JP2024004369A - Analysis system for pyroclastic material component particles and analysis program for pyroclastic material component particles - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an analysis system and an analysis program for pyroclastic material component particles, that have excellent flexibility in optical designing and allow pyroclastic material component particles to be analyzed in a robust and efficient manner even for such crowded or contacting particles or small particles.

SOLUTION: A system comprises: a digital camera 2 or the like which images pyroclastic material component particles for forming a particle image; a detection model 10 for outputting a rectangle surrounding the pyroclastic material component particles by inputting the particle image; first analysis means 11 which acquires the rectangle by inputting the particle image into the detection model 10, and cuts out individual pieces of the pyroclastic material component particles surrounded by the rectangle from the particle image for forming an individual particle image; a classification model 20 for outputting a class of the pyroclastic material component particles by inputting the individual particle image; and second analysis means 21 which acquires the class by inputting the individual particle image into the classification model 20, for classifying the pyroclastic material component particles.

SELECTED DRAWING: Figure 2

COPYRIGHT: (C)2024,JPO&INPIT

Description

Application for application of Article 30, Paragraph 2 of the Patent Act filed August 27, 2021 https://confit. atlas. jp/guide/event/geosocjp128/proceedings/list https://confit. atlas. jp/guide/event-img/geosocjp128/1poster18-20-03/public/pdf? type=in [Publications, etc.] 128th Academic Conference of the Geological Society of Japan, September 4, 2021

The present invention relates to a pyroclastic material particle analysis system and a pyroclastic material particle analysis program for classifying pyroclastic material particles.

Particles that make up pyroclastic materials (hereinafter referred to as pyroclastic material constituent particles) are thought to hold information such as eruption style and magma properties, and analysis of their constituent fractions and shapes are fundamental and important in geological surveys. It's work. However, the task of visually identifying the type of particles during analysis is difficult to standardize and is labor-intensive, so there is a need to improve efficiency.

Therefore, research is being conducted aiming at automating analysis based on image processing using machine learning (for example, see Non-Patent Document 1). Many of these studies use particle analysis equipment to efficiently collect and utilize individual particle images from images containing a large amount of pyroclastic particles. However, in previous research, color information in particle images is lost when particles are detected, making it difficult to link judgments similar to those made with naked eye identification during post-process image processing. Furthermore, due to the characteristics of the equipment, there are restrictions on the upper limit size of particles, and analysis of particles larger than the size of volcanic lapilli remains difficult to replace manually.

Shoji, D., Noguchi, R., Otsuki, S. and Hino, H. (2018) Classification of volcanic ash particles using aconvolutional neural network and probability. Scientific Reports.

In view of the above circumstances, the present invention has a high degree of freedom in optical design, and is capable of analyzing pyroclastic material constituent particles robustly and efficiently even in the case of particle crowding, contact, and small particles. The purpose is to provide a system and an analysis program for particles constituting pyroclastic materials.

One aspect of the present invention for achieving the above object includes an imaging means for forming a particle image by imaging particles constituting the pyroclastic material, and outputting a rectangle surrounding the particles constituting the pyroclastic material when the particle image is input. The rectangle is obtained by inputting the particle image to a first trained model and the first trained model, and individual particles of pyroclastic material surrounded by the rectangle are cut out from the particle image. a first analysis means that forms an image; a second trained model that outputs a classification of pyroclastic material constituent particles captured in the individual particle image when the individual particle image is input; and the second learning model. and a second analysis means for classifying the pyroclastic material particles by inputting the individual particle images into a finished model and obtaining the classification.

