Recycled raw material characteristic evaluation system and recycled raw material characteristic evaluation method

Abstract

To estimate, from an image, the characteristics of recycled raw material such as the content of artifacts, ignition loss, the amount of elution of chemical substances and the like.SOLUTION: Provided is a system and a method for estimating the characteristics of recycled raw material such as the content of artifacts, ignition loss, the amount of elution of chemical substances and the like, by using a learned model generated with reference material with known characteristics and inputting an image of the recycled raw material to the learned model. The recycled raw material characteristic evaluation system 1 comprises an image acquisition unit 11 that is image acquisition means, the learned model 13, and an estimation unit 14 that is estimation means. The learned model 13 is formed of a plurality of learned models different in the water content in raw material, which are learned models 13a, 13b, 13c when the water content is divided into three standards. Provided is a technique capable of dealing with raw material, the image of which changes according to a difference in the water content.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a system and method for easily evaluating characteristics such as the content of artificial matter, the amount of eluted chemical substances, and the loss on heating of recycled materials such as mixed construction waste from images. [Background technology]

[0002] The amount of construction waste generated in fiscal year 2018 was a massive 74.4 million tons. The recycling and reduction rate for mixed construction waste is currently only 63%, which is low compared to other items, and there is a need to improve the recycling rate. The average lifespan of a Japanese home is said to be around 30 years, and recycling construction waste, which will continue to increase in the future, will become even more important in order to realize a resource-circulating society.

[0003] Intermediate processing companies that recycle construction waste receive waste from a variety of building demolition sites, resulting in a wide variety of characteristics, such as the amount of artificial material content, thermal loss, and chemical leaching. This posed a challenge to quality control for recycled materials made from mixed construction waste. During waste acceptance inspections, assessors visually determined the proportions of wood and concrete. However, when it comes to recycled materials made by crushing and classifying this waste, the waste is sorted and divided into product groups according to its intended use during the manufacturing process, making it impossible to know the characteristics of products with specific particle sizes on-site.

[0004] Thermogravimetric analysis and statistical image analysis have been proposed as conventional methods for analyzing properties. Thermogravimetric analysis uses a specialized testing machine to measure the change in weight of a sample while heating it at a constant temperature increase rate. Statistical image analysis separates waste into particles, takes images of them, extracts geometric attributes such as the perimeter and diameter of each particle, and estimates the material through regression analysis.

[0005] When producing recycled materials from waste, it is also important to ensure environmental compatibility. For example, when recycled materials are used outdoors in civil engineering and construction, the Soil Contamination Countermeasures Act requires quality control of the materials to prevent the leaching of chemical substances for which environmental standards have been established, i.e., environmental pollutants.

[0006] Standards for pyrolysis loss have been established as an indicator of organic contaminants when landfilling the residue from the processing of mixed construction waste. Pyrolysis loss is expressed as the weight loss rate when a dried sample is ignited at 600°C for three hours. Materials with a pyrolysis loss of more than 5% have restrictions on their use in stable landfills and civil engineering materials, so it is important to manage them appropriately.

[0007] Patent Document 1 discloses a technology for sorting waste that uses gravity sensors and cameras to identify and sort objects using artificial intelligence. This technology targets relatively large waste that is larger than 50 mm and visible after particles of 50 mm or less have been removed using a sieve. Patent Document 2 discloses an image diagnostic system and learning method for managing chemical reaction rates and the mixing ratio of mixed samples. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] Japanese Patent Application Publication No. 2017-109197 [Patent Document 2] Japanese Patent Application Publication No. 2023-58429 Summary of the Invention [Problem to be solved by the invention]

[0009] Thermogravimetric analysis requires specialized testing equipment and takes a long time to complete, making it unsuitable for on-site measurements. Statistical image analysis cannot distinguish materials from images that are not separated into individual particles. The technology in Patent Document 1 does not mention the moisture content of the sample, and does not estimate the loss on heating or the amount of chemical elution. The technology in Patent Document 2 uses machine learning to estimate the blending ratio of multiple substances, but does not consider the moisture content of the sample, and does not estimate the loss on heating or the amount of chemical elution.

[0010] The amount of eluted environmental pollutants must usually be measured on the test solution obtained from the sample elution treatment using dedicated analytical equipment such as an ion chromatograph, spectrophotometer, high-frequency inductively coupled plasma atomic emission spectrometer, etc., which makes it difficult to measure on-site. In addition, the loss on heating must also be measured in the laboratory using equipment such as an electric furnace, which makes it difficult to measure on-site. [Means for solving the problem]

[0011] The present invention has been made in consideration of the above points, and provides a technology for quickly estimating characteristics of recycled materials such as mixed construction waste, such as the content of artificial matter, amount of eluted chemical substances, and thermal loss, from images acquired by an imaging device at the manufacturing site, etc. The technology is also applicable to materials whose captured images change depending on the moisture content.

[0012] That is, the first invention of the present invention is: An evaluation system for characteristics of recycled materials, comprising an image acquisition means, a trained model, and an estimation means, and capable of being mounted on at least a portable terminal device, The characteristic is the content of artificial substances contained in the raw material, the image acquisition means is an imaging means of the portable terminal device or a means for acquiring an image of the recycled material by selecting a saved image, The trained model is a trained model created by machine learning the relationship between a plurality of reference images obtained by capturing a reference object having known characteristics and moisture content and the characteristics, The estimation means is a means for estimating characteristics by inputting the image obtained by the image acquisition means into the trained model. Furthermore, an output means for outputting the estimation result can be preferably used.

