JP7718860B2 - Aggregate particle size distribution estimation system - Google Patents

Description

The present invention relates to an aggregate particle size distribution estimation system.

BACKGROUND ART Conventionally, when constructing a concrete structure, aggregate is mixed into ready-mixed concrete in order to reduce shrinkage, cracks, and the like that occur when the concrete hardens.
In other words, concrete is composed of four materials: water, cement, coarse aggregate (gravel), and fine aggregate (sand), and quality control of the aggregate, which accounts for the majority of the weight, is important.
The particle size distribution of the coarse aggregate mixed into concrete varies depending on the use, purpose, and strength of the concrete, and the type of concrete.
For this reason, the grading of coarse aggregate and fine aggregate is used as an important parameter in calculating the strength of concrete structures.

On the other hand, aggregates are sieved in an attempt to make the particle size of the aggregate uniform, but even aggregates sieved to the same particle size (particle diameter of the aggregate) have a particle size distribution (distribution of particle diameters of aggregate particle groups) in which aggregates of different particle sizes are mixed.
For this reason, the method for measuring the particle size of aggregate is specified in the Japanese Industrial Standards (JIS A1102:2014) as the "aggregate sieving test method."
However, the above-mentioned "aggregate sieving test method" places a heavy burden on the worker conducting the test, and it is difficult to carry out the test frequently.

Therefore, in order to reduce the burden on the worker compared to the "aggregate sieving test method," there is a system in which coarse aggregate is spread in a single layer on a measuring plate, and the captured image of the aggregate is analyzed to measure the particle size distribution (see, for example, Patent Document 1).
There is also a system that scatters coarse aggregate onto a belt conveyor, captures images of the coarse aggregate on the belt conveyor, extracts the contours of each of the captured aggregates, and measures the particle size distribution (see, for example, Patent Document 2).

There is also a system that measures particle size distribution by scattering coarse aggregate in a cylindrical shape and allowing it to fall, capturing images of the falling coarse aggregate, extracting the outline of each of the captured coarse aggregates, and measuring the particle size of each of the aggregates using these outlines (see, for example, Patent Document 3).
There is a system that measures particle size by converting the shape of each fine aggregate into an equivalent circular shape by image processing of a microphotograph of the fine aggregate (see, for example, Non-Patent Document 1).

Japanese Patent Application Laid-Open No. 2008-268051 Japanese Patent Application Laid-Open No. 2003-010726 JP 2016-070714 A

Ito, K., Tsujimoto, K., Suzuki, K., Tsuji, Y., "Development of a simple inspection method for fine aggregate particle size using an image processing system," Cement Science and Concrete Technology. 63, 227-232, 2009.

However, unlike the "aggregate sieving test method," each of the above-mentioned prior arts can reduce the burden on the worker conducting the test, but each requires special equipment when imaging the aggregate and analyzing the particle size distribution.
For this reason, it is difficult to install the above-mentioned equipment at construction sites where concrete structures are actually created, or at ready-mixed concrete (ready-mixed concrete) manufacturing plants where concrete is produced.

The present invention was made in consideration of the above circumstances, and aims to provide an aggregate particle size distribution estimation system that reduces the burden on workers who measure the particle size distribution of aggregates (coarse aggregate, fine aggregate) mixed in when making concrete, and can easily determine the particle size distribution of the aggregate.

To solve the above problem, the aggregate particle size distribution estimation system of the present invention is an aggregate particle size distribution estimation system that estimates the particle size distribution of aggregate (coarse aggregate or fine aggregate) mixed in when producing concrete, and is characterized by comprising: a shape feature extraction unit that extracts shape feature quantities calculated from at least the contour length, contour area, and circularity of the aggregate as feature quantities of each of the captured aggregate from an aggregate surface image, which is an image of the surface of piled aggregate; a color feature extraction unit that extracts color feature quantities calculated from at least the gradient, hue, saturation, and brightness of the aggregate's color components as feature quantities of each of the captured aggregate from the aggregate surface image; and a particle size distribution estimation unit that receives the shape feature quantities and color feature quantities and estimates the particle size distribution of the aggregate using a predetermined machine learning model.