Another aspect of the present invention for achieving the above object is a first analysis means for forming an individual particle image in which each pyroclastic material constituent particle is cut out from a particle image in which the pyroclastic material constituent particle is imaged; A program for analyzing pyroclastic material particles that causes a computer to function as a second analysis means for classifying individual particle images, wherein the computer outputs a rectangle surrounding the pyroclastic material particles when the particle image is input. and a second trained model that outputs a class of particles constituting pyroclastic material captured in the individual particle image when the individual particle image is input, and the first analysis means inputs the particle image into the first trained model to obtain the rectangle, cuts out individual pyroclastic material constituent particles surrounded by the rectangle from the particle image to form individual particle images, A program for analyzing particles constituting pyroclastic materials, characterized in that the second analysis means classifies the particles constituting pyroclastic materials by inputting the individual particle images into the second trained model and acquiring the classes. It is in.

According to the present invention, there is a system for analyzing particles constituting pyroclastic materials, which has a high degree of freedom in optical design, and is capable of robustly and efficiently analyzing particles constituting pyroclastic materials even in the case of particle crowding, contact, and small particles. An analysis program for particles constituting crushed materials is provided.

1 is a diagram showing a schematic configuration of an analysis system according to Embodiment 1. FIG. 1 is a diagram showing functions of an analysis system and an analysis program according to a first embodiment; FIG. 3 is a diagram showing an example of a particle image according to Embodiment 1. FIG. FIG. 3 is a diagram showing an example of a rectangle according to the first embodiment. 3 is a diagram showing individual particle images and classification according to Embodiment 1. FIG. FIG. 3 is a diagram showing functions related to learning of the analysis program according to the first embodiment. FIG. 3 is a diagram showing a rectangle attached to a particle image for creating training data. FIG. 3 is a diagram showing functions of an analysis system and an analysis program according to a second embodiment.

<Embodiment 1>
Pyroclastic material is a general term for ejecta ejected from a crater by an eruption, excluding lava flows. Depending on the particle size, pyroclastic materials include volcanic rock masses, volcanic lapilli, and volcanic ash, and these are collectively referred to as pyroclastic material particles. Pyroclastic particles generally have the following particle sizes. Volcanic ash has a particle size of less than 2 mm, and volcanic ash, also called fine volcanic ash, has a particle size of about 1/16 mm or less. Volcanic lapilli has a particle size of 2 mm to 64 mm. A volcanic rock mass is a particle with a particle size larger than 64 mm. Here, a case will be exemplified in which particles constituting pyroclastic material collected around volcano V1 or volcano V2 are analyzed.

FIG. 1 is a diagram showing a schematic configuration of an analysis system, and FIG. 2 is a diagram showing functions of an analysis system and an analysis program. As shown in FIG. 1, the analysis system 1 includes a digital camera 2, a scanner 3, a microscope camera 4, and a computer 5 as hardware. An analysis program 6 is executed on the computer 5. The analysis system 1 and analysis program 6 of the present invention detect individual pyroclastic material particles from an image in which a considerable amount of pyroclastic material particles are captured, regardless of the particle size of the pyroclastic material particles, Furthermore, analysis processing is performed to classify the constituent particles of pyroclastic materials.

The digital camera 2, the scanner 3, and the microscope camera 4 are all examples of imaging means, and are devices that take images of particles constituting the pyroclastic material to form an image. Pyroclastic particles collected at the site have a variety of particle sizes, and the imaging means is selected according to the particle size that accounts for the majority. For example, if the particles are mainly volcanic lapilli or volcanic rock blocks with a particle size of 20 mm or more, the digital camera 2 is used to capture the image. If the particles are mainly volcanic ash or lapilli with a particle size of 0.5 mm to 20 mm, images are taken using the scanner 3. If the volcanic ash is mainly composed of particles with a particle size of less than 0.5 mm, images are taken with the microscope camera 4.

The image obtained by the imaging means is called a particle image. The particle images are taken for each volcano V1 or volcano V2. FIG. 3 shows an example of a particle image. The particle image is a color image, and various formats such as JPEG and PNG can be adopted. The particle image includes a considerable amount of particles constituting the pyroclastic material to be analyzed. Furthermore, it is desirable that the background of the particle image be blue. This is because pyroclastic material constituent particles exhibiting a blue color are rare, and therefore individual pyroclastic material constituent particles can be easily detected from the image.