[0013] A second invention is a system for evaluating characteristics of recycled materials, which includes an image acquisition means, a trained model, and an estimation means, and which can be installed in at least a portable terminal device, the characteristic is heat loss; the image acquisition means is an imaging means of the portable terminal device or a means for acquiring an image of the recycled material by selecting a saved image, The trained model is a trained model created by machine learning the relationship between a plurality of reference images obtained by capturing a reference object having known characteristics and moisture content and the characteristics, The estimation means is a means for estimating characteristics by inputting the image obtained by the image acquisition means into the trained model. Furthermore, an output means for outputting the estimation result can be preferably used.

[0014] A third invention is a system for evaluating characteristics of recycled materials, which includes an image acquisition means, a trained model, and an estimation means, and which can be installed in at least a portable terminal device, the characteristic is the amount of chemical substance elution, the image acquisition means is an imaging means of the portable terminal device or a means for acquiring an image of the recycled material by selecting a saved image, The trained model is a trained model created by machine learning the relationship between a plurality of reference images obtained by capturing a reference object having known characteristics and moisture content and the characteristics, the amount of eluted chemical substance is at least the amount of eluted fluorine or the amount of eluted hexavalent chromium, The estimation means is a means for estimating characteristics by inputting the image obtained by the image acquisition means into the trained model. Furthermore, an output means for outputting the estimation result can be preferably used.

[0015] The fourth invention is a system for evaluating the characteristics of recycled raw materials, characterized in that at least one of resolution optimization and illumination unevenness correction is performed as a preprocessing means on the image acquired by the image acquisition means.

[0016] A fifth invention is a method for evaluating characteristics of recycled materials, comprising an image acquisition step, a learning step, and an estimation step, The characteristic is the content of artificial substances contained in the raw material, the image acquisition step is a step of acquiring an image of the recycled material by capturing an image or selecting a stored image; The learning step is a step of creating a trained model by machine learning a relationship between a plurality of reference images obtained by capturing images of a reference object having known characteristics and moisture content and the characteristics, The method for evaluating the characteristics of recycled raw materials is characterized in that the estimation process is a process of inputting the image obtained in the image acquisition process into the trained model to estimate the characteristics.

[0017] A sixth aspect of the present invention is a method for evaluating characteristics of recycled materials, comprising an image acquisition step, a learning step, and an estimation step, the characteristic is heat loss; the image acquisition step is a step of acquiring an image of the recycled material by capturing an image or selecting a stored image; The learning step is a step of creating a trained model by machine learning a relationship between a plurality of reference images obtained by capturing images of a reference object having known characteristics and moisture content and the characteristics, The method for evaluating the characteristics of recycled raw materials is characterized in that the estimation process is a process of inputting the image obtained in the image acquisition process into the trained model to estimate the characteristics.

[0018] A seventh aspect of the present invention is a method for evaluating characteristics of recycled materials, comprising an image acquisition step, a learning step, and an estimation step, the characteristic is the amount of chemical substance elution, the image acquisition step is a step of acquiring an image of the recycled material by capturing an image or selecting a stored image; The learning step is a step of creating a trained model by machine learning a relationship between a plurality of reference images obtained by capturing images of a reference object having known characteristics and moisture content and the characteristics, the amount of eluted chemical substance is at least the amount of eluted fluorine or the amount of eluted hexavalent chromium, The method for evaluating the characteristics of recycled raw materials is characterized in that the estimation process is a process of inputting the image obtained in the image acquisition process into the trained model to estimate the characteristics.

[0019] The eighth invention is a method for evaluating the characteristics of recycled raw materials, characterized in that at least one preprocessing step of resolution optimization and illumination unevenness correction is performed on the image acquired in the image acquisition process. [Effects of the Invention]

[0020] According to the present invention, it is possible to quickly determine the amount of artificial matter contained in recycled materials, the amount of heat loss, and the amount of chemical elution from images of the recycled materials taken with a portable terminal device such as a smartphone. [Brief explanation of the drawings]

[0021] [Figure 1] Diagram explaining the configuration of the recycled material characteristics evaluation system [Figure 2] A diagram illustrating the configuration of a system for evaluating the characteristics of recycled materials, including image preprocessing. [Figure 3] Diagram explaining how to create a trained model [Figure 4] Diagram explaining how to evaluate the properties of recycled materials [Figure 5]FIG. 1 shows the estimated results of the gypsum content of recycled raw materials in Example 1. [Figure 6] FIG. 1 shows the estimated results of the wood content of recycled raw materials in Example 2. [Figure 7] FIG. 10 shows the estimated results of thermal loss of recycled raw materials in Example 3. [Figure 8] FIG. 10 shows the estimated results of the amount of fluorine eluted from recycled raw materials in Example 4. [Figure 9] FIG. 10 shows the estimated results of the gypsum content of recycled raw materials in Example 5. [Figure 10] FIG. 10 shows the estimated results of the wood content of recycled raw materials in Example 6. [Figure 11] FIG. 10 shows the estimated results of thermal loss of recycled raw materials in Example 7. [Figure 12] FIG. 13 is a diagram showing the estimation result of the image magnification ratio in Example 8. [Figure 13] FIG. 10 shows the estimated results of the amount of fluorine eluted from recycled raw materials in Example 8. [Figure 14] FIG. 10 shows the estimated results of the gypsum content of recycled raw materials in Example 9. [Figure 15] FIG. 10 is a diagram showing the estimated results of the amount of fluorine eluted from recycled raw materials in Example 10. [Figure 16] FIG. 16 shows the estimated results of the amount of hexavalent chromium eluted from recycled raw materials in Example 11. DETAILED DESCRIPTION OF THE INVENTION