The present invention provides an aggregate particle size distribution estimation system that reduces the burden on workers who measure the particle size distribution of aggregates (coarse aggregate, fine aggregate) mixed into concrete when making it, and can easily determine the particle size distribution of the aggregate.

1 is a block diagram showing an example of the configuration of an aggregate particle size distribution estimation system according to an embodiment of the present invention. 10A and 10B are diagrams illustrating the capturing of an aggregate surface image in this embodiment. 3 is a diagram showing an example of the structure of a feature table written and stored in a feature storage unit 19 in this embodiment. FIG. 10 is a diagram showing an example of the configuration of a table showing the correspondence between aggregate image identification information and particle size distribution of coarse aggregate. FIG. 10 is a flowchart showing an example of the operation of a process for classifying the particle size distribution of coarse aggregate, the particle size distribution of which is unknown, into each of known particle size distributions, performed by the aggregate particle size distribution estimation system according to the present embodiment.

Figure 1 is a block diagram showing an example configuration of an aggregate particle size distribution estimation system according to one embodiment of the present invention. In Figure 1, the aggregate particle size distribution estimation system 10 includes a data input/output unit 11, an image preprocessing unit 12, a shape feature extraction unit 13, a color feature extraction unit 14, a machine learning model generation unit 15, a particle size distribution estimation unit 16, a feature importance evaluation unit 17, an image data storage unit 18, a feature storage unit 19, a machine learning model storage unit 20, and a training data storage unit 21. Below, we will use coarse aggregate as an example of aggregate, but similar processing can also be performed on fine aggregate.

The data input/output unit 11 transmits and receives data to and from external devices. The data input/output unit 11 also writes and stores aggregate surface image data, which is an image of coarse aggregate supplied from an external device, in the image data storage unit 18, along with aggregate surface image identification information. The aggregate surface image is captured in a closed space, such as a dark box that blocks external light.

FIG. 2 is a diagram illustrating how to capture an aggregate surface image in this embodiment.
Fig. 2(a) is a diagram showing how coarse aggregate, which is the object to be imaged, is piled up in a container with an open top, such as a bucket, and the surface of the piled coarse aggregate is imaged to obtain an image 101. Fig. 2(b) shows how the acquired image 101 is trimmed so that the image area is limited to the surface of the coarse aggregate, and this trimming result is an aggregate surface image 102.

Figure 2(c) shows an example of a method for capturing an aggregate surface image. A bucket 201 containing a pile of coarse aggregate 202 is placed in a container 200, such as a dark box that blocks external light. The surface of the accumulated coarse aggregate 202 is then captured by an imaging device 210 at the timing when a flash 211 is emitted. This allows the imaging environment for capturing aggregate surface images to be constant, and multiple aggregate surface images can be captured under the same conditions. Furthermore, in this embodiment, the camera parameters of the imaging device are also the same when capturing each aggregate surface image.

Returning to Figure 1, the captured image pre-processing unit 12 performs the trimming process described above to obtain an aggregate surface image, and also performs processes such as erasing the shadows of other coarse aggregates in the image area of each coarse aggregate in the aggregate surface image. The captured image pre-processing unit 12 then writes and stores the processed aggregate surface image in the image data storage unit 18 in association with the aggregate surface image identification information.

The shape feature extraction unit 13 performs a process to extract the contours of each of the coarse aggregates in the aggregate surface image. The shape feature extraction unit 13 then calculates the perimeter length of each of the extracted contours as the contour length, adds up the contour lengths of all the extracted contours, and obtains the total contour length value, which is the sum of the contour lengths, as one of the shape feature quantities. The shape feature extraction unit 13 also calculates the average value of the contour lengths of each of the contours as the contour length average value, and calculates the standard deviation of the contour lengths of each of the contours as the contour length standard deviation, as shape feature quantities.

Similarly, the shape feature extraction unit 13 calculates the area inside the extracted contour as the contour area, adds up the contour areas of all the extracted contours, and obtains the total contour area value, which is the sum of the contour areas, as one of the shape feature quantities. Furthermore, the shape feature extraction unit 13 calculates the average value of the contour areas of each contour as the contour area average, and calculates the standard deviation of the contour areas of each contour as the contour area standard deviation, as shape feature quantities.

