JP7724688B2 - Particle size distribution measuring system and particle size distribution measuring method - Google Patents

Description

The present invention relates to a particle size distribution measurement system and a particle size distribution measurement method.

Batcher plants and CSG mixing plants are used in the process of producing concrete and CSG (Cemented Sand and Gravel) in the construction of dams and other structures. These plants require quality control to ensure that the particle size of locally generated materials (aggregates, soil, sand, etc.) used falls within a specified range (see Figure 12). For CSG quality control, it is necessary to understand the particle size distribution of locally generated materials. The particle size distribution is controlled, for example, into five classifications (0-5 mm, 5-10 mm, 10-20 mm, 20-40 mm, and 40-80 mm).
Traditionally, quality control tests have been conducted manually, with samples being sampled intermittently at a set frequency (once per hour in the early stages of construction) and particle size measurements being carried out (see, for example, JIS A 1102 "Sieving Test Method for Aggregate"). Manual quality control tests involve, for example, sieving samples into particle sizes, drying them, and then measuring the weight of the samples to determine the particle size distribution. Manual quality control tests require a lot of testing time, making it difficult to respond in a timely manner to fluctuations in particle size distribution.
Additionally, a technology has been proposed that captures images of materials (e.g., locally generated materials) moving along a conveying line and measures the particle size distribution of the material in real time based on the captured images. The general process for creating particle size distributions using images is as follows: image capture → image binarization → contour extraction → particle size determination → mass conversion → particle size distribution calculation. The most important step in this process is image binarization, which requires dispersing the material particles before capturing the image to prevent them from appearing as a single particle when binarized. Therefore, an image capture device has been proposed that includes a flow straightening plate through which the material flows from upstream to downstream, a screen located downstream of the flow straightening plate, and a capture unit located on the front side of the screen (see, for example, Patent Document 1). In this image capture device, multiple protrusions are formed on the surface of the flow straightening plate. This allows the material flowing down the surface of the flow straightening plate to come into contact with the protrusions and be dispersed. This eliminates overlapping and contact between the particles, allowing for accurate determination of the particle size of the material particles and improving the accuracy of particle size distribution measurement.

JP 2015-010952 A

The technology described in Patent Document 1 had the problem of increased costs due to the use of a flow rectifier plate to rectify the flow of material and a screen for photography. Furthermore, manual labor and conventional analysis methods required the installation of a dispersion mechanism such as a flow rectifier plate and the need to separate the material to limit the material flow, resulting in a problem of a small ratio of test samples to the material used in the actual construction.
From this perspective, the present invention provides a particle size distribution measurement system and a particle size distribution measurement method that can measure particle size distribution in a timely manner while keeping costs down. Furthermore, the present invention also makes it possible to easily increase the proportion of test samples.

A particle size distribution measurement system according to the present invention is a particle size distribution measurement system for measuring the particle size distribution of an object, and includes an imaging unit that images the object, and a processing unit that calculates the particle size distribution of the imaged object.
The processing unit includes a component detection unit that detects components of the object in the captured image, an area calculation unit that calculates the area of the components in the captured image, and a particle size distribution estimation unit that estimates the particle size distribution of the object from the detected components.
The component detection unit individually detects the components appearing in the captured image using a trained model that has learned about the components using a learning model that supports instance segmentation.
The particle size distribution estimation unit estimates the particle size distribution based on the total area for each particle size range in the captured image and a mass conversion coefficient that is the mass per predetermined area for each particle size range.
The learning model is, for example, MASK R-CNN, and the trained model is preferably one that has been machine-trained to detect the components by inputting the captured image.
The particle size distribution measurement system according to the present invention uses a learning model that supports instance segmentation, making it possible to separately detect overlapping or contacting components on a pixel-by-pixel basis. Here, segmentation outputs the class to which each pixel in an image belongs, while instance segmentation further divides the region into individual units (extracts the object region of each individual). As a result, for example, there is no need to rectify the object before photographing, which reduces equipment costs. Furthermore, since the particle size distribution is measured using a mass conversion coefficient obtained from a photographed image of the object, the particle size distribution can be measured with high accuracy.