[0022] The configuration of the recycled material characteristic evaluation system of the present invention will be explained using Figure 1. The recycled material characteristic evaluation system 1 comprises an image acquisition unit 11, which is an image acquisition means, a trained model 13, and an estimation unit 14, which is an estimation means. It is also preferable to have a data output unit 15 and a screen display unit 16. The trained model 13 comprises multiple trained models that differ depending on the moisture content of the raw material, and when the moisture content is classified into three levels, it comprises 13a, 13b, and 13c.

[0023] The configuration of characteristic evaluation system 2, which is a system for evaluating the characteristics of recycled materials including preprocessing means, will be described using Figure 2. Similar to characteristic evaluation system 1, it includes an image acquisition unit 11, a trained model 13, and an estimation unit 14, but an image preprocessing unit 12, which is an image preprocessing means, is connected between image acquisition unit 11 and trained model 13.

[0024] The image acquisition unit 11 can be a general imaging device, such as a smartphone or tablet terminal with a built-in camera, a portable digital camera, a digital camera with interchangeable lenses, a fixed digital camera, or a USB camera. Also, imaging devices capable of measuring depth, such as a digital camera with a LiDAR function or a stereo camera, can be preferably used. Cameras capable of emitting near-infrared light, infrared light, or ultraviolet light, which are wavelengths other than visible light, can also be used. The image acquisition unit 11 can also select any image from stored images.

[0025] The image obtained by the image acquisition unit 11 preferably has information for the red (R), green (G), and blue (B) channels, and may additionally have information for the depth channel, near-infrared channel, infrared channel, and ultraviolet channel.

[0026] The image preprocessing unit 12 has the function of processing images by performing at least one of image resolution optimization and illumination unevenness correction. For example, a portable terminal device such as a smartphone or tablet, or a personal computer can be used. Alternatively, a computer connected to a network, a server, or a service built on the cloud can also be used.

[0027] The trained model 13 is a machine learning model trained and created by combining image data and the characteristics of recycled materials using a supervised learning algorithm or the like. When an image of a recycled material with unknown characteristics is input into the trained model 13, the characteristics of the recycled material are output. Specific examples of supervised learning algorithms include, but are not limited to, multilayer perceptrons (MLPs), convolutional neural networks (CNNs), and Vision Transformers (ViTs). It is also preferable to perform additional learning (transfer learning) on ​​the trained model that has already been trained using other data. The trained model 13 may reside in a portable terminal device such as a smartphone or tablet, or in a personal computer, or it may be a service built on a network-connected computer, server, or cloud.

[0028] The trained model 13 uses different trained models depending on the moisture content of the recycled material. This is because the moisture content of recycled materials can vary depending on the weather, and this is reflected in the image. For example, recycled materials such as mixed construction waste are collected at demolition sites and brought in as is without drying, so their moisture content tends to be lower on sunny days and higher on rainy days. Because the overall brightness and gloss of images of recycled materials change with changes in moisture content, using a trained model built with images of low-moisture content materials for images of high-moisture content materials tends to reduce estimation accuracy. Therefore, it is preferable to use a trained model built with images of "low-moisture content" raw materials for low-moisture content materials, and a trained model built with images of "high-moisture content" raw materials for high-moisture content materials. This is an example in which moisture content is divided into two levels, low and high, but the same applies when the number of levels is three or more. Furthermore, the final characteristic estimate can be obtained by weighting and summing the estimates obtained from multiple trained models with different moisture contents. Recyclable materials collected at recycling sites are generally weighed when they are transported to recycling plants, and since the heavier the weight, the higher the transportation costs, care is taken to keep the materials as dry as possible when they are transported or collected.

[0029] The moisture content of recycled materials can be measured using known methods, such as gravimetric methods that calculate the difference in weight before and after drying, electrical conductivity methods, time-domain reflectometry (TDR) and time-domain transmittance (TDT) methods that utilize dielectric properties, capacitance methods, and near-infrared methods. It can also be estimated from images of the recycled materials. Moisture content estimation from images can be performed based on the average values ​​of the brightness, saturation, and hue of the entire image. Alternatively, it is preferable to construct a trained model that estimates moisture content from images and input the image into the trained model to estimate the moisture content.

[0030] The estimation unit 14 estimates the characteristics of recycled materials by inputting images of the recycled materials with unknown characteristics into the trained model 13. The estimation unit 14 may be present in a portable terminal device such as a smartphone or tablet terminal, a personal computer, or may be a computer connected to a network, a server, or a service built on the cloud.

[0031] The characteristics of the recycled raw materials obtained by the estimation unit 14 may be displayed on a screen display unit 16, such as the screen of a smartphone or personal computer, or may be recorded via the data output unit 15 in the built-in storage device of the smartphone or personal computer or in a cloud service for use.