Regardless of whether a digital camera 2, scanner 3, or microscope camera 4 is used for the particle image, the size of the long axis of the pyroclastic material particles should be approximately 2.5-8% of the side of the particle image. This image was taken in . In other words, images may be taken using the optical zoom function, digital zoom function, or resolution setting function of a digital camera 2, etc., or digital The distance between the camera 2 and the particles constituting the pyroclastic material is adjusted to take an image. In addition, particle images are taken so that the particles constituting the pyroclastic material do not come into contact with each other as much as possible and do not overlap.

The size of the long axis of the pyroclastic material particles may be determined by sizing the pyroclastic material particles using a sieve or the like. Further, although there is no particular limitation on the size of the particle image used for inference, it is preferable that the particle image used for learning be approximately 1024×1024 pixels. Further, the particle image does not need to be square, but may be rectangular. Furthermore, the size of the long axis of the pyroclastic material constituent particles is approximately 2.5-8% of the short side of the particle image. The density (number of particles/image) of particles constituting the pyroclastic material per particle image is preferably about 1 to 700 particles.

The means for storing particle images in the computer 5 from an imaging means such as the digital camera 2 is not particularly limited. For example, the particle image may be stored in the storage device of the computer 5 from the imaging means via a USB interface, removable media, wireless or wired LAN, or the like.

The computer 5 is a general computer equipped with a processing device such as a CPU or GPU, a storage device (USB flash drive, solid state drive, hard disk drive, etc.), an input device (keyboard, mouse, etc.), an output device (display, etc.), a communication means, etc. It is a typical information processing device. The computer 5 uses the particle image captured by the imaging means as input data, and detects and classifies each of the particles constituting the pyroclastic material captured in the particle image.

The storage device of the computer 5 stores an analysis program 6, a detection model 10, and a classification model 20. The analysis program 6 causes the computer 5 to function as a first analysis means 11 and a second analysis means 21.

The detection model 10 is a model that has been trained to output a rectangle surrounding the particles constituting the pyroclastic material when a particle image is input. Specifically, it is an implementation of an algorithm for detecting objects in images based on a convolutional neural network (referred to as CNN), and is trained using training data. For example, the detection model 10 may be "You look only once" (referred to as YOLO), "Single Shot Multibox Detector" (referred to as SSD), "Region CNN" (referred to as RCNN), etc. Can be used.

The detection model 10 may be created and learned for each volcano that is the source of the pyroclastic particles. Alternatively, the detection model 10 may be created and learned in common for a plurality of volcanoes. Here, a case will be described in which the detection model 10 created and learned in common for the volcano V1 and the volcano V2 is used.

The rectangle surrounding the pyroclastic material particles is information that surrounds the individual pyroclastic material particles captured in the particle image and specifies the position and size of the pyroclastic material particles in the particle image. The detection model 10 learns rectangles based on annotation rules described later. FIG. 4 shows a rectangle obtained by giving a particle image to the detection model 10 trained in this way and superimposed on the particle image. In the figure, a rectangle output by the detection model 10 is displayed in pink. It can be seen that each rectangle surrounds individual pyroclastic constituent particles. Pyroclastic material particles within 0.2% of the edge of the particle image are excluded from the output even if they are detected by the detection model 10. Therefore, rectangles are not displayed for particles constituting pyroclastic materials near the edges of the particle image.

The first analysis means 11 inputs a particle image as test data to which a rectangle is not attached to the detection model 10 and performs inference to obtain a rectangle. Then, the first analysis means 11 forms an image in which pyroclastic material constituent particles surrounded by a rectangle are cut out from the particle image. This image is called an individual particle image. The individual particle image is a color image like the original particle image, and is in a format such as JPEG or PNG. Further, the size of the individual particle image is made the same as the size of the individual particle image (hereinafter referred to as a predetermined size) used when the classification model 20 is trained.

