size adjustment unit 34 generates multiple images of the image size accepted by the input unit 4 by dividing a single image acquired by the image acquisition unit 1 into a two-dimensional matrix by the image areas of the image size accepted by the input unit 4. Then, for each of the multiple images generated by division in the image size adjustment unit 34, the particle size class processing unit 33 inputs the image into the machine-learned second machine learning model corresponding to the size of the image received by the input unit 4 as the machine-learned first machine learning model, thereby determining the probability of each of the multiple particle size classes output from the machine-learned second machine learning model, and the control unit 31 outputs these probabilities to the output unit 5.

The control processing unit 3, input unit 4, output unit 5, IF unit 6, and memory unit 7 can be configured, for example, by a desktop or notebook computer.

Next, the operation of this embodiment will be described. FIG. 6 is a flowchart showing the operation of the particle size class determination device related to machine learning. FIG. 7 is a flowchart showing the operation of the particle size class determination device related to image size determination support. FIG. 8 is a graph showing the correlation between the powder rate determined using a sieve and the powder rate estimated using the particle size class determination device. FIG. 9 is a flowchart showing the operation of the particle size class determination device related to particle size class determination. FIG. 10 is an example of an image to be determined, in which the particle size class corresponding to the target particle size value is indicated by a mark. FIGS. 10A to 10D show the image after determination, FIG. 10E shows the image before determination corresponding to FIG. 10A, FIG. 10F shows the image before determination corresponding to FIG. 10B, FIG. 10G shows the image before determination corresponding to FIG. 10C, and FIG. 10H shows the image before determination corresponding to FIG. 10D.

When the power is turned on, the particle size class determination device D with image size determination support function configured as described above initializes each necessary unit and begins operation. By executing the control processing program, the control processing unit 3 functionally configures a control unit 31, machine learning unit 32, particle size class processing unit 33, and image size adjustment unit 34.

In machine learning using the second machine learning model, the user (operator) inputs a predetermined training dataset to the particle size class determination device D with image size determination support function via the input unit 4 or IF unit 6. In the above example, a predetermined training dataset including first to fourth data groups corresponding to the first to fourth image sizes, respectively, is input to the particle size class determination device D.

In Figure 6, first, the particle size class determination device D acquires the predetermined learning data set input via the input unit 4 or IF unit 6 using the control unit 31 of the control processing unit 3, and stores it in the memory unit 7 (S11).

Next, the particle size class determination device D uses the machine learning unit 32 of the control processing unit 3 to train a second machine learning model (not yet trained) for one data group in the predetermined training dataset, using multiple images (image data) in that data group (S12). For example, a second machine learning model (not yet trained) for the first data group is trained using multiple images in a first data group (first image size).

Next, when the machine learning is completed, the particle size class determination device D causes the machine learning unit 32 to store the machine-learned second machine learning model trained for the data group in the memory unit 7 in association with (linked to) the data group (S13). In the above example, the machine-learned second machine learning model trained for the first data group is stored in the memory unit 7 in association with the first data group (e.g., an identifier for specifying and identifying the first data group (data group ID, image size ID)). The machine-learned second machine learning model may be verified using a verification dataset.

Next, the particle size class determination device D uses the machine learning unit 32 to determine whether or not each of steps S12 and S13 has been completed for all data groups (all image sizes) in the specified learning dataset (S14). If the result of this determination is that each of the steps has not been completed for all data groups (No), the particle size class determination device D returns to step S12 to execute each of the steps for the next data group. On the other hand, if the result of the determination is that each of the steps has been completed for all data groups (Yes), the particle size class determination device D terminates this process. Note that, if necessary, each machine-learned machine learning model corresponding to each data group may be output from the IF unit 6 to an external device.

As a result, in the above example, a machine-trained second machine learning model is generated for the first data group (for the first image size), a machine-trained second machine learning model is generated for the second data group (for the second image size), a machine-trained second machine learning model is generated for the third data group (for the third image size), and a machine-trained second machine learning model is generated for the fourth data group (for the fourth image size), and these are stored in memory unit 7.