[0032] The image resolution in the image preprocessing unit 12 will now be described. Image resolution is a concept expressed as the number of pixels per unit length of the object to be imaged. At the time of image capture, image resolution is primarily determined by the number of pixels in the image capture device, the focal length of the lens of the image capture device, and the distance between the object and the image capture device. However, even after image capture, image resolution can be adjusted by resizing and cropping the image. Estimation accuracy increases when the image resolution values ​​are close in the learning process and the estimation process. Using the same image capture device in the learning process and the estimation process results in equivalent image resolutions, which is effective in improving estimation accuracy, but it is also possible to use different image capture devices in the learning process and the estimation process. In this case, it is desirable to make the effective image resolutions equivalent.

[0033] The illumination unevenness correction performed by the image preprocessing unit 12 can be performed using known methods such as shading correction, flat field correction, rolling ball algorithm, and morphological calculation.

[0034] The trained model creation process in step S1 will be described with reference to FIG. 3. The trained model used in the present invention is created by machine learning using images of multiple training recycled materials (reference objects) as input information, with information on the characteristics of the recycled materials associated with the input information. Characteristics of recycled materials include the content of artificial matter, the amount of eluted chemical substances, and the weight loss from heating. First, in the image acquisition process of step S11, images of recycled materials with known characteristics are captured. In step S11, images may be acquired by selecting from stored images. Next, in the image preprocessing process of step S12, the images are preprocessed. This image preprocessing process may be omitted. Next, in the learning process of step S13, a trained model 13 is created by performing machine learning by associating images with characteristics using a supervised learning algorithm or the like.

[0035] A method for evaluating the characteristics of recycled material with unknown characteristics in step S2, which is performed after the trained model creation step in step S1, is described using Figure 4. First, in the image acquisition step of recycled material with unknown characteristics in step S21, an image of the recycled material with unknown characteristics is taken to acquire an image. In step S21, an image may be acquired by selecting from stored images. Next, in the image preprocessing step in step S22, the image is preprocessed. This image preprocessing step may be omitted. Next, in the estimation step in step S23, the image is input into the trained model 13, and the characteristics are estimated. (Type of recycled material)

[0036] The types of recycled materials targeted by this invention are preferably applicable to materials that are composed of or may contain artificial materials, such as mixed construction waste, construction materials, crushed home appliances and automobiles, disaster waste, food residue, waste plastics, scrap rubber, scrap metal, scrap glass, scrap concrete, scrap ceramics, scrap wood, and scrap textiles. Among these, this method is particularly suitable for mixed construction waste, construction materials, and disaster waste, which are often stored outdoors and have fluctuating moisture contents. While there are no particular restrictions on the particle size of the raw materials, this method is preferably applicable to relatively fine materials of 50 mm or less, and even more preferably to even finer materials of 5 mm or less. (Estimation of artificial content)

[0037] The types of artificial materials contained in recycled materials targeted by the present invention include, but are not limited to, industrially produced materials such as gypsum, wood, paper, plastic, glass, concrete, roofing tiles, metal, rubber, ceramics, synthetic fibers, and natural fibers. The present invention can estimate the content of any of these artificial materials as long as they can be distinguished from other materials in an image. The estimation results can be preferably displayed visually or saved as data using an output means. (Estimated heat loss)

[0038] The present invention also provides a system and method for estimating the loss due to thermal cracking from an image of recycled materials. Thermal cracking loss is used as an index for evaluating the organic matter content, and this value increases as the content of combustible materials such as wood and paper increases. In other words, thermal cracking loss is a characteristic that can be associated with organic matter such as wood or paper if it can be recognized in an image of recycled materials. Estimating the loss due to thermal cracking from an image is preferably done by directly associating the image with the loss due to thermal cracking when creating a trained model. Alternatively, the loss due to thermal cracking can be calculated by adding up the amount of organic matter, such as wood, paper, and plastic, using the estimated amount of artificial matter. The estimated results can be visually displayed or saved as data using an output device. (Estimation of chemical substance leaching amount)

[0039] The present invention also provides a system and method for estimating the amount of chemical leaching from images of recycled materials. This is because artificial materials in recycled materials contain specific chemicals. For example, gypsum, especially artificial gypsum, often contains fluorine. Therefore, a trained model that can recognize the presence of gypsum can estimate the amount of fluorine leaching. Additionally, concrete and cement often contain hexavalent chromium. Chemical substances targeted by the present invention include, for example, cadmium and its compounds, for which leaching amounts are subject to standard limits set by the Soil Contamination Countermeasures Act, hexavalent chromium compounds, simazine, cyanide compounds, thiobencarb, carbon tetrachloride, 1,2-dichloroethane, 1,1-dichloroethylene, cis-1,2-dichloroethylene, 1,3-dichloropropene, dichloromethane, mercury and its compounds, selenium and its compounds, tetrachloroethylene, thiuram, 1,1,1-trichloroethane, 1,1,2-trichloroethane, trichloroethylene, lead and its compounds, arsenic and its compounds, fluorine and its compounds, benzene, boron and its compounds, polychlorinated biphenyls, and organophosphorus compounds. In addition to these chemical substances, the present invention can also be used to measure the leaching amounts of other substances as long as they can be observed as changes in images. When estimating the amount of chemical substance leaching from an image, it is preferable to directly set the amount of chemical substance leaching as the objective variable when creating a trained model, or alternatively, the amount of chemical substance leaching may be calculated by using the estimated result of the artificial substance content and multiplying that value by a predetermined coefficient.These estimation results can preferably be made visible using output means or saved as data. [Example]