The target object may be, for example, a locally generated material for CSG (Cemented Sand and Gravel), and the apparatus may include a transport unit for transporting the locally generated material. The photographing unit may continuously photograph the locally generated material being transported, and the particle size distribution estimating unit may calculate a particle size distribution for each photographed image from each photographed image and statistically process the results to estimate the particle size distribution of the locally generated material.
This makes it possible to detect all locally generated materials transported by the transport section, thereby increasing the number of samples used to estimate particle size distribution, and thus enabling more accurate particle size distribution measurements.

A particle size distribution measuring method according to the present invention is a particle size distribution measuring method for measuring the particle size distribution of an object, and includes an imaging step of imaging the object, and a processing step of calculating the particle size distribution of the imaged object.
The processing step includes a component detection step of detecting components of the object in the captured image, an area calculation step of calculating the area of the components in the captured image, and a particle size distribution estimation step of estimating the particle size distribution of the object from the detected components.
In the component detection process, the components appearing in the captured image are individually detected using a trained model that has learned about the components using a learning model that supports instance segmentation.
In the particle size distribution estimation step, the particle size distribution is estimated based on the total area for each particle size range in the captured image and a mass conversion coefficient that is the mass per predetermined area for each particle size range.
The particle size distribution measurement method according to the present invention uses a learning model that supports instance segmentation, making it possible to separately detect overlapping or contacting components on a pixel-by-pixel basis. As a result, for example, there is no need to rectify the object before photographing it, which can reduce equipment costs. Furthermore, since the particle size distribution is measured using a mass conversion factor obtained from a photographed image of the object, the particle size distribution can be measured with high accuracy.

This invention makes it possible to measure particle size distribution in a timely manner while keeping costs down.

1 is a schematic configuration diagram of a particle size distribution measuring system according to a first embodiment of the present invention. 10 is an image of an image captured by a photographing unit. FIG. 2 is a functional configuration diagram of a processing unit (particle size distribution measuring device). This is an image of the detection result in which the detection result is reflected in the captured image. FIG. 5 is an enlarged view of an area K in FIG. 4 . 10 is a flowchart illustrating an example of processing in a calibration mode. 10 is a flowchart illustrating a process in a learning mode. 10 is a flowchart illustrating a process in a particle size distribution estimation mode. FIG. 10 is a schematic configuration diagram of a particle size distribution measuring system according to a second embodiment of the present invention. 10 is an image of an image captured by a photographing unit. 1A is a diagram illustrating an image of a masonry pile as an example of an object for which a particle size distribution is to be estimated, and FIG. 1B is a diagram illustrating a detection result. This is an example of information used for quality control of locally generated materials.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Each drawing is merely a schematic illustration to allow a sufficient understanding of the present invention. Therefore, the present invention is not limited to the illustrated examples. In each drawing, common or similar components are designated by the same reference numerals, and redundant explanations thereof will be omitted.
[First embodiment]
<Configuration of particle size distribution measuring system according to the first embodiment>
The configuration of a particle size distribution measuring system 1 according to the first embodiment will be described with reference to Figures 1 and 2. Figure 1 is a schematic diagram of the particle size distribution measuring system 1 according to the first embodiment. Figure 2 is an image of an image captured by the image capturing unit 20.
The particle size distribution measurement system 1 is a system that measures the particle size distribution of an object based on an image acquired by photographing the object. The object is composed of granular components, such as a mixed material containing coarse particles (aggregate, etc.) and fine particles (soil, sand, etc.). The size of the components is not particularly limited and may be, for example, from several millimeters to several meters.
The particle size distribution measurement system 1 can be used in various situations requiring measurement of particle size distribution, and the object to be measured is not particularly limited in type, use, etc. In this embodiment, it is assumed that the particle size distribution measurement system 1 is used in CSG (Cemented Sand and Gravel) dam construction, and material 9 (on-site generated material for CSG) used in the CSG dam construction is the object to be measured for particle size distribution (hereinafter, may be referred to as "measurement object").