The shape feature extraction unit 13 calculates the contour circularity, which is the circularity of the shape of the contour line of the extracted contour, and then adds up the contour circularities of all the extracted contours to obtain a contour circularity total value, which is the total value of the contour circularities, as one of the shape feature quantities. Here, the shape feature extraction unit 13 calculates the circularity circ of each contour line of the coarse aggregate using the following equation (1). In this equation (1), area is the contour area of the coarse aggregate, and length is the contour length of the coarse aggregate.
circ=4π×area/(length) ² …(1)

Furthermore, the shape feature extraction unit 13 calculates the average value of the contour circularity for each contour as a shape feature, and calculates the standard deviation of the contour circularity for each contour as a contour circularity standard deviation.The shape feature extraction unit 13 then writes and stores each of the acquired shape feature values in the feature storage unit 19.

The color feature extraction unit 14 extracts color information for each pixel in the aggregate surface image and acquires color feature quantities for the aggregate surface image. For example, the color information is calculated by calculating the average gradient values for each of the color components R, G, and B for each pixel in the aggregate surface image, to acquire the R average value, G average value, and B average value. At this time, the color feature extraction unit 14 calculates the grayscale for each pixel in the aggregate surface image from the color components R, G, and B, and calculates the grayscale average value, which is the average value of the grayscales. The color feature extraction unit 14 then writes and stores the acquired R average value, G average value, and B average value in the feature storage unit 19.

For color information, the gradients of the color components R, G, B, and grayscale are each divided into multiple ranges. For example, if there are 256 gradations, they are classified into eight classes: 0-31, 32-63, 64-95, 96-127, 128-159, 160-191, 192-223, and 224-255. The color feature extraction unit 14 then divides the number of pixels included in each class by the total number of pixels in the aggregate surface image, and uses the division result for each class as the occurrence rate of that class.

For example, the color feature extraction unit 14 divides the number of pixels in the eight classes of color component R by the total number of pixels in the aggregate surface image, obtaining division values for classes from 0-31 to 224-255, and then obtains the occurrence rate for each class of color component R's gradient. The color feature extraction unit 14 also standardizes the occurrence rates so that the maximum and minimum occurrence rates in the eight classes of color component R are 1 and 0, respectively, and writes and stores each of the obtained occurrence rates of color component R (color component R occurrence rate for each of the eight classes) in the feature storage unit 19. The color feature extraction unit 14 also performs the same process on each of color components G, B, and grayscale as for color component R, and writes and stores each of the obtained color component G occurrence rate, color component B occurrence rate, and grayscale occurrence rate in the feature storage unit 19.

In addition, the color feature extraction unit 14 calculates the average values of hue H, saturation S, and lightness V in the HSV display for each pixel of the aggregate surface image as other color feature quantities, thereby obtaining the average hue, average saturation, and average lightness values.The color feature extraction unit 14 then writes and stores the obtained average hue, average saturation, and average lightness values in the feature storage unit 19.The color feature extraction unit 14 divides the range of values for hue H, saturation S, and lightness V into eight classes, similar to the gradient, and classifies each pixel into one of these classes.

For example, the color feature extraction unit 14 converts the values of the hue scale (0 to 359) indicating the hue H into 256 gradations (0 to 255), classifies them into eight classes, and, as with the gradations, classifies each pixel according to the hue H and divides the number of pixels classified into each class by the total number of pixels in the aggregate surface image. The color feature extraction unit 14 then standardizes the occurrence rates so that the maximum and minimum occurrence rates in the hue H class are 1 and 0, respectively, to determine the standardized hue occurrence rates. The color feature extraction unit 14 writes and stores each of the acquired hue occurrence rates (the hue occurrence rate for each of the eight classes) in the feature storage unit 19.

Similarly, the color feature extraction unit 14 converts the numerical values representing saturation S, from 0% to 100%, into 256 gradations from 0 to 255, and classifies them into eight classes. As with the gradations, it classifies each pixel according to its saturation S and divides the number of pixels classified into each class by the total number of pixels in the aggregate surface image. The color feature extraction unit 14 then standardizes the occurrence rates so that the maximum and minimum occurrence rates in the saturation S classes are 1 and 0, respectively, to obtain standardized saturation occurrence rates. The color feature extraction unit 14 writes and stores each of the acquired saturation occurrence rates (the saturation occurrence rates for each of the eight classes) in the feature storage unit 19.

The color feature extraction unit 14 also converts the values from 0% to 100% indicating the brightness V into 256 gradations from 0 to 255, classifying them into eight classes. As with the gradients, it classifies each pixel according to its brightness V and divides the number of pixels classified into each class by the total number of pixels in the aggregate surface image. The color feature extraction unit 14 then standardizes the occurrence rates so that the maximum and minimum occurrence rates in the brightness V classes are 1 and 0, respectively, to determine the standardized brightness occurrence rate. The color feature extraction unit 14 writes and stores each of the acquired brightness occurrence rates (the brightness occurrence rate for each of the eight classes) in the feature storage unit 19.

Figure 3 shows an example of the structure of a feature table written and stored in the feature storage unit 19 in this embodiment. Figure 3(a) shows an example of the structure of a shape feature table, which is one of the feature tables. For each record, the shape feature table has columns for the above-mentioned aggregate image identification information, total contour length, average contour length, standard deviation of contour length, total contour area, average contour area, standard deviation of contour area, total contour circularity, average contour circularity, and standard deviation of contour circularity as feature groups of shape features.

Figure 3(b) shows an example of the configuration of a color feature table, which is one of the feature tables. The color feature table has columns for each record as a group of color feature values: the aggregate image identification information described above, R average value, R color component appearance rate, G average value, G color component appearance rate, B average value, B color component appearance rate, grayscale average value, grayscale appearance rate, hue average value, hue appearance rate, saturation average value, saturation appearance rate, lightness average value, and lightness appearance rate. Furthermore, the R color component appearance rate, G color component appearance rate, B color component appearance rate, grayscale appearance rate, hue appearance rate, saturation appearance rate, and lightness appearance rate are each written in a column as a set of eight appearance rate values for each class.

Returning to FIG. 1 , the machine learning model generation unit 15 inputs each of the above-mentioned shape features and color features and generates a machine learning model that outputs an estimated value of the aggregate particle size distribution (hereinafter simply referred to as particle size distribution) of the coarse aggregate. That is, the machine learning model in this embodiment is a classifier that classifies multiple particle size distributions. By inputting a group of feature data for each of the shape features and color features, the machine learning model outputs a confidence level (an estimated value indicating the likelihood, the estimated value p described below) when the particle size distribution of the coarse aggregate captured in the aggregate surface image is classified into each label. For example, in this embodiment, three types of particle size distributions, particle size distributions #1, #2, and #3, are used as labels. Therefore, the machine learning model outputs a confidence level for each of particle size distributions #1, #2, and #3, and the particle size distribution with the highest output confidence level is estimated to be the most likely to be the particle size distribution of the coarse aggregate captured in the aggregate surface image.

The machine learning model generation unit 15 captures images of the surface of coarse aggregate with a known particle size distribution and generates a data group for training the machine learning model. In this embodiment, for example, a multilayer perceptron classifier is used as the machine learning model, but any other classifier using a neural network or the like may also be used. Furthermore, in this embodiment, the data group is divided at a predetermined ratio, and, for example, 80% of the data in the data group is used as a training data group for learning (training) the machine learning model. The remaining 20% of the data group is then used as test data for determining the accuracy rate of the created machine learning model.

The machine learning model generation unit 15 also divides the teacher data group into a training data group and a validation data group at a predetermined ratio, and, for example, designates each of 80% of the data in the teacher data group as the training data group to be used for actual learning (training) of the machine learning model. The machine learning model generation unit 15 then designates each of the remaining 20% of the data in the teacher data group as the validation data group to be used for verifying the accuracy rate of the machine learning model trained with the data in the training data group. Here, the machine learning model generation unit 15 extracts a portion of data at a predetermined rate (20%) from the training data group, replaces the same number of teacher data pieces with those in the validation data group, trains the machine learning model with the data in the new training data group after the replacement, and uses the data in the new validation data group to verify the accuracy rate of the machine learning model. The machine learning model generation unit 15 then performs the data replacement operation in each of the above-mentioned training data group and validation data group a predetermined number of times to train the machine learning model.