If the cut out individual particle image seems to protrude from the predetermined size, it is reduced, and if the cut out individual particle image seems to be smaller than the predetermined size, it is enlarged. If a margin occurs in an individual particle image due to an aspect ratio different from the predetermined size, the first analysis means 11 can change the aspect ratio or pad the margin with pixels of the background color, so that the first analysis means 11 can change the size to pink as shown in FIG. Each of the pyroclastic material constituent particles surrounded by a rectangle with a solid color line is formed as an individual particle image of a predetermined size.

The classification model 20 is a model that is trained to output the classification of the pyroclastic material constituent particles captured in the individual particle image when the individual particle image is input. Specifically, it is an implementation of an algorithm for classifying images based on CNN, and is trained using training data.

A classification model 20 for each volcano may be created from individual particle images for each volcano, or a single classification model 20 may be created from individual particle images for a plurality of volcanoes. Here, a case will be described in which the former classification model 20, that is, the classification model 20 for each volcano (classification model 20 for volcano V1 and volcano V2) learned from individual particle images for each volcano is used.

Classification refers to classifying discrete input values into a plurality of predefined classes, and in the present invention, classes that would be assigned to individual particle images extracted from imaged pyroclastic materials It refers to estimating the A specific example of the class is shown in FIG.

FIG. 5(a) is an individual particle image of pyroclastic material constituent particles whose class is scoria. FIG. 5(b) is an individual particle image of pyroclastic constituent particles whose class is pumice. FIG. 5C is an individual particle image of pyroclastic material constituent particles whose class is rock fragments. Note that the three classes are just examples, and are not limited to scoria etc., and there is no particular limitation on the number of classes.

The classification model 20 is trained to predict and output each class of the individual particle images when the individual particle images as shown in FIG. 5 are input. Since the individual particle images are color images, the classification of particles constituting pyroclastic materials is realized using color information as well.

The second analysis means 21 selects one from a plurality of classification models 20 learned for each volcano. Specifically, if input data is an individual particle image of particles constituting pyroclastic material obtained from the volcano V1, the classification model 20 for the volcano V1 is selected. Then, the second analysis means 21 inputs the individual particle images to the classification model 20 for the volcano V1, thereby performing inference calculations to obtain the pyroclastic material constituent particles and their classification. The second analysis means 21 outputs the inference result to the storage device of the computer 5 or displays it on the display device.

As described above, according to the analysis system 1 and the analysis program 6 of the present invention, as shown in FIG. It is possible to analyze whether the volcano V1 or the volcano V2 belongs to any of the scoria, pumice, and rock fragment classes.

Learning of the detection model 10 and classification model 20 will be explained using FIG. 6. As shown in FIG. 6, the analysis program 6 includes a first learning means 12 for learning the detection model 10 and a second learning means 22 for learning the classification model 20. Here, a case will be described in which one detection model 10 is created and trained for the volcanoes V1 and V2, and a classification model 20 is created and trained for each of the volcanoes V1 and V2.

The first learning means 12 causes the detection model 10 to learn using the particle image and the rectangle attached to the particle image as training data.

Table 1 shows an example of the teacher data.

Particle image A in Table 1 represents a particle image (file) of volcano V1, and particle image B represents a particle image (file) of volcano V2. Xleft is the X coordinate of the vertex located at the upper left of the rectangle in the particle image, and Ytop is the Y coordinate of the vertex located at the upper left of the rectangle in the particle image. Xright is the X coordinate of the vertex located at the lower right of the rectangle in the particle image, and Ybottom is the Y coordinate of the vertex located at the lower right of the rectangle in the particle image. A plurality of rectangles are attached to each of particle image A and particle image B, and particle image A and particle image B and their rectangles are used in the detection model 10 as training data. Of course, the training data in Table 1 is just an example, and in reality, it is preferable to prepare an appropriate number of particle images in order to improve learning accuracy. Note that the amount of training data may be increased by data augmentation. Further, although the coordinate notation is Xleft, Ytop, Xright, and Ybottom, it is not limited thereto. It may be selected as appropriate depending on the algorithm for detecting and classifying objects. For example, a rectangle may be represented by coordinate notation consisting of Xleft, Ytop, width in the X direction, and height in the Y direction.