[0040] The present invention will be explained in more detail below by showing examples, but the present invention is not limited to the following examples. (Moisture content of recycled raw materials in examples)

[0041] In Examples 1 to 11 shown below, the moisture content of the recycled materials was classified into three levels: low, medium, and high. Low moisture content was defined as 0 to less than 10% by weight, medium moisture content was defined as 10 to less than 20% by weight, and high moisture content was defined as 20 to less than 50% by weight. Unless otherwise specified, the moisture contents of the recycled materials for study and estimation were the same (same category). (Imaging recycled materials and creating trained models in examples)

[0042] Unless otherwise noted, the same imaging conditions and imaging device were used to acquire the training images and estimation images, and the image resolution was made equivalent. A convolutional neural network was used to create the trained model. Images different from those used during training were used to estimate the characteristics after the trained model was created. Example 1

[0043] This Example 1 describes an example in which the artificial matter content of construction waste mixed soil (hereinafter referred to as mixed soil), a recycled raw material, is estimated using a smartphone with a built-in camera. The mixed soil was a sample that passed through a 5 mm mesh sieve. The smartphone's built-in camera had a horizontal resolution of 3024 pixels and a vertical resolution of 4032 pixels. The smartphone used was an iPhone (registered trademark) 14 Pro. Prior to creating the trained model, mixed soils with known gypsum contents, containing 2 or 5% gypsum by weight, were prepared. The moisture content of these mixed soils was between 4 and 8% by weight, and they were classified as low-moisture raw materials. Thirty samples of the mixed soil, each approximately 8 g in weight, were collected for each gypsum content and placed evenly across the entire surface of an 8 cm diameter Petri dish. Photographs were taken to obtain an image group P11. To create the trained model, data corresponding to the image group P11 and the gypsum content was used as input information to create a trained model M11. To confirm the performance of the trained model M11, a group of images P12 of mixed soil (low moisture content) with a known gypsum content was input as test data into the trained model M11 to estimate the gypsum content, and the gypsum content was displayed on the smartphone screen. Figure 5 shows the relationship between the known gypsum content and the estimated value. The estimated gypsum content had an error of between -2% and 43% of the known content, which was a good result. This Example 1 is an example in which the gypsum content, one of the characteristics of recycled raw materials, can be accurately estimated from a trained model created using images captured using a smartphone with a built-in camera. Example 2

[0044] Example 2 illustrates the estimation of wood content in recycled materials using the same smartphone as Example 1. Similar to Example 1, an image group P21 of mixed soil with a known wood content, containing 0.5 or 1.5% wood chips by weight, was obtained. The moisture content of this mixed soil was between 4 and 8% by weight, and it was classified as a low-moisture material. A trained model M21 was created to output the wood content using images obtained by performing Flat Field correction, a method of correcting for uneven lighting, on the image group P21 as image preprocessing. To verify the performance of the trained model M21, an image group P22 obtained from mixed soil with a known wood content (low moisture content) was subjected to Flat Field correction, as in Example 1, and input into the trained model M21 to estimate the wood content. The wood content was then displayed on the smartphone screen. Figure 6 shows the relationship between the known wood content and the estimated value. The estimated wood content was a satisfactory result, with an error of between -5% and 43% relative to the known content. This Example 2 is an example in which the wood content, which is one of the characteristics of recycled raw materials, can be accurately estimated from a trained model created using preprocessed images taken using a smartphone with a built-in camera. Example 3

[0045] This Example 3 describes an example of estimating the sintering loss of mixed soil using an interchangeable-lens digital camera. The digital camera had a horizontal resolution of 3,456 pixels and a vertical resolution of 2,304 pixels. Prior to creating a trained model, 60 8g samples of mixed soil were prepared, each with different organic matter content (e.g., wood and paper), i.e., different sintering losses. The moisture content of the mixed soil was between 4 and 8% by weight, and the mixed soil was classified as a low-moisture raw material. Each mixed soil was evenly distributed over the entire surface of an 8cm diameter petri dish and photographed to obtain image group P31. After photographing, the mixed soil was subjected to sintering loss measurement using an electric furnace. A trained model M31 was created using data corresponding to the image group P31 and the sintering loss measurement values ​​as input information. Next, an image group P32 of another mixed soil (low moisture content) was input to the trained model M31 to estimate the sintering loss. After the images were taken, the mixed soil was subjected to a measurement of the loss in thermal energy using an electric furnace. Figure 7 shows the relationship between the measured and estimated loss in thermal energy. The estimated loss in thermal energy was a good result, with an error of between -20% and 2% of the measured value. This Example 3 is an example in which the loss in thermal energy, one of the characteristics of recycled materials, can be accurately estimated from a trained model created using images taken with an interchangeable lens digital camera. Example 4