Feature values used as training data are written and stored in the training data storage unit 21 as a training data table. This training data table has a similar structure to the shape feature value table and color feature value table shown in Figure 3. The training data table stores each of the shape feature values and color feature values obtained by analyzing training data images, which are images of the aggregate surface captured from coarse aggregate with a known particle size distribution.

The feature importance evaluation unit 17 selects either a shape feature or a color feature in the training data used to train the trained model, and evaluates the impact on the confidence level when the data of that feature is swapped between aggregate surface images.
For example, the contour length data in the shape feature is randomly shuffled among the aggregate surface image identification information, and the shuffled feature set is input to the trained model. The importance of the contour length as a feature for training the machine learning model is determined based on the level of confidence reduction. By shuffling the feature data in this way for all feature values, the importance of the feature in training the machine learning model can be confirmed. For example, using this result, it is possible to compact the machine learning model by not using feature values with confidence reduction values that are less than a predetermined percentage of the largest confidence reduction value for training the machine learning model.
In this embodiment, the feature importance evaluation unit 17 processed the following to extract the top 10 features: total contour circularity value, total contour length value, average saturation value, hue appearance rate (gradation 96-127), hue appearance rate (160-191), brightness appearance rate (gradation 192-223), color component B appearance rate (gradation 192-223), average contour length value, color component B appearance rate (gradation 0-31), and hue appearance rate (0-31).

Figure 4 shows an example of a table showing the correspondence between aggregate image identification information and the particle size distribution of coarse aggregate. Figure 4(a) shows an example of the configuration of the training data information table in the training data storage unit 21. The training data information table has columns for aggregate image identification information, particle size distribution type, and application type for each record. This aggregate image identification information is identification information that individually identifies each aggregate surface image of coarse aggregate with a known particle size distribution used as training data. The particle size distribution indicates the type of particle size distribution of the coarse aggregate from which the aggregate surface image was captured (the type of particle size distribution that serves as a label during learning). In this embodiment, for example, three particle size distribution types are used, each represented as particle size distribution #1, particle size distribution #2, and particle size distribution #3. The application type indicates whether each feature of the aggregate surface image (the training data) is used as data in the training data group or the validation data group, or as test data, respectively, in the training data group.

Figure 4(b) shows an example of the configuration of the estimation result table in the feature storage unit 19. The estimation result table has columns for aggregate image identification information, particle size distribution #1, particle size distribution #2, and particle size distribution #3 for each record. This aggregate image identification information is identification information that individually identifies each aggregate surface image for which the particle size distribution of coarse aggregate is unknown. The particle size distribution #1 column shows an estimated value p (e.g., 0 < p < 1) that indicates the confidence (likelihood) that the particle size distribution of the coarse aggregate, estimated by the machine learning model when the feature values of the coarse aggregate with an unknown particle size distribution are input, is classified into the particle size distribution #1 of the label. Similarly, the particle size distribution #2 column shows an estimated value p that indicates the confidence (likelihood) that the particle size distribution of the coarse aggregate, estimated by the machine learning model when the feature values of the coarse aggregate with an unknown particle size distribution are input, is classified into the particle size distribution #2 of the label. The Particle Size Distribution #3 column shows the estimated value p, which indicates the confidence level that the particle size distribution of the coarse aggregate is classified as particle size distribution #3, as estimated by the machine learning model when the features of the coarse aggregate with an unknown particle size distribution are input.

Figure 5 is a flowchart showing an example of the operation of the aggregate particle size distribution estimation system according to this embodiment, which classifies the particle size distribution of coarse aggregate with an unknown particle size distribution into each of the known particle size distributions. In this embodiment, as an example, there are three known particle size distributions, and the process of generating a machine learning model (classifier) that classifies coarse aggregate with an unknown particle size distribution into these three particle size distributions (labels) and the classification process that classifies coarse aggregate with an unknown particle size distribution into the three labels using the generated machine learning model are explained.

The user piles coarse aggregate with a known particle size distribution into a bucket, and in a closed space that blocks external light, uses an imaging device to capture aggregate surface images, which are surface images of the piled coarse aggregate to be used as training data (step S1). The captured image preprocessing unit 12 then performs preprocessing on the captured aggregate surface images, such as trimming and removing shadows.

The shape feature extraction unit 13 and the color feature extraction unit 14 extract shape features and color features from each of the aggregate surface images taken of a plurality of coarse aggregates, i.e., coarse aggregates having three particle size distributions, and write and store the extracted shape features and color features in the teacher data table in the teacher data storage unit 21. Here, for example, several hundred images of the aggregate surface are taken for each of the three known particle size distributions of the coarse aggregate, and are acquired as a feature data group of the teacher data (step S2).
In addition, the machine learning model generation unit 15 sets 80% of the feature data group in the teacher data table in the teacher data storage unit 21 as teacher data and 20% as test data, and enters these in the usage type column of the teacher data table (steps S3 and S4).

The machine learning model generation unit 15 sets 80% (i.e., 64% of the data group) of the feature data group (i.e., the teacher data group) in the teacher data table in the teacher data storage unit 21 as a training data group, and 20% (i.e., 16% of the data group) as a validation data group, and enters these in the usage type column of the teacher data table (steps S5 and S6).
The machine learning model generation unit 15 then performs learning by adjusting the parameters of the machine learning model (classifier) using data whose use type is the training data group in the training data table in the training data storage unit 21 (steps S7, S8, and S9). At this time, the machine learning model generation unit 15 calculates the accuracy rate of the classification of the machine learning model to be trained using data whose use type is the validation data in the training data table in the training data storage unit 21 (model verification, correct/incorrect determination). The machine learning model generation unit 15 also exchanges the data in the validation data group with a portion of the data in the training data group (e.g., 20% of the training data group), and re-trains the generated machine learning model using the new data in the exchanged training data group. This re-training (re-parameter adjustment) of the machine learning model in which the data in the validation data group and a portion of the data in the training data group have been exchanged is performed a predetermined number of times, thereby generating a trained machine learning model for estimating aggregate particle size distribution (trained model), which is then written and stored in the machine learning model storage unit 20.

The machine learning model generation unit 15 also uses data in the training data table in the training data storage unit 21 with the application type set to test data to calculate the accuracy rate, which is the confidence level (correct/incorrect judgment) of the classification of the machine learning model being trained as the trained model for aggregate particle size distribution estimation (trained machine learning model) (steps S10 and S11). Here, if the accuracy rate of the trained model is low, the machine learning model generation unit 15 may be configured to return to step S3 and train the machine learning model again.

The user piles coarse aggregate with an unknown particle size distribution into a bucket, and in a closed space that blocks external light, an image of the surface of the piled coarse aggregate is captured using an imaging device (step S12). The captured image pre-processing unit 12 then performs pre-processing on the captured aggregate surface image, such as trimming and removing shadows.

The shape feature extraction unit 13 and the color feature extraction unit 14 extract shape features and color features from each of the aggregate surface images taken of coarse aggregate with unknown particle size distribution, and write and store the feature data group in the feature table in the feature storage unit 19 in correspondence with the assigned aggregate image identification information (step S13).
The particle size distribution estimation unit 16 reads out the trained model from the machine learning model storage unit 20, and also reads out a feature data group from the feature table of the feature storage unit 19. Then, the particle size distribution estimation unit 16 inputs the feature data group into the trained model, and writes and stores the confidence levels of each output label in the label column of the result table of the feature storage unit 19 (steps S14 and S15).

According to this embodiment, by inputting shape features and color features extracted from an aggregate surface image, which is a surface image of piled aggregate (e.g., coarse aggregate), to a trained model that has been trained using shape features and color features extracted from an aggregate surface image of aggregate with a known particle size distribution, it is possible to produce concrete without requiring a large-scale particle size distribution measuring device as in the past, reduce the burden on workers who measure the particle size distribution of the aggregate to be mixed in, and easily determine the particle size distribution of the aggregate (coarse aggregate and fine aggregate).
Furthermore, according to this embodiment, it is possible to easily estimate the particle size distribution of aggregates (coarse aggregate and fine aggregate) at construction sites or at businesses that manufacture ready-mixed concrete, as described above, which makes it possible to easily ensure (provide) quality control in the construction of concrete structures and traceability of data on aggregate particle size distribution in the construction of concrete structures, thereby preventing rework due to poor quality.

The program for implementing the functions of the aggregate particle size distribution estimation system 10 of the present invention shown in Figure 1 may be recorded on a computer-readable recording medium, and the program may be loaded into a computer system and executed to estimate the particle size distribution of coarse aggregate. Note that the term "computer system" as used herein includes hardware such as an operating system (OS) and peripheral devices. It also includes a World Wide Web (WWW) system equipped with a homepage environment (or display environment). The term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs (Read Only Memory), and CD-ROMs (Compact Disc-Read Only Memory), as well as storage devices such as hard disks built into computer systems. Furthermore, the term "computer-readable recording medium" also includes devices that retain a program for a certain period of time, such as volatile memory (RAM (Random Access Memory)) within computer systems that function as servers or clients when a program is transmitted via a network such as the Internet or a communication line such as a telephone line.

The above program may also be transmitted from a computer system that stores the program in a storage device or the like to another computer system via a transmission medium, or by transmission waves within the transmission medium. Here, the "transmission medium" that transmits the program refers to a medium that has the function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The above program may also be one that realizes part of the above-mentioned functions. Furthermore, it may be a so-called differential file (differential program) that can realize the above-mentioned functions in combination with a program already recorded on the computer system.

10...Aggregate particle size distribution estimation system 11...Data input/output unit 12...Captured image preprocessing unit 13...Shape feature extraction unit 14...Color feature extraction unit 15...Machine learning model generation unit 16...Particle size distribution estimation unit 17...Feature importance evaluation unit 18...Image data storage unit 19...Feature storage unit 20...Machine learning model storage unit 21...Teacher data storage unit

Claims (5)

This is an aggregate particle size distribution estimation system that estimates the particle size distribution of aggregate mixed in when producing concrete.
a shape feature extraction unit that extracts shape feature quantities calculated from at least a contour length, a contour area, and a circularity of each of the aggregates as feature quantities of each of the captured aggregates from an aggregate surface image that is an image of the surface of the piled aggregates;
a color feature extraction unit that extracts color feature amounts calculated from at least the gradation, hue, saturation, and brightness of color components of the aggregate as feature amounts of each of the captured aggregates from the aggregate surface image;
a particle size distribution estimation unit that estimates the particle size distribution of the aggregate by using a predetermined machine learning model when the shape feature amount and the color feature amount are input ,
The shape feature extraction unit
the sum, average and standard deviation of the feature amounts of the contour length, the contour area and the circularity of each of the aggregates in the aggregate surface image are calculated as the shape feature amounts;
The color feature extraction unit classifies each feature amount of the gradient, hue, saturation, and brightness of each color component of the aggregate in the aggregate surface image into a plurality of ranges of feature size, and determines the occurrence rate of each range as the color feature amount . 2. The aggregate particle size distribution estimation system according to claim 1, wherein the color components are each of a color component R, a color component G, and a color component B, and a gray scale calculated from each of the color components R, G, and B. 3. The aggregate particle size distribution estimation system according to claim 1, further comprising a machine learning model generation unit that trains the machine learning model using, as training data, each of the shape features and the color features extracted from the aggregate surface image obtained by capturing an image of the surface of a pile of aggregate whose particle size distribution is known. The machine learning model generation unit
The aggregate particle size distribution estimation system according to claim 3, characterized in that the training data is divided into training data and verification data at a predetermined ratio, the machine learning model is trained using the training data, and the accuracy rate of the estimation results of the machine learning model is calculated using the verification data, and this learning process is performed a predetermined number of times by replacing a portion of the training data with the verification data. The aggregate particle size distribution estimation system according to any one of claims 1 to 4, further comprising a feature importance evaluation unit that, in training the machine learning model, evaluates the contribution of each of the shape feature and the color feature to the training result of the machine learning model using a sorting importance technique.