Normally, algorithms such as YOLO that detect and classify objects from images use the classification as training data along with a rectangle (region) surrounding the object included in the image. In the present invention, a class is also assigned to the rectangle surrounding the pyroclastic material particles included in the particle image, but the same class is assigned regardless of the actual class of the pyroclastic material particles. In the example of Table 1, the class of all rectangles is "grain".

The rectangle (teacher data) attached to the particle image described above is created, for example, as follows. The first learning means 12 displays on the screen of the computer 5 the particle image to which a rectangle has not yet been attached. Next, the first learning means 12 performs a process of drawing a rectangle superimposed on the particle image in response to operations such as mouse movement and clicking. Then, the first learning means 12 calculates a rectangle consisting of the coordinates in the particle image as exemplified in Table 1 based on the information of the drawn rectangle (coordinates on the screen of the computer 5), and stores it as training data. Store it in the device. Generally, adding information tags (metadata) to data is called annotation, but here, the operation of adding a rectangle to a particle image is called annotation.

By using the function realized by the first learning means 12, the user can visually check the particle image and the particles constituting the pyroclastic material captured therein, and use an input device such as a mouse to create a rectangle surrounding the particle constituting the pyroclastic material. You can create training data by inputting When creating a rectangle of training data, that is, annotating it, it is preferable to consider the rules described in FIG. 7 and below.

FIG. 7 shows a particle image in which a rectangle made up of green lines connecting four corner points is added as an annotation. The rectangle formed by the red line displayed in FIG. 7(a) is shown to be enlarged in each of FIG. 7(b) to FIG. 7(f).
・Figure 7(a): Each of the many pyroclastic material constituent particles is surrounded by a rectangle.
・Figure 7(b) - Figure 7(f): Create a rectangle with an interval of about 0-5 pixels from the outer shape of the pyroclastic material constituent particles.
・Figure 7(c) White arrow: If the long side of the rectangle surrounding the pyroclastic material particles is less than 1%, no rectangle is created.
・Figure 7(d): If the particles constituting the pyroclastic material are close to each other, in contact with each other, or overlap each other, separate them one by one and create a rectangle.
・Figure 7(e): A rectangle is also created for the pyroclastic material constituent particles cut by the edges of the particle image.
・Figure 7(f): Even if the long side of the rectangle surrounding the pyroclastic material particles is 1% or more of the entire image, a rectangle is not created if the short side is 0.5% or less of the entire image.

As for how to train the detection model 10, a known algorithm for detecting and classifying objects such as YOLO or SSD can be used, so a detailed explanation will be omitted. No matter which algorithm is used, the detection model 10 trained using training data consisting of particle images, rectangles, and the same classification as shown in Table 1 will output a rectangle as illustrated in FIG. 4 from the particle image. be able to.

The second learning means 22 causes the classification model 20 to learn using the individual particle images and the classes assigned to the individual particle images as training data. The classification model 20 is created for each volcano V1 and volcano V2.

Table 2 shows an example of the teacher data. Individual particle images XZ in Table 2 represent individual particle images (files) of volcano V1. The class represents the class of one pyroclastic material constituent particle imaged in the individual particle image. Although not particularly illustrated, the same applies to the volcano V2.

The training data used for learning the classification model 20, that is, the classes assigned to individual particle images, are created, for example, as follows. The second learning means 22 displays individual particle images to which no classes have been assigned yet on the screen of the computer 5. Next, the second learning means 22 performs a process of having the user select one of the multiple classes displayed on the screen using the mouse, or inputting a class using the keyboard into the input field displayed on the screen. Then, the second learning means 22 stores the individual particle images and their classes in the storage device as teacher data. By performing such processing for each of the volcanoes V1 and V2, teacher data for each of the volcanoes V1 and V2 is created. Note that the amount of training data may be increased by data augmentation.