As shown in FIG. 1, the particle size distribution measuring system 1 is made up of a transport unit 10, an imaging unit 20, and a processing unit 30 (particle size distribution measuring device).
The conveying unit 10 conveys the object to be measured. In this embodiment, the conveying unit 10 is a belt conveyor, and includes upstream and downstream pulleys 11, an endless belt 12 wound around both pulleys 11, a motor (not shown) for rotating the upstream pulley 11, and an inverter (not shown) for changing the rotation speed of the motor, and is capable of continuously conveying on-site generated materials for CSG. A hopper, for example, is installed downstream of the conveying unit 10, and the material 9 flowing down from the conveying unit 10 is stored in the hopper.
The photographing unit 20 photographs the object to be measured. The photographing unit 20 is positioned so as to photograph the material 9. For example, as shown in FIG. 1, the photographing unit 20 is positioned downstream of the conveying unit 10 and configured to photograph the material 9 flowing vertically downward (free fall) from the downstream end of the belt conveyor from a horizontal direction (front direction). The photographing unit 20 is preferably a video camera or a high-resolution digital camera capable of acquiring digital images. FIG. 2 shows an image 8 photographed by the photographing unit 20. As shown in FIG. 2, the photographed image 8 shows the material 9 flowing down from the conveying unit 10. The photographing unit 20 continuously photographs the object to be measured and outputs the photographed images 8 to the processing unit 30. The photographing unit 20 may output a portion of the continuously photographed photographed images 8 to the processing unit 30.
The processing unit 30 is a computer consisting of a CPU (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), etc., and acquires the image 8 captured by the photographing unit 20 and calculates the particle size distribution of the material 9 based on this captured image 8.

Next, the processing of the processing unit 30 (particle size distribution measuring device) will be described with reference to Fig. 3. Fig. 3 is a functional configuration diagram of the processing unit 30 (particle size distribution measuring device).
The processing unit 30 learns the constituents of the measurement object (material 9) by machine learning and detects the constituents that appear in the captured image 8 captured by the photographing unit 20. The processing unit 30 then calculates the size (particle size) of each constituent based on the detection results and determines the particle size distribution of the measurement object in the captured image 8. The processing unit 30 repeatedly calculates the particle size distribution for multiple captured images 8 and estimates the particle size distribution of the entire object by statistically processing (e.g., averaging) the results.
In this embodiment, it is assumed that components are detected from the captured image 8 using deep learning, which is one of the machine learning techniques, and a case will be described in which a MASK R-CNN (Regions with Convolutional Neural Network) that supports instance segmentation is used.
Here, segmentation outputs the class to which each pixel in an image belongs, while instance segmentation further divides the area into individual objects (extracting the object area of each individual object). Therefore, instance segmentation makes it possible to separately detect overlapping or touching objects of the same class on a pixel-by-pixel basis.

MASK R-CNN is an extension of Faster R-CNN. Faster R-CNN simplifies and speeds up the processing in R-CNN. R-CNN applies a technique called selective search to extract regions of interest (ROIs), which are then transformed to the input size of the convolutional neural network and input individually. Faster R-CNN uses a region proposal network (RPN) instead of selective search. RPN is composed of a convolutional neural network, processes up to the intermediate layers of the network to obtain a feature map, and determines whether each element position in the feature map is an object or not. MASK R-CNN not only performs object recognition but also segmentation simultaneously on the RPN of Faster R-CNN. The processing of MASK R-CNN is publicly known, so a detailed description will be omitted. MASK R-CNN is explained in, for example, the following document:
Kaiming He et al., "Mask R-CNN," Internet <https://arxiv.org/pdf/1703.06870.pdf>

In this way, MASK R-CNN supports instance segmentation, making it possible to determine whether or not a pixel is a target component. The processing unit 30 then counts the pixels in the captured image that are determined to be components, calculates the area and size (particle size) of each component, and calculates the particle size distribution based on the area corresponding to the particle size classification.
MASK R-CNN supports the detection of multiple types of objects (classification into multiple classes), and can recognize the type (class) of each object when multiple types of objects are captured in a single captured image. In this embodiment, a mixed material is assumed as the object for which particle size distribution is to be measured, and a case where coarse particles such as aggregate are detected will be described. In other words, the only class detected is the "coarse particle class" related to coarse particles, and fine particles such as soil and sand are not detected. Note that the class classification may be other than that described in this embodiment.