[0046] Example 4 describes an example in which the amount of chemical substances leached from mixed soil (fluorine leaching amount) is estimated using a stationary digital camera. To create a trained model, a stationary USB camera with a horizontal resolution of 2592 pixels and a vertical resolution of 1944 pixels was used to obtain an image group P41 in the same manner as in Example 1 for mixed soil with a known measured fluorine leaching amount of 0.1 mg / L or more but less than 0.8 mg / L. The moisture content of this mixed soil was 4% or more but less than 8% by weight, and it was classified as a low-moisture-content raw material. A trained model M41 was created using data corresponding to the image group P41 and the fluorine leaching amount as input information. To verify the performance of the trained model M41, an image group P42 of mixed soil (low moisture content) with a known fluorine leaching amount was input into the trained model M41 to estimate the fluorine leaching amount. Figure 8 shows the relationship between the measured and estimated fluorine leaching amounts. The fluorine leaching amount was estimated with a good error of -21% or more but less than 44% of the measured value. This Example 4 is an example in which the amount of fluorine elution, which is one of the characteristics of recycled raw materials, can be accurately estimated from a trained model created using images captured using a stationary USB camera. Example 5

[0047] This Example 5 describes an example in which the gypsum content of mixed soil is estimated using a smartphone with a built-in camera. Mixed soil containing 2 or 5 wt% gypsum was uniformly sprayed with water to adjust the moisture content to 25 to 30%. This was classified as a high-moisture raw material. To create a trained model, the same smartphone as in Example 1 was used to obtain an image group P51 of the mixed soil with a known gypsum content. Next, a trained model M51 was created using data corresponding to the image group P51 and the gypsum content as input information. To verify the performance of the trained model M51, an image group P52 of mixed soil (high moisture content) with a known gypsum content was input into the trained model M51 as test data. The gypsum content was estimated and displayed on the smartphone screen. Figure 9 shows the relationship between the known and estimated gypsum content. The estimated gypsum content was a good result, with an error of between -26% and 37% relative to the known content. This Example 5 is an example in which the gypsum content, which is one of the characteristics of recycled raw materials, can be estimated from a trained model created from images captured using a smartphone with a built-in camera, and is an example in which the gypsum content can be accurately estimated for raw materials with high moisture content. Example 6

[0048] Example 6 describes an example of estimating the wood content of mixed soil using a tablet device (portable terminal device) with a built-in camera. Mixed soil containing 0.5 or 1.5% wood by weight was uniformly sprayed with water to adjust the moisture content to 25% to 30%. This was classified as a high-moisture raw material. To create a trained model, a tablet device with a built-in camera measuring 3024 pixels horizontally and 4032 pixels vertically was used to capture images of the mixed soil in the same manner as in Example 1, obtaining an image group P61. Next, shading correction for uneven lighting was performed on the image group P61 as image preprocessing. Next, a trained model M61 was created using data corresponding to the image group P61 and the wood content as input information. To verify the performance of the trained model M61, an image group P62 of mixed soil (high moisture content) with a known wood content was input as test data into the trained model M61, and the wood content was estimated. The wood content was then displayed on the tablet device screen. The relationship between the known wood content and the estimated value is shown in Figure 10. The estimated wood content had an error of between -4% and 33% of the known content, which was a good result. Example 6 is an example in which the wood content, which is one of the characteristics of recycled raw materials, can be estimated from a trained model created from images captured using a tablet device with a built-in camera, and is an example in which the wood content can be accurately estimated even for raw materials with a high moisture content. Example 7

[0049] This Example 7 describes an example of estimating the calcination loss of mixed soil using the same tablet device as Example 6. Mixed soils with different organic matter contents (i.e., calcination losses) were prepared in the same manner as Example 3, and images were taken of the mixed soil in the same manner as Example 6 to obtain an image group P71. The moisture content of the mixed soil was between 4% and 8%, which indicated that it was a low-moisture material. After imaging, the mixed soil was subjected to calcination loss measurement using an electric furnace. Next, a trained model M71 was created using data corresponding to the image group P71 and the measured calcination loss values ​​as input information. To verify the performance of the trained model M71, an image group P72 of mixed soil with known calcination losses was input as test data into the trained model M71, and the calcination loss was estimated and displayed on the tablet device screen. The relationship between the measured and estimated calcination losses is shown in Figure 11. The estimated value of the loss due to thermal cracking was a good result, with an error of between -19% and 11% of the measured value. This Example 7 is an example in which the loss due to thermal cracking, which is one of the characteristics of recycled raw materials, can be accurately estimated from a trained model created using images captured using a tablet terminal with a built-in camera. Example 8

[0050] This Example 8 describes an example in which the amount of chemical substance elution (fluorine elution amount) from mixed soil is estimated using the same smartphone as in Example 1. An image group P81 was obtained by capturing images of mixed soil with a measured fluorine elution amount of 0.1 mg / L to 0.8 mg / L. The moisture content of this mixed soil was 4% to 8%, and this was classified as a raw material with a low moisture content. Next, a trained model M81 was created using data corresponding to the image group P81 and the fluorine elution amount as input information.

[0051] On the other hand, a trained model M82 for estimating image resolution was constructed using a different set of mixed soil images P83. This trained model was created by enlarging the image at magnifications of 25%, 50%, 75%, and 100%, cropping out an arbitrary area, resizing the image to a specified image size, and associating the image with the magnification. Figure 12 shows the results of estimating the magnification ratio when images of mixed soil different from the images used for training, with magnification ratios of 25, 50, 75, and 100%, were input into the trained model M82. The estimated magnification ratio was between -15% and 27% of the true value, demonstrating favorable results. Because the magnification ratio has a linear relationship with image resolution, the magnification ratio, and therefore the image resolution, can be obtained by inputting images obtained under imaging conditions with different focal lengths and subject distances into the trained model M82.