By using the function realized by the second learning means 22, the user can classify the pyroclastic particles using an input device such as a mouse or a keyboard while visually recognizing the pyroclastic particles captured in the individual particle image. You can create training data by inputting

A known algorithm based on CNN can be used to train the classification model 20, so a detailed explanation will be omitted. No matter which algorithm is used, the classification model 20 trained using training data consisting of individual particle images and classes as shown in Table 2 is capable of outputting the classes of particles constituting pyroclastic materials captured in individual particle images. can.

As explained above, the analysis system 1 and the analysis program 6 detect individual pyroclastic material particles from the captured particle images of the pyroclastic material particles, and create individual particle images in which the pyroclastic material particles are imaged. The individual particle images can then be classified. Since the particles constituting the pyroclastic material are classified individually in this way, it is possible to streamline the analysis of the constituent fraction and shape of the particles constituting the pyroclastic material in geological surveys.

Classification of particles constituting pyroclastic materials was conventionally achieved by visual inspection, but standardization of judgment was difficult and labor-intensive. However, according to the analysis system 1 and the analysis program 6 of the present invention, classification of particles constituting pyroclastic materials is achieved by the detection model 10 and classification model 20 that have learned the appraisal work, so it depends on the experience and skill of the operator. There is no need to do this, and the labor required for classification work can be significantly reduced.

Furthermore, the analysis system 1 can employ any device such as a digital camera 2 as an imaging means. In other words, the digital camera 2 or the like can be employed in accordance with the size of the particles constituting the pyroclastic material, providing a high degree of freedom in optical design. From the above, it is robust against particle crowding, contact, and small particles, and can efficiently analyze volcanic rock masses, lapilli, and volcanic ash of different sizes. Furthermore, since there is a high degree of freedom in optical design, it is possible to construct an analysis system without using expensive and large analysis equipment.

In general, algorithms such as YOLO can simultaneously detect and classify particles constituting pyroclastic materials from particle images. However, when such a model is used, the cost of relearning and creating training data increases when trying to classify particles constituting pyroclastic material from a new source (volcano). Specifically, it is necessary not only to assign new classifications to particles constituting pyroclastic materials from volcanoes, but also to assign rectangles to particle images and retrain the model.

However, the present invention has the advantage that it is only necessary to create and learn a new classification model 20 corresponding to a new source (volcano), and there is no need to change or re-learn the detection model 10.

In the analysis system 1 and analysis program of the present invention, particle images are formed for each source of pyroclastic material particles (volcano V1, volcano V2), and the detection model 10 is trained in common even if the sources are different. It is what was done. Such a detection model 10 can obtain rectangles with sufficient precision even for particle images of particles constituting volcanic pyroclastic materials that are not used in the learning stage. Furthermore, in the learning stage, there is no need to create and train the detection model 10 for each volcano, and the learning cost can be reduced.

In the analysis system 1 and analysis program 6 of the present invention, the classification model 20 is learned for each source. As a result, individual particle images can be classified with higher accuracy than when a commonly learned classification model is used even if the sources are different.

Pyroclastic particles have various sizes ranging from μm to cm order. If particles constituting pyroclastic material were to be imaged at the same scale to form a particle image, it would be particularly difficult to detect volcanic ash on the order of μm. However, in the analysis system 1 and analysis program 6 of the present invention, the particle images are captured such that the long sides of the particles constituting the pyroclastic material are 2.5-8% of the long sides of the particle images. Even if the pyroclastic material constituent particles are of various sizes, the detection model 10 can detect rectangular shapes by using a sieve or the like to divide the particle size distribution into approximately 2φ increments, and separate the individual pyroclastic material constituent particles. Individual particle images of particles can be obtained.

In the analysis system 1 and analysis program 6 of the present invention, particle images are formed for each source of pyroclastic particles (volcano V1, volcano V2), and a common detection model 10 is used even if the sources are different. Let them learn. It is possible to create a detection model 10 that outputs rectangles with sufficient precision even for particle images of volcanic pyroclastic particles that are not used in the learning stage.

The analysis system 1 and analysis program 6 of the present invention train the classification model 20 for each source. Thereby, it is possible to create a classification model 20 that classifies individual particle images with high accuracy compared to a commonly learned classification model even if the generation sources are different.

<Embodiment 2>
The analysis system 1 and analysis program 6 of the second embodiment will be explained using FIG. 8. Components that are the same as those in Embodiment 1 are given the same reference numerals, and overlapping explanations will be omitted. The analysis system 1 and analysis program 6 of the second embodiment include a pass/fail determination model 30.

The quality determination model 30 is a model that has been trained to output a quality determination of whether the pyroclastic material constituent particles are good or bad when an individual particle image is input. Specifically, it is an implementation of an algorithm for classifying images based on CNN, and is trained using training data. As training data, a set of individual particle images and pass/fail judgments is used. The pass/fail judgment model 30 is trained by a known method using such training data.

A quality determination model 30 for each volcano may be created from individual particle images of each volcano, or a single quality determination model 30 may be created from individual particle images of a plurality of volcanoes. Here, a case will be described in which the latter quality judgment model 30, that is, the quality judgment model 30 created and learned in common for the volcano V1 and the volcano V2 is used.

A defective determination is given to an individual particle image in which one pyroclastic material particle does not fit, such as when a plurality of pyroclastic material particles are imaged in contact or overlapping, or most of the area is not imaged. A good judgment is given to an individual particle image in which one pyroclastic material constituent particle is contained.

The second analysis means 21A inputs the individual particle images into the quality determination model 30 and obtains the quality determination. Then, the pyroclastic material constituent particles are classified by inputting the individual particle images judged to be good to the classification model 20 and acquiring the classes. On the other hand, the second analysis means 21A does not classify the pyroclastic material constituent particles for individual particle images determined to be defective.

As described above, the analysis system 1 and the analysis program 6 of the second embodiment use the quality determination model to determine the quality of individual particle images before classifying the individual particle images. This makes it possible to avoid classification of defective individual particle images in which a single pyroclastic material constituent particle is not contained, and a decrease in classification accuracy. The first analysis means 11 uses the detection model 10 to create individual particle images, but not all individual particle images contain one pyroclastic material particle. However, since defective individual particle images are excluded by the quality determination model 30, it is possible to improve the accuracy of classification of finally obtained pyroclastic material constituent particles.

In addition, as one of the classes output by the classification model 20, in addition to scoria, pumice, etc., for example, "defect shape" may be added and the classification model 20 may be trained. However, since the number of individual particle images that are judged to be defective is extremely small, there is a risk that the accuracy will not improve even if the learning is performed. Therefore, it is preferable to have the classification model 20 perform only the classification of particles constituting the pyroclastic material without making a quality determination, and to have the quality determination model 30 perform the quality determination separately from the classification model 20.

Although the embodiments of the present invention have been described, the present invention is not limited to the above-described embodiments. For example, although the detection model 10 is common to a plurality of volcanoes (volcano V1, volcano V2), a different detection model 10 may be created and trained for each of a plurality of volcanoes. Although the classification model 20 was created and trained for each of a plurality of volcanoes, a classification model 20 common to a plurality of volcanoes may be created and trained.

Although the volcanoes V1 and V2 have been cited as examples of sources of pyroclastic particles, the present invention is not limited to these examples. The source may be one, or two or more.

The number of computers is not limited to one, but may be multiple. The learning of each model and the inference calculations for providing input data to each model and obtaining output data may be distributed and executed by multiple computers. . Furthermore, model learning and inference calculations do not need to be executed on the same computer, but may be executed on different computers.

Furthermore, although the digital camera 2 and the like which are the imaging means in Embodiments 1 and 2 are devices that directly form particle images, the imaging means is not limited thereto. For example, the image capturing means may be a camera that captures images on film, a scanner that converts photographs taken by the camera into digital data, or the like.