As shown in FIG. 3 , the processing unit 30 operates in three modes: a "calibration mode," a "learning mode," and a "particle size distribution estimation mode," and has the functions necessary for these operations. The calibration mode is a mode for making adjustments according to the situation in which the photographing unit 20 is installed. The learning mode is a mode for learning the relationship between the object shown in the photographed image 8 and the constituents of the material 9 to be detected. The processing unit 30 performs learning using the photographed image 8 taken by the photographing unit 20. The particle size distribution estimation mode is a mode for detecting the constituents of the material 9 shown in the photographed image 8 and estimating the particle size distribution of the material 9 based on the detection results. The processing unit 30 estimates the particle size distribution using the photographed image 8 taken by the photographing unit 20.
Operation in the calibration mode is performed before operation in the learning mode and particle size distribution estimation mode. Operation in the particle size distribution estimation mode is performed after operation in the learning mode. Note that after adjustment in the calibration mode is completed, the positional relationship between the photographing unit 20 and the transport unit 10 is not changed (particularly, the positional relationship between the photographing unit 20 and the transport unit 10 in the approaching direction).

The processing unit 30 includes a calibration image input unit 41, a calibration parameter generation unit 42, and a storage unit 43. These functions are mainly related to the "calibration mode."
The processing unit 30 also includes a learning video input unit 51, a teacher data generation unit 52, a deep learning unit 53, and a storage unit 54. These functions are mainly related to the "learning mode."
The processing unit 30 also includes an estimation image input unit 61, an aggregate detection unit 62 (component detection unit), an area calculation unit 63, a diameter determination unit 64, a mass conversion unit 65, and a particle size distribution estimation unit 66. These functions are mainly related to the "particle size distribution estimation mode."

(Calibration mode functions)
The calibration image input unit 41 has a function for inputting an image to be used for calibration (a calibration image, an “image correction image”) from the imaging unit 20 .
The calibration parameter generator 42 has a function for calculating calibration parameters from the input calibration image (image-corrected image). The calibration parameters include, for example, information necessary for image correction and coefficients indicating the relationship between the real space length and the pixels of the captured image. The calibration parameters are stored in the memory 43 and are used, for example, to determine the diameter of the material 9 when estimating the particle size distribution.

(Learning mode functions)
In the learning mode, learning is performed so that object detection (including position identification) and object recognition can be performed from the captured image 8.
The learning video input unit 51 has a function for inputting captured images 8 (learning videos) used for learning from the imaging unit 20. The learning videos are preferably captured under the same conditions as those used to estimate the particle size distribution, for example, videos of material 9 flowing down while the conveying unit 10 is operating. In addition to the learning videos, information related to the imaging rate of the imaging unit 20 is input to the learning video input unit 51, and the learning video input unit 51 outputs, for example, at least a portion of the frames (captured images 8) that make up the learning videos to the teacher data generation unit 52.

The teacher data generation unit 52 has a function for generating teacher data. The teacher data generation unit 52 displays the captured image 8 acquired from the learning video input unit 51 on a screen and accepts tags (e.g., information indicating that it is a composition) set by a person (e.g., an examiner). The teacher data generation unit 52 creates teacher data by tagging the captured image 8 with, for example, the following information:
- Information on the frame (polyline) surrounding the aggregate, which is a component - Information on the class of the aggregate, which is a component - Information on the mask of the aggregate, which is a component The teacher data generation unit 52 generates teacher data that associates the captured image 8 with information on the tag, and outputs the generated teacher data to the deep learning unit 53.
The deep learning unit 53 has a function for training a learning model (network). The deep learning unit 53 of this embodiment trains MASK R-CNN using the training data created by the training data generation unit 52. The deep learning unit 53 trains the learning model by machine learning so that the learning model can detect coarse grain material 9a (aggregate, etc.) contained in the material 9 by inputting the photographed image 8 of the material 9 photographed by the photographing unit 20. The learning model (trained model) that has completed training is stored in the memory unit 54.