In terms of machine learning, the particle size class determination device D operates as follows:

To assist in determining image size, after generating a second machine learning model that has undergone machine learning, the user (operator) inputs a predetermined size determination dataset to the particle size class determination device D with image size determination assistance function via the input unit 4 or IF unit 6. In the above example, a predetermined size determination dataset comprising first to fourth data groups corresponding to the first to fourth image sizes, respectively, is input to the particle size class determination device D.

In Figure 7, first, the particle size class determination device D acquires the predetermined size determination data set input via the input unit 4 or IF unit 6 using the control unit 31 of the control processing unit 3, and stores it in the memory unit 7 (S21).

Next, the particle size class determination device D, by using the particle size class processing unit 33 of the control processing unit 3, determines the probability of each of the plurality of particle size classes by applying the plurality of images (image data) in one data group in the predetermined size determination dataset to a second machine learning model trained by the machine learning unit 32 as described above using a data group in the predetermined training dataset with the same image size as the image in the data group, and stores these probabilities in the memory unit 7 (S22). Each probability is associated with the data group (data group ID, image size ID), the image (input image) input into the machine-learned second machine learning model to determine this probability (an identifier (image ID) for specifying and identifying the image), and the content rate associated with this input image. For example, for each of the plurality of images (image data) in the first data group (first image size), the image is input into the machine-learned second machine learning model for the first data group, and the small particle size probability, medium particle size probability, and large particle size probability for that image are determined. These probabilities are then stored in the memory unit 7 in association with the data group ID, image ID, and powder content.

Next, the particle size class determination device D uses the particle size class processing unit 33 to determine whether process S22 has been completed for all data groups (all image sizes) in the specified size determination dataset (S23). If the result of this determination is that process S22 has not been completed for all data groups (No), the particle size class determination device D returns to process S22 to execute process S22 for the next data group. On the other hand, if the result of the determination is that process S22 has been completed for all data groups (Yes), the particle size class determination device D then executes process S24.

In process S24, the particle size class determination device D causes the control unit 31 of the control processing unit 3 to output to the output unit 5 each of the multiple probabilities calculated in process S22 and stored in the memory unit 7, along with the associated data group (data group ID, image size ID), input image (image ID), and content rate (powder rate in the above example), and then terminates this process. If necessary, each of these multiple probabilities may be output from the IF unit 6 to an external device together with the data group (data group ID, image size ID), etc.

To assist in determining image size, the particle size class determination device D operates in this manner.

In this embodiment, as described above, two size determination datasets were prepared, and the above-described processes were performed on each size determination dataset, resulting in the respective probabilities for each size determination dataset. The results are shown in Figure 8. The horizontal axis of Figure 8 is the powder fraction (content of small particle diameters) (mass %) measured using a JIS-standard sieve. The vertical axis of Figure 8 is the powder fraction (area %) estimated using the size determination dataset in the particle diameter class determination device D, which is the average value of the small particle diameter probability among the small particle diameter probability, medium particle diameter probability, and large particle diameter probability obtained by the particle diameter class determination device D through the above-described processes S21 to S24. For example, in the first image size, the horizontal axis represents the fineness ratio measured using a JIS sieve for a sintered ore aggregate with a fineness ratio of 20% used to generate one of the two size determination data sets. For each of the 50 images in the first data group generated from this sintered ore aggregate with a fineness ratio of 20%, a small particle size probability, a medium particle size probability, and a large particle size probability for that image are calculated using the second machine learning model for the first data group. The vertical axis represents the average of the small particle size probabilities calculated for each of the 50 images. For example, in the third image size, the horizontal axis represents the fineness ratio measured using a JIS sieve for a sintered ore aggregate with a fineness ratio of 40% used to generate one of the two size determination data sets. For each of the 50 images in the third data group generated from this sintered ore aggregate with a fineness ratio of 40%, a small particle size probability, a medium particle size probability, and a large particle size probability for that image are calculated using the second machine learning model for the third data group. The vertical axis represents the average of the small particle size probabilities calculated for each of the 50 images. Results for the first image size (first data group) are indicated with ◆, results for the second image size (second data group) are indicated with ●, results for the third image size (third data group) are indicated with ×, and results for the fourth image size (fourth data group) are indicated with ▲.

As can be seen from the results for the third image size indicated by an x and the fourth size indicated by a ▲ in Figure 8, the average small particle size probability determined by particle size class determination device D for both measured powder rates was approximately 33.3%. Therefore, with the third and fourth image sizes, particle size class determination device D determines each particle size class equally, and is unable to distinguish between small, medium, and large particles, making it unable to determine the particle size class. Therefore, it is considered that the third and fourth image sizes are too small for particle size class determination.

On the other hand, as can be seen from the results for the first image size indicated by ◆ in Figure 8, the average small particle size probability calculated by particle size class determination device D for all measured powder rates is approximately 100%. Therefore, at the first image size, particle size class determination device D determines all particles as being in the small particle size class, and is unable to distinguish between small, medium, and large particles, making it unable to determine the particle size class. Therefore, it is considered that the first image size is too large for particle size class determination.

For this reason, when determining particle size classes, the image size must be greater than the third image size and smaller than the first image size. In other words, if the image size input to the machine learning model is a size obtained by multiplying the target particle size value by a predetermined constant, then in the above example, the target particle size value is 2 mm, and the predetermined constant is greater than 7 and less than 12. Preferably, the predetermined constant is greater than 8 and less than 11. Preferably, the predetermined constant is 10.

To determine the particle size class, the user (operator) inputs the image size (or the predetermined constant) to the particle size class determination device D via the input unit 4. For example, a second image size (or 10 as the predetermined constant) is input to the particle size class determination device D. Note that it is assumed that a second machine learning model that has been trained and corresponds to each of multiple image sizes (or the predetermined constants) is pre-stored in the memory unit 7.

In Figure 9, first, the particle size class determination device D receives input of the image size (or the predetermined constant) at the input unit 4 via the control unit 31 of the control processing unit 3, and stores this input image size (or the predetermined constant) in the memory unit 7 (S31). The illumination unit 2 is also turned on.

Next, the particle size class determination device D uses the control unit 31 of the control processing unit 3 to select a machine-learned second machine learning model that corresponds to the input image size (or the predetermined constant) from among multiple machine-learned second machine learning models stored in the memory unit 7, and sets this selected machine-learned second machine learning model as the first machine learning model to be used in the particle size class processing unit 33 (S32).

Next, the particle size class determination device D controls the control unit 31 to acquire an image using the image acquisition unit 1 and store it in the memory unit 7 (S33).

Next, the particle size class determination device D uses the image size adjustment unit 34 of the control processing unit 3 to adjust the image acquired in process S33 to the image size accepted in process S31 (S34).

Next, the particle size class determination device D, using the particle size class processing unit 33 of the control processing unit 3, inputs the image whose size has been adjusted in step S34 into the machine learning model that has been set up in step S32 to determine each probability for the image, and determines the particle size class with the highest probability from among the probabilities as the particle size class for the image, and stores the result of this determination in the memory unit 7 (S35).

Next, the particle size class determination device D outputs the determination result determined in process S35 to the output unit 5 via the control unit 31 (S36).

More specifically, in steps S34 through S36, regions of the image size accepted in step S31 are extracted sequentially from the image acquired in step S33, starting from one corner (e.g., the upper left corner of the image) and moving diagonally to the opposite corner (in this example, the lower right corner of the image) in a manner similar to raster scanning. Each time a region is extracted, the image of the extracted region is input into the machine-learning model established in step S32, probabilities are calculated for the image of the extracted region, and a particle size class is determined for the image of the extracted region. The results of this determination are superimposed and displayed on the image acquired in step S33 at the position of the extracted region. Here, the calculated probabilities are used directly as the probabilities for each particle size class for the image of the extracted region. For example, as shown in Figures 10A to 10D, for an image of a region where the small particle size probability is 50% or more, a mark (□) indicating that the position of the region is determined to be in the small particle size class with a probability of 50% or more is superimposed and displayed on the image acquired in process S33. The image shown in Figure 10A is an image in which the particle size class determination result (□) for the image with a 20% powder rate shown in Figure 10E is superimposed on the image with a 20% powder rate shown in Figure 10E. The image shown in Figure 10B is an image in which the particle size class determination result (□) for the image with a 30% powder rate shown in Figure 10F is superimposed on the image with a 30% powder rate shown in Figure 10F. The image shown in Figure 10C is an image in which the particle size class determination result (□) for the image with a 40% powder rate shown in Figure 10G is superimposed on the image with a 40% powder rate shown in Figure 10G. The image shown in Figure 10D is an image in which the particle size class determination result (□) for the image with a 50% powder content shown in Figure 10H is superimposed on the image with a 50% powder content shown in Figure 10H. The image acquired in process S33 is divided into regions of the image size accepted in process S31 in a two-dimensional matrix, and the particle size class is determined for each of the images in these divided regions, with the determination result (□) being marked. As shown in Figure 10, the number of square marks increases as the powder content increases. In other words, as the powder content increases, the number of regions determined to be in the small particle size class with a probability of 50% or more increases.