Note that the detection model 10, classification model 20, and pass/fail determination model 30 described in Embodiment 1 and Embodiment 2 are the first trained model, second trained model, and third trained model in the claims, respectively. This is an example.

DESCRIPTION OF SYMBOLS 1...Analysis system, 2...Digital camera, 3...Scanner, 4...Microscope camera, 5...Computer, 6...Analysis program, 10...Detection model (first learned model), 11...First analysis means, 12 ...first learning means, 20...classification model (second trained model), 21, 21A...second analysis means, 22...second learning means, 30...pass/fail determination model (third trained model) )

Claims (8)

an imaging means for imaging pyroclastic material constituent particles to form a particle image;
a first trained model that outputs a rectangle surrounding the pyroclastic material constituent particles when the particle image is input;
A first analysis in which the particle image is input to the first trained model to obtain the rectangle, and individual pyroclastic material constituent particles surrounded by the rectangle are cut out from the particle image to form individual particle images. means and
a second trained model that outputs a class of pyroclastic material constituent particles captured in the individual particle image when the individual particle image is input;
a second analysis means for classifying pyroclastic material constituent particles by inputting the individual particle images into the second trained model and acquiring the classes;
An analysis system for particles constituting pyroclastic materials, characterized by comprising: The analysis system for particles constituting pyroclastic material according to claim 1,
The particle image is formed for each source of the pyroclastic material constituent particles,
An analysis system for particles constituting pyroclastic materials, characterized in that the first trained model is one that has been learned in common even if the generation sources are different. The analysis system for particles constituting pyroclastic material according to claim 1,
The particle image is formed for each source of the pyroclastic material constituent particles,
The second trained model is trained for each source. A system for analyzing particles constituting pyroclastic materials. The analysis system for particles constituting pyroclastic material according to claim 1,
The particle image is captured such that the long side of the pyroclastic material particle is 2.5% to 8% of the long side of the particle image. The analysis system for particles constituting pyroclastic material according to claim 1,
comprising a third trained model that outputs a determination of whether the pyroclastic material constituent particles are good or bad when the individual particle image is input;
The second analysis means inputs the individual particle images into the third trained model to obtain the pass/fail judgment, obtains the class for the individual particle images judged to be good, and analyzes the particles constituting the pyroclastic material. An analysis system for particles constituting pyroclastic materials, characterized in that the individual particle images that are classified and determined to be defective are not classified as particles constituting pyroclastic materials. The analysis system for particles constituting pyroclastic material according to claim 1,
The particle image is formed for each source of the pyroclastic material constituent particles,
comprising a first learning means for learning the first trained model using the rectangle attached to the particle image,
A system for analyzing particles constituting pyroclastic materials, wherein the first learning means uses the particle images and the rectangles of different sources to learn the common first trained model. The analysis system for particles constituting pyroclastic material according to claim 1,
The particle image is formed for each source of the pyroclastic material constituent particles,
comprising a second learning means for learning the second trained model using the class assigned to the individual particle image,
A system for analyzing particles constituting pyroclastic materials, wherein the second learning means uses the individual particle images of different sources and the classes to learn the second trained model for each different source. . The computer functions as a first analysis means for forming an individual particle image in which each pyroclastic material constituent particle is extracted from a captured particle image of the pyroclastic material constituent particle, and a second analysis means for classifying the individual particle image. An analysis program for particles constituting pyroclastic materials,
The computer is
a first trained model that outputs a rectangle surrounding the pyroclastic material constituent particles when the particle image is input;
a second trained model that outputs a class of pyroclastic material constituent particles captured in the individual particle image when the individual particle image is input;
Equipped with
The first analysis means inputs the particle image into the first trained model to obtain the rectangle, and cuts out individual pyroclastic material constituent particles surrounded by the rectangle from the particle image to separate individual particles. form an image,
The second analysis means classifies the pyroclastic material particles by inputting the individual particle images into the second trained model and acquiring the class. program.

Images