(Functions related to particle size distribution estimation mode)
In the particle size distribution estimation mode, object detection (including position identification) and object recognition are performed from a captured image 8 of an object, and the particle size distribution of the object is estimated from the results.
The estimation video input unit 61 has a function for inputting captured images 8 (estimation video) used for estimating particle size distribution from the imaging unit 20. The estimation video is preferably captured under the same conditions as the learning video used in learning, for example, video of material 9 flowing down while the conveying unit 10 is operating. In addition to the estimation video, information regarding the imaging rate of the imaging unit 20 is input to the estimation video input unit 61, and at least a portion of the frames (captured images 8) that make up the estimation video is output to the aggregate detection unit 62, for example.
The aggregate detection unit 62 has a function of detecting coarse granular material 9a (aggregate, etc.) contained in the material 9 from the captured image 8 captured by the imaging unit 20. In this embodiment, the aggregate detection unit 62 detects the coarse granular material 9a (aggregate, etc.) using a trained model trained by the deep learning unit 53. The aggregate detection unit 62 inputs the captured image 8 for estimation to the trained model and obtains the detection result (information on the detected coarse granular material 9a (aggregate, etc.)) output from the trained model.
The detection results by the aggregate detection unit 62 are shown in Figures 4 and 5. Figure 4 is an image of the detection result in which the detection results are reflected in the photographed image. Figure 5 is an enlarged view of area K in Figure 4. In Figure 4, the detected objects (coarse granular materials 9a (aggregates, etc.)) are highlighted by hatching or the like. As shown in Figure 5, even if multiple coarse granular materials 9a are photographed overlapping or touching, they are detected as different objects. The aggregate detection unit 62 outputs the detection results to the area calculation unit 63.

The area calculation unit 63 has a function of calculating the area of the coarse granular material 9a (aggregate, etc.) detected by the aggregate detection unit 62. The area calculation unit 63 calculates the area (number of pixels) of each coarse granular material 9a in the photographed image 8, and outputs the calculation result to the diameter determination unit 64 and the mass conversion unit 65.
The diameter determination unit 64 has a function of determining the diameter of the coarse granular material 9a (aggregate, etc.) detected by the aggregate detection unit 62. The diameter determination unit 64 adds diameter information to information on the detected coarse granular material 9a and outputs the information to the mass conversion unit 65.
The mass conversion unit 65 has a function for calculating a mass conversion coefficient (also referred to as a "weight conversion coefficient") for each particle size classification. The mass conversion coefficient is the mass per predetermined area of the coarse granular material 9a (aggregate, etc.) in the captured image 8. The particle size classification is set as five classifications, for example, 5 mm or less, greater than 5 mm and less than 10 mm, greater than 10 mm and less than 20 mm, greater than 20 mm and less than 40 mm, and greater than 40 mm and less than 80 mm. The mass conversion unit 65 has known values for calculating the mass conversion coefficients by some method (e.g., prior registration by the tester). The mass conversion unit 65 calculates the mass conversion coefficient for each particle size classification based on the total area of the coarse granular material 9a (aggregate, etc.) for each particle size classification and the known values for calculating the mass conversion coefficients. The mass conversion unit 65 outputs the calculated mass conversion coefficients to the particle size distribution estimation unit 66. The mass conversion factor may be calibrated according to the moisture content (surface moisture content) of the aggregate, etc. The greater the moisture content, the larger the particle size of the aggregate.
The particle size distribution estimation unit 66 calculates the particle size distribution of the material 9 based on the total area of each particle size division of the coarse granular material 9a (aggregate, etc.) in each captured image 8 and the mass conversion coefficient. The particle size distribution estimation unit 66 preferably calculates the particle size distribution for each captured image 8 and estimates the particle size distribution of the material 9 by statistically processing (e.g., averaging) the results. This makes it possible to detect all of the material 9 transported by the transport unit 10, thereby increasing the number of samples used to estimate the particle size distribution. This allows for more accurate particle size distribution measurements. The particle size distribution estimation unit 66 then displays the calculated particle size distribution on a display unit (not shown), for example.