Returning to FIG. 9 , the particle size class determination device D then determines, via the control unit 31, whether or not the process has ended. If the result of this determination is that the process has ended (Yes), the particle size class determination device D ends the process. If the result of this determination is that the process has not ended (No), the particle size class determination device D then executes process S38. For example, if, during the execution of each of the above-mentioned processes S33 to S36, the input of a predetermined command instructing the process to end is received via the input unit 4 (if a predetermined input switch, etc., is operated), it is determined that the process has ended; otherwise, it is determined that the process has not ended.

In process S38, the particle size class determination device D uses the control unit 31 to determine whether it is time for the next determination. If the result of this determination is that it is time for the next determination (Yes), the particle size class determination device D returns to process S33 to determine the particle size class of the next image. On the other hand, if the result of the determination is that it is not time for the next determination (No), the particle size class determination device D returns to process S38. Therefore, process S38 is repeated until it is time for the next determination. For example, when determining the particle size class of an image at a predetermined sampling interval, it is determined that it is time for the next determination when the predetermined sampling interval has elapsed since the previous determination.

In the above description, the image size is input in step S31, and a machine-learned second machine learning model corresponding to this input image size is set in the particle size class processing unit 33 in step S32. However, the image size may be specified in advance, and a machine-learned second machine learning model corresponding to this may be set in advance in the particle size class processing unit 33, and steps S31 and S32 may be omitted.

With regard to particle size class determination, the particle size class determination device D operates as follows:

As described above, the particle size class determination device D with image size determination assistance function in this embodiment, and the particle size class determination method and particle size class determination program implemented therein, use a first machine learning model to determine the particle size class of an image of multiple particles. Therefore, by performing machine learning on images of various particles, it is possible to determine the particle size class of a wider range of particles. During this determination, the size of the image input to the first machine learning model is a predetermined constant multiplied by the particle size value of the determination target. Therefore, the particle size class determination device D, particle size class determination method, and particle size class determination program can appropriately determine the particle size class. In particular, the particle size class determination device D, particle size class determination method, and particle size class determination program are suitable when the particle size to be determined is a pseudoparticle.

The particle size class determination device D, particle size class determination method, and particle size class determination program further include an illumination unit 2 (2-1 to 2-4) and use an imaging unit 1 for the image acquisition unit 1. Therefore, the particle size class determination device D, particle size class determination method, and particle size class determination program can be used, for example, on a production line within a factory, allowing particle size classes to be determined during production.

The above-described embodiment provides the particle size class determination device D, particle size class determination method, and particle size class determination program, in which the predetermined constant is set to a value greater than 7 and less than 12.

The above-described embodiment provides the particle size class determination device D, particle size class determination method, and particle size class determination program for determining pseudo particles. Because the particle size class determination device D, particle size class determination method, and particle size class determination program are used to determine the particle size class of pseudo particles, the particle size class of the pseudo particles can be appropriately determined.

In this embodiment, the particle size class determination device D with image size determination support function, and the image size determination support method and image size determination support program implemented therein, output the probability for each of the multiple particle size classes obtained by inputting a predetermined size determination dataset into a machine-learned second learning model. The user can then determine the optimal image size from among multiple image sizes by referencing the output probability for each of the multiple particle size classes and the content rate associated with the image (in this embodiment, the powder content is an example of this). Therefore, the particle size class determination device D with image size determination support function, image size determination support method, and image size determination support program can effectively assist the user in determining the image size. Therefore, this embodiment provides the particle size class determination device D with image size determination support function, image size determination support method, and image size determination support program that assist in determining the image size used for particle size class determination.

In order to express the present invention, the present invention has been appropriately and sufficiently described above through embodiments with reference to the drawings. However, it should be recognized that those skilled in the art could easily modify and/or improve the above-described embodiments. Therefore, unless modifications or improvements made by those skilled in the art deviate from the scope of the claims set forth in the claims, such modifications or improvements are deemed to be encompassed within the scope of the claims.

D. Particle size class determination device with image size determination support function 1 Image acquisition unit 2 (2-1 to 2-4) Illumination unit 3 Control processing unit 4 Input unit 5 Output unit 6 Interface unit (IF unit)
7 Memory unit 31 Control unit 32 Machine learning unit 33 Particle size class processing unit 34 Image size adjustment unit

Claims (6)

an image acquisition unit that acquires an image of a plurality of particles including particles of different particle sizes;
a particle size class processing unit that, when an image is input and particle sizes are divided into a plurality of particle size classes into a plurality of different ranges, calculates the probability of each of the plurality of particle size classes in the image acquired by the image acquisition unit using a machine learning model that has undergone machine learning and outputs, for each of the plurality of particle size classes, the probability that the input image will be classified into that particle size class;
Any of the boundary values in the plurality of ranges is a particle size value that is a predetermined target for judgment,
The size of the image input to the machine learning model is a size obtained by multiplying the particle size value of the judgment target by a predetermined constant,
the predetermined constant is greater than 7 and less than 12 ;
The particles are pseudo-particles each comprising a first particle serving as a core and a second particle attached to the surface of the first particle and smaller than the first particle .
Particle size class determination device. an illumination unit that illuminates the plurality of particles;
The image acquisition unit is an imaging unit that generates an image.
The particle size class determination device according to claim 1 . 1. An image size determination support device that supports the determination of an image size to be used for determining a particle size class when particle sizes are divided into a plurality of particle size classes each having a different range, the device comprising:
a first data acquisition unit that acquires a predetermined learning dataset;
a second data acquisition unit that acquires a predetermined size determination data set;
a machine learning unit that uses a predetermined learning dataset acquired by the first data acquisition unit to perform machine learning on a second machine learning model that receives an input image and outputs, for each of the plurality of particle size classes, a probability that the input image is classified into the particle size class;
an output unit that outputs the probabilities of each of the plurality of particle size classes output from the second machine learning model trained by the machine learning unit by inputting the predetermined size determination data set acquired by the second data acquisition unit into the second machine learning model trained by the machine learning unit,
the predetermined learning dataset and the predetermined size determination dataset each include the same number of data groups each including image data of a plurality of images;
The plurality of data groups have different image sizes,
In each of the plurality of data groups in the predetermined learning dataset, the plurality of images in the data group are a plurality of images including images of each of the plurality of particle size classes, and each of the plurality of images in the data group is linked to the particle size class of the image;
In each of the plurality of data groups in the predetermined size determination data set, the plurality of images in the data group are a plurality of images including images with different content rates of particles belonging to a particle size class with a smallest particle size, and the content rate of each of the plurality of images in the data group is linked,
the number of the unmachine-learned second machine learning models is the same as the number of the plurality of data groups, and the unmachine-learned second machine learning models are provided corresponding to each of the plurality of data groups;
the machine learning unit performs machine learning on a second machine learning model corresponding to each of the plurality of data groups in the predetermined training dataset, the second machine learning model being an untrained model corresponding to the data group, using a plurality of images in the data group;
The output unit outputs the probabilities of each of the plurality of particle size classes output from the second machine learning model trained by the machine learning unit by inputting, for each of the plurality of data groups in the predetermined size determination data set, a plurality of images in the data group into the second machine learning model trained by the machine learning unit using a data group in the predetermined training data set having the same image size as the image size in the data group.
Image size determination support device. The image size determination support device according to claim 3 is further provided,
As the machine-learned machine learning model, a second machine learning model trained by the machine learning unit is used.
The particle size class determination device according to claim 1 or 2 . 1. An image size determination support method for supporting the determination of an image size to be used for determining a particle size class when particle sizes are divided into a plurality of particle size classes each having a different range, the method comprising:
a first data acquisition step of acquiring a predetermined training data set;
a second data acquisition step of acquiring a predetermined sizing data set;
a machine learning process in which an image is input and a second machine learning model, which is not yet machine-learned, is trained by machine using the predetermined training data set acquired in the first data acquisition process, and which outputs, for each of the plurality of particle size classes, the probability that the input image is classified into that particle size class;
an output step of inputting the predetermined size determination data set acquired in the second data acquisition step into a second machine learning model trained by machine learning in the machine learning step, and outputting the probabilities of each of the plurality of particle size classes output from the second machine learning model trained by machine learning in the machine learning step;
each of the predetermined training dataset and the predetermined size determination dataset includes a plurality of data groups each including image data of a plurality of images;
The plurality of data groups have different image sizes,
In each of the plurality of data groups in the predetermined learning dataset, the plurality of images in the data group are a plurality of images including images of each of the plurality of particle size classes, and each of the plurality of images in the data group is linked to the particle size class of the image;
In each of the plurality of data groups in the predetermined size determination data set, the plurality of images in the data group are a plurality of images including images with different content rates of particles belonging to a particle size class with a smallest particle size, and the content rate of each of the plurality of images in the data group is linked,
the number of the unmachine-learned second machine learning models is the same as the number of the plurality of data groups, and the unmachine-learned second machine learning models are provided corresponding to each of the plurality of data groups;
the machine learning step performs machine learning of an untrained second machine learning model corresponding to each of the plurality of data groups in the predetermined training dataset using a plurality of images in the data group;
The output step outputs the probabilities of each of the plurality of particle size classes output from the second machine learning model trained in the machine learning step by inputting, for each of the plurality of data groups in the predetermined size determination data set, a plurality of images in the data group into the second machine learning model trained in the machine learning step using a data group in the predetermined training data set having the same image size as the image size in the data group.
A method to help determine image size. 1. An image size determination support program executed by a computer, which supports the determination of a size of an image to be used for determining a particle size class when particle sizes are divided into a plurality of particle size classes each having a different range, the program comprising:
a first data acquisition step of acquiring a predetermined training data set;
a second data acquisition step of acquiring a predetermined sizing data set;
a machine learning process in which an image is input and a second machine learning model, which is not yet machine-learned, is trained by machine using the predetermined training data set acquired in the first data acquisition process, and which outputs, for each of the plurality of particle size classes, the probability that the input image is classified into that particle size class;
an output step of inputting the predetermined size determination data set acquired in the second data acquisition step into a second machine learning model trained by machine learning in the machine learning step, and outputting the probabilities of each of the plurality of particle size classes output from the second machine learning model trained by machine learning in the machine learning step;
each of the predetermined training dataset and the predetermined size determination dataset includes a plurality of data groups each including image data of a plurality of images;
The plurality of data groups have different image sizes,
In each of the plurality of data groups in the predetermined learning dataset, the plurality of images in the data group are a plurality of images including images of each of the plurality of particle size classes, and each of the plurality of images in the data group is linked to the particle size class of the image;
In each of the plurality of data groups in the predetermined size determination data set, the plurality of images in the data group are a plurality of images including images with different content rates of particles belonging to a particle size class with a smallest particle size, and the content rate of each of the plurality of images in the data group is linked,
the number of the unmachine-learned second machine learning models is the same as the number of the plurality of data groups, and the unmachine-learned second machine learning models are provided corresponding to each of the plurality of data groups;
the machine learning step performs machine learning of an untrained second machine learning model corresponding to each of the plurality of data groups in the predetermined training dataset using a plurality of images in the data group;
The output step outputs the probabilities of each of the plurality of particle size classes output from the second machine learning model trained in the machine learning step by inputting, for each of the plurality of data groups in the predetermined size determination data set, a plurality of images in the data group into the second machine learning model trained in the machine learning step using a data group in the predetermined training data set having the same image size as the image size in the data group.
A program to help determine image size.