[0052] Next, image group P82 was obtained by capturing images of mixed soil with a fluorine elution rate of 0.1 mg / L to 0.8 mg / L at a focal length and subject distance different from those of image group P81. Image group P82 was input into trained model M82, and the magnification ratio was obtained as output information. If the magnification ratio is output as 100%, the input image can be said to have the same image resolution as image group P81 and can be input directly to trained model M81. If the magnification ratio is less than 100%, it can be determined that the input image captures a wider range of raw materials than image group P81, for example, because the focal length is short or the subject is farther away from the imaging device. By resizing image group P82 using this magnification ratio, the image resolution can be achieved to be equivalent to that of image group P81, thereby improving the estimation accuracy when input to trained model M81.

[0053] As described above, image group P82 was input into trained model M82 using the magnification ratio at which an estimate was obtained. Image group P82 was resized, i.e., image preprocessing was performed, to obtain image group P84, which was adjusted to the same image resolution as P81. To confirm the performance of trained model M81, image group P84 was input into trained model M81 as test data to estimate the amount of fluorine elution. The amount of fluorine elution was then displayed on a smartphone screen. Figure 13 shows the relationship between the measured and estimated amounts of fluorine elution. The estimated amount of fluorine elution was satisfactory, with an error of between -14% and 55% of the true value. Example 8 demonstrates how a smartphone with a built-in camera can be used to preprocess images using a trained model that estimates image resolution, followed by a trained model that estimates calcination loss. This demonstrates how the amount of fluorine elution, a characteristic of recycled materials, can be accurately estimated even from images captured under conditions different from those used during training. Example 9

[0054] Example 9 describes an example in which the same smartphone as in Example 1 is used to estimate the gypsum content of mixed soil with a low moisture content. Mixed soil containing 2 or 5% gypsum by weight was prepared. The moisture content of the mixed soil was 4% to 8%, which was classified as a low-moisture raw material. To create a trained model, the same smartphone as in Example 1 was used to obtain an image group P91 of the mixed soil with a known gypsum content. Next, the image group P91 was subjected to a Rolling Ball Algorithm correction for uneven lighting as image preprocessing. Next, a trained model M91 was created using data corresponding to the image group P91 and the gypsum content as input information. To verify the performance of the trained model M91, an image group P92 of mixed soil (low moisture content) with a known gypsum content was subjected to a Rolling Ball Algorithm correction for uneven lighting as test data and input into the trained model M91. The gypsum content was estimated and displayed on the smartphone screen. The relationship between the known gypsum content and the estimated value is shown in Figure 14. The estimated gypsum content was a good result, with an error of between -1% and 32% of the known content. Example 9 is an example in which a trained model is created from images captured using a smartphone with a built-in camera and preprocessed, and the gypsum content, which is one of the characteristics of recycled raw materials, can be estimated from images captured under the same imaging conditions, and the gypsum content can be well estimated by preprocessing the images. Example 10

[0055] This Example 10 describes an example in which a trained model for estimating the amount of fluorine elution from mixed soil with a low moisture content was created using a stationary camera, and the trained model was then used to estimate the amount of fluorine elution from images of the mixed soil captured with a smartphone equipped with a built-in camera. After the trained model M41 was created in the same manner as in Example 4, a group of images P102 of mixed soil (low moisture content) with a known amount of fluorine elution was acquired using the same smartphone as in Example 1. These images were then input into the trained model M41 to estimate the amount of fluorine elution. Figure 15 shows the relationship between the measured and estimated amounts of fluorine elution. The amount of fluorine elution could be estimated with an error of between -17% and 66% of the measured amount. This Example 10 illustrates how the amount of fluorine elution, one of the characteristics of recycled materials, can be accurately estimated from captured images, even if the imaging devices used during training and estimation are different, as long as the image resolution is equivalent. Example 11

[0056] Example 11 describes an example in which the same smartphone as in Example 1 was used to estimate the amount of hexavalent chromium elution (amount of chemical substance elution) from mixed soil. Mixed soil with a known hexavalent chromium elution level of 0.01 mg / L or more but less than 0.05 mg / L was uniformly sprayed with water to adjust the moisture content to 13% or more but less than 18%. This was classified as a medium-moisture content raw material. To create a trained model, the smartphone was used to obtain an image group P111 of the mixed soil. A trained model M111 was created using data corresponding to the image group P111 and the amount of hexavalent chromium elution as input information. To verify the performance of the trained model M111, an image group P112 of mixed soil (low moisture content) with a known amount of hexavalent chromium elution was input into the trained model M111 to estimate the amount of hexavalent chromium elution. Figure 16 shows the relationship between the measured and estimated amounts of hexavalent chromium elution. The amount of hexavalent chromium elution was estimated with a good error of -25% or more but less than 38% of the measured value. This Example 11 is an example in which the amount of hexavalent chromium elution, which is one of the characteristics of recycled raw materials, can be accurately estimated from a trained model created using an image captured using a smartphone.

[0057] Table 1 shows a summary of the imaging devices, estimation target characteristics, moisture content, and image preprocessing in Examples 1 to 11. [Table 1] [Industrial Applicability]

[0058] The ability to estimate the amount of artificial matter, heat loss, and eluted chemicals in recycled materials will ensure quality and expand the applications of recycled materials. Furthermore, based on the amount of eluted chemicals estimated by this invention, the amount of chemicals to be added to suppress elution of chemicals can be optimized, contributing to the prevention of environmental pollution and the reduction of chemical costs. [Explanation of symbols]

[0059] 1 ... Evaluation system for the properties of recycled materials 2 ... Evaluation system for characteristics of recycled materials including image preprocessing unit 11... Image acquisition section 12 ... Image preprocessing section 13 … Pre-trained model 13a, 13b, 13c ... Trained models created using raw materials with different moisture contents 14 … Estimation part 15...Data output section 16...Screen display section S1: Trained model creation process S2: Evaluation method for the characteristics of recycled materials S11... Image acquisition process S12: Image pre-processing S13: Learning process S21: Image acquisition process for recycled materials with unknown properties S22: Image pre-processing S23 … Estimation process

Claims

1. An evaluation system for characteristics of recycled materials, comprising an image acquisition means, a trained model, and an estimation means, and capable of being mounted on at least a portable terminal device, the characteristic is the content of artificial matter; the image acquisition means is an imaging means of the portable terminal device or a means for acquiring an image of the recycled material by selecting a saved image, The trained model is a trained model created by machine learning the relationship between a plurality of reference images obtained by capturing a reference object having known characteristics and moisture content and the characteristics, A system for evaluating the characteristics of recycled raw materials, characterized in that the estimation means is a means for inputting images obtained by the image acquisition means into the trained model to estimate the characteristics.

2. An evaluation system for characteristics of recycled materials, comprising an image acquisition means, a trained model, and an estimation means, and capable of being mounted on at least a portable terminal device, the characteristic is heat loss; the image acquisition means is an imaging means of the portable terminal device or a means for acquiring an image of the recycled material by selecting a saved image, The trained model is a trained model created by machine learning the relationship between a plurality of reference images obtained by capturing a reference object having known characteristics and moisture content and the characteristics, A system for evaluating the characteristics of recycled raw materials, characterized in that the estimation means is a means for inputting images obtained by the image acquisition means into the trained model to estimate the characteristics.

3. An evaluation system for characteristics of recycled materials, comprising an image acquisition means, a trained model, and an estimation means, and capable of being mounted on at least a portable terminal device, the characteristic is the amount of chemical substance elution, the image acquisition means is an imaging means of the portable terminal device or a means for acquiring an image of the recycled material by selecting a saved image, The trained model is a trained model created by machine learning the relationship between a plurality of reference images obtained by capturing a reference object having known characteristics and moisture content and the characteristics, the amount of eluted chemical substance is at least the amount of eluted fluorine or the amount of eluted hexavalent chromium, A system for evaluating the characteristics of recycled raw materials, characterized in that the estimation means is a means for inputting images obtained by the image acquisition means into the trained model to estimate the characteristics.

4. The system for evaluating the characteristics of recycled raw materials according to claim 1, 2 or 3, characterized in that at least one of resolution optimization and illumination unevenness correction is performed as a preprocessing means for the image acquired by the image acquisition means.

5. A method for evaluating characteristics of recycled materials, comprising an image acquisition step, a learning step, and an estimation step, the characteristic is the content of artificial matter; the image acquisition step is a step of acquiring an image of the recycled material by capturing an image or selecting a stored image; The learning step is a step of creating a trained model by machine learning a relationship between a plurality of reference images obtained by capturing images of a reference object having known characteristics and moisture content and the characteristics, A method for evaluating the characteristics of recycled raw materials, characterized in that the estimation process is a process of inputting the image obtained in the image acquisition process into the trained model to estimate the characteristics.

6. A method for evaluating characteristics of recycled materials, comprising an image acquisition step, a learning step, and an estimation step, the characteristic is heat loss; the image acquisition step is a step of acquiring an image of the recycled material by capturing an image or selecting a stored image; The learning step is a step of creating a trained model by machine learning a relationship between a plurality of reference images obtained by capturing images of a reference object having known characteristics and moisture content and the characteristics, A method for evaluating the characteristics of recycled raw materials, characterized in that the estimation process is a process of inputting the image obtained in the image acquisition process into the trained model to estimate the characteristics.

7. A method for evaluating characteristics of recycled materials, comprising an image acquisition step, a learning step, and an estimation step, the characteristic is the amount of chemical substance elution, the image acquisition step is a step of acquiring an image of the recycled material by capturing an image or selecting a stored image; The learning step is a step of creating a trained model by machine learning a relationship between a plurality of reference images obtained by capturing images of a reference object having known characteristics and moisture content and the characteristics, the amount of eluted chemical substance is at least the amount of eluted fluorine or the amount of eluted hexavalent chromium, A method for evaluating the characteristics of recycled raw materials, characterized in that the estimation process is a process of inputting the image obtained in the image acquisition process into the trained model to estimate the characteristics.

8. The method for evaluating the characteristics of recycled raw materials according to claim 5, 6 or 7, characterized in that at least one preprocessing step is performed on the image acquired in the image acquisition step, which preprocessing step is performed on the image acquired in the image acquisition step.

Images