<Regarding the particle size distribution estimation method of the particle size distribution measurement system according to the first embodiment>
A particle size distribution estimation method of the particle size distribution measurement system 1 will be described with reference to Fig. 6 to Fig. 8 (and Fig. 1 to Fig. 5 as appropriate). Fig. 6 is an example of a flowchart showing processing in a calibration mode. Fig. 7 is an example of a flowchart showing processing in a learning mode. Fig. 8 is an example of a flowchart showing processing in a particle size distribution estimation mode.
(Calibration mode operations)
The operation of the calibration mode will be described with reference to FIG. 6 . First, the tester operates the photographing unit 20 to photograph a calibration image, and the calibration image input unit 41 accepts input of the photographed calibration image (step S11). Next, the calibration parameter generation unit 42 creates calibration parameters based on the input calibration image (step S12). The created calibration parameters are stored in the storage unit 43 and used during operation in the particle size distribution estimation mode. Note that after the calibration parameters are created, the environment related to photographing by the photographing unit 20 (such as the position of the photographing unit 20) should not be changed.

(Learning mode operations)
The operation of the learning mode will be described with reference to FIG. 7 . The tester performs the operation in the learning mode after completing the operation in the calibration mode. First, the tester operates the image capture unit 20 to capture learning video, and the learning video input unit 51 begins inputting the learning video (step S21). Next, the teacher data generation unit 52 displays frames of the learning video (captured images 8) acquired from the learning video input unit 51 on the screen and creates teacher data by accepting tags (e.g., information indicating that the material is a coarse-grained material 9a) set by the tester (step S22). The teacher data generation unit 52 creates multiple pieces of teacher data, for example, using frames (captured images 8) captured at different times. Next, the deep learning unit 53 trains a learning model (e.g., MASK R-CNN) through deep learning using the teacher data created in step S22 (step S23). This creates a trained model, which is then stored in the memory unit 54. The trained model detects coarse grain material 9a (aggregate, etc.) contained in the material 9 by inputting a photographed image 8 of the material 9 photographed by the photographing unit 20.

(Operations related to particle size distribution estimation mode)
The operation of the particle size distribution estimation mode will be described with reference to FIG. 8 . The tester performs the operation in the particle size distribution estimation mode after completing the operations in the calibration mode and the learning mode. First, the tester operates the image capture unit 20 to capture an estimation video, and the estimation video input unit 61 begins inputting the estimation video (step S31). Next, the aggregate detection unit 62 detects coarse granular material 9a (aggregates) from the estimation video using the trained model trained by deep learning in step S23 (step S32). In the detection in step S32, frames (captured images 8) constituting the estimation video are input to a trained model (e.g., MASK R-CNN), and the detection result (information on the detected coarse granular material 9a (aggregates, etc.)) output from the trained model is obtained.
Next, the diameter determination unit 64 determines the diameter of the coarse granular material 9a (aggregate) in the photographed image 8 based on the coarse granular material 9a (aggregate) detected in step S32 (step S33A). Note that the diameter determination unit 64 may, for example, display the detection result on a display unit to have the tester select both ends of the coarse granular material 9a, and determine the actual diameter of the material 9 (aggregate) based on the number of pixels between the selected ends and the calibration parameters. Furthermore, the area calculation unit 63 calculates the area of each of the coarse granular material 9a (aggregate) detected in step S32 (step S33B).
Next, the mass conversion unit 65 calculates a mass conversion coefficient for each particle size range (step S34). The particle size distribution estimation unit 66 then estimates the particle size distribution based on the total area and mass conversion coefficient for each particle size range of the coarse granular material 9a (aggregate) (step S35). The particle size distribution estimation unit 66 then displays the estimated particle size distribution on a display unit (not shown).

As described above, the particle size distribution measuring system 1 according to the first embodiment uses a learning model that supports instance segmentation, making it possible to separately detect overlapping coarse granular materials 9a (aggregates) and contacting coarse granular materials 9a (aggregates) on a pixel-by-pixel basis. Here, segmentation outputs the class to which each pixel in an image belongs, while instance segmentation further divides the region into individual units (extracts the object region of each individual unit). As a result, for example, there is no need to rectify the material 9 before capturing an image, which can reduce equipment costs.

Second Embodiment
<Configuration of particle size distribution measuring system according to the second embodiment>
The configuration of a particle size distribution measuring system 101 according to the second embodiment will be described with reference to Fig. 9 and Fig. 10. Fig. 9 is a schematic diagram of the particle size distribution measuring system 101 according to the second embodiment. Fig. 10 is an image of an image captured by the image capturing unit 20.
The components of a particle size distribution measuring system 101 according to the second embodiment are the same as those of the particle size distribution measuring system 1 according to the first embodiment. As shown in Fig. 9, the particle size distribution measuring system 101 is composed of a transport unit 10, an imaging unit 20, and a processing unit 30 (particle size distribution measuring device). In the second embodiment, the arrangement of the imaging unit 20 differs from that in the first embodiment.
The photographing unit 20 in the second embodiment is disposed above the conveying unit 10 and is configured to photograph the material 9 moving horizontally facing vertically downward. The photographing unit 20 is preferably a video camera or a high-resolution digital camera capable of acquiring digital images. FIG. 10 shows an image 8 photographed by the photographing unit 20. As shown in FIG. 10, the photographed image 8 shows the material 9 being conveyed on the conveying unit 10. The photographing unit 20 outputs the photographed results to the processing unit 30.

The particle size distribution measuring system 101 according to the second embodiment described above can also achieve substantially the same effects as those of the first embodiment.
<Modification>
In each embodiment, MASK R-CNN has been described as an example of a learning model (network) that supports instance segmentation. However, other learning models can be used as long as they can individually detect the estimation target pixel by pixel from the captured image 8.
In each embodiment, the object for which the particle size distribution is to be estimated is material 9 (on-site generated material for CSG) used in CSG dam construction. However, the object for which the particle size distribution is to be estimated is not limited to on-site generated material, and the particle size distribution of various objects can be estimated. For example, as shown in FIG. 11 , it is also possible to detect stones 109a in masonry 109 and estimate its particle size distribution. FIG. 11 is a diagram explaining the object for which the particle size distribution is to be estimated, where (a) is an image of a photographed image of masonry 109 as an example of the object, and (b) is a diagram showing the detection results.

1,101 Particle size distribution measurement system 8 Photographed image 9 Material (object)
9a Coarse grain material (constituent)
10 Transport unit 20 Photography unit 30 Processing unit 41 Calibration video input unit 42 Calibration parameter generation unit 43 Storage unit 51 Learning video input unit 52 Teacher data generation unit 53 Deep learning unit 54 Storage unit 61 Estimation video input unit 62 Aggregate detection unit (component detection unit)
63 Area calculation unit 64 Diameter determination unit 65 Mass conversion unit 66 Particle size distribution estimation unit

Claims (4)

A particle size distribution measurement system for measuring the particle size distribution of an object,
an imaging unit that images the object;
a processing unit that calculates the particle size distribution of the photographed object,
The processing unit
a component detection unit that detects components of the object in the captured image;
an area calculation unit that calculates an area of the component in the captured image;
a particle size distribution estimation unit that estimates a particle size distribution of the object from the detected constituents,
the component detection unit individually detects the components appearing in the captured image using a trained model that has learned the components using a learning model corresponding to instance segmentation;
the particle size distribution estimation unit estimates the particle size distribution based on a total area for each particle size range in the captured image and a mass conversion coefficient that is a mass per predetermined area for each particle size range.
A particle size distribution measuring system characterized by: The learning model is MASK R-CNN,
The trained model is machine-trained to detect the component by inputting the captured image.
2. The particle size distribution measuring system according to claim 1 . The object is a locally generated material for CSG,
a conveying unit for conveying the on-site generated materials;
The photographing unit continuously photographs the on-site generated materials being transported,
the particle size distribution estimation unit obtains a particle size distribution for each photographed image from the photographed image, and estimates the particle size distribution of the on-site generated material by statistically processing the results.
3. The particle size distribution measuring system according to claim 1 or 2 . A particle size distribution measurement method for measuring the particle size distribution of an object, comprising:
an imaging step of imaging the object;
and a processing step of calculating the particle size distribution of the photographed object,
The processing step comprises:
a component detection step of detecting components of the object in the captured image;
an area calculation step of calculating an area of the component in the captured image;
a particle size distribution estimation step of estimating a particle size distribution of the object from the detected constituents,
In the component detection step, the components appearing in the captured image are individually detected using a trained model that has learned the components using a learning model corresponding to instance segmentation;
In the particle size distribution estimation step, the particle size distribution is estimated based on a total area for each particle size division in the captured image and a mass conversion coefficient which is a mass per predetermined area for each particle size division.
A particle size distribution measuring method characterized by: