Scrap sales support system, scrap sales support method, scrap sales support program, and manufacturing method of learning model

Abstract

To facilitate a selection of a buyer by assisting a seller's scrap sale.SOLUTION: A scrap sales support system includes: an acquisition unit (21) that acquires a captured image of a set of scraps to be sold; and an estimation unit (22) that estimates, from the captured image using one or more learning models constructed by machine learning, a determination result of each of a plurality of buyer candidates for the set, the determination result showing a ratio of scraps belonging to each class among a plurality of classes.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a scrap sales support system, a scrap sales support method, a scrap sales support program, and a method for manufacturing a learning model. [Background technology]

[0002] Amid growing global interest in reducing CO2 emissions, the steel industry is turning its attention to steel recycling, and domestic scrap circulation is expected to increase. Heavy scrap accounts for the majority of steel scrap, and is graded (HS, H1, H2, etc.) according to its thickness and other dimensions. For example, heavy scrap with a thickness of 6 mm or more, a width (or height) x length of 500 mm or less x 700 mm or less, and a unit weight of 600 kg or less is classified as the highest grade, HS. Heavy scrap with a thickness of 6 mm or more, a width (or height) x length of 500 mm or less x 1200 mm or less, and a unit weight of 1000 kg or less is classified as the next highest grade, H1. Heavy scrap is then similarly classified into H2, H3, and H4.

[0003] In distribution, the transaction price (yen / ton) varies depending on this grade. When buying and selling, workers at the buying company visually inspect the scrap brought in by the seller to determine the grade and the transaction price. This process (grading) requires skill, and variations between individuals are also an issue.

[0004] To address this issue, the following Patent Documents 1 to 5 disclose techniques for capturing an image of the appearance of scrap with a camera, inputting the image into a learning model, and automatically determining the grade. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 2020-95709 [Patent Document 2] Patent Publication No. 2021-157725 [Patent Document 3] International Publication No. WO2021-220987 [Patent Document 4] Patent No. 7036298 [Patent Document 5] Japanese Patent Application Publication No. 2023-70671 Summary of the Invention [Problem to be solved by the invention]

[0006] Typically, there are multiple potential buyers for a seller. In the past, sellers would select the company that was most likely to offer the highest price based on experience and intuition, taking into consideration the price lists for each grade presented by each buyer and their track record in grading.

[0007] One aspect of the present disclosure aims to facilitate the selection of buyer companies. [Means for solving the problem]

[0008] A scrap sales support system according to one embodiment of the present disclosure is a scrap sales support system that supports sellers in selling scrap, and includes an acquisition unit that acquires an image of a collection of scrap to be sold, and an estimation unit that uses one or more learning models constructed by machine learning to estimate, from the image, the judgment results of each of a plurality of potential buyers regarding the collection, including the judgment results of the proportion of scrap belonging to each of a plurality of grades.

[0009] A scrap sales support method according to one embodiment of the present disclosure includes an acquisition step of acquiring an image of a collection of scrap to be sold, and an estimation step of estimating, from the image, the judgment results of each of a plurality of potential buyers regarding the proportion of scrap belonging to each grade in the collection using one or more learning models constructed by machine learning.

[0010] A scrap sales support program according to one embodiment of the present disclosure is a scrap sales support program that supports sellers in selling scrap, and causes a computer to perform an acquisition process of acquiring an image of a collection of scrap to be sold, and an estimation process of using one or more learning models constructed by machine learning to estimate, from the image, the judgment results of each of a plurality of potential buyers regarding the proportion of scrap belonging to each grade in the collection.

[0011] A learning model manufacturing method according to one aspect of the present disclosure is a learning model manufacturing method for manufacturing a learning model for estimating the judgment results of the proportion of scrap belonging to each grade in a collection of scrap for sale from an image of the collection of scrap, and the learning model is manufactured using training data including image data representing an image of the collection of scrap taken by the seller and labels representing the proportion of scrap belonging to each grade in the collection of scrap judged by the buyer for the collection of scrap.

[0012] The scrap sales support system according to each aspect of the present disclosure may be realized by a computer, in which case the control program of the scrap sales support system that realizes the scrap sales support system on a computer by causing the computer to operate as each part (software element) of the scrap sales support system, and the computer-readable recording medium on which it is recorded, also fall within the scope of the present disclosure. [Effects of the Invention]

[0013] According to one aspect of the present disclosure, the selection of a buyer company is facilitated. [Brief explanation of the drawings]

[0014] [Figure 1] 1 is a block diagram showing a functional configuration of a scrap sales support system according to a first embodiment of the present disclosure. [Figure 2] FIG. 1 is a block diagram showing a configuration of a machine learning device according to a first embodiment of the present disclosure. [Figure 3]A flow chart showing the flow of a manufacturing method for a learning model according to the present disclosure. [Figure 4] 1 is a flowchart showing the flow of a scrap sales support method according to a first embodiment of the present disclosure. [Figure 5] 10 is an example of an image of a collection of captured scraps. [Figure 6] 10 is an example of an image showing a determination result displayed on a display unit. [Figure 7] 10 is another example of an image showing a determination result displayed on the display unit. [Figure 8] FIG. 10 is a block diagram showing the functional configuration of a scrap sales support system according to a second embodiment. [Figure 9] 10 is a flowchart showing the flow of a scrap sales support method according to a second embodiment. [Figure 10] 1 is an exemplary price chart showing scrap purchase prices by grade. [Figure 11] 10 is an example of an image showing a table displayed on a display unit. [Figure 12] 10 is an example of an image showing a table displayed on a display unit. [Figure 13] 10 is an example of an image showing a table displayed on a display unit. [Figure 14] 10 is an example of an image showing a table displayed on a display unit. [Figure 15] 10 is an example of an image showing a table displayed on a display unit. [Figure 16] FIG. 1 is a diagram illustrating an example of the installation of each component of a scrap sales support system. [Figure 17] FIG. 1 is a diagram illustrating an example of the installation of each component of a scrap sales support system. [Figure 18] FIG. 1 is a diagram illustrating an example of the installation of each component of a scrap sales support system. DETAILED DESCRIPTION OF THE INVENTION

[0015] [Embodiment 1] An embodiment of the present disclosure will be described in detail below.

[0016] (Configuration of scrap sales support system) In a scrap supply chain, from the upstream side, there may be dismantlers, scrap dealers, and steel companies. Therefore, there are multiple possible combinations of sellers and buyers according to the present disclosure. For example, if the seller is a dismantler and the buyer is a scrap dealer, the dismantler will desire to select the scrap dealer that will purchase the scrap at the highest price from among the multiple scrap dealers as the buyer. Also, if the seller is a scrap dealer and the buyer is a steel company, the scrap dealer will desire to select the steel company that will purchase the scrap at the highest price from among the multiple steel companies as the buyer. The scrap sales support system according to the present disclosure is a system that supports sellers in selling scrap in such a scrap supply chain.

[0017] The scrap sales support system of the present disclosure estimates, from an image of a collection of scrap to be sold, the judgment results of each of multiple buyers regarding the collection, including the proportion of scrap belonging to each of multiple grades.

[0018] For simplicity, the scrap sales support system will be referred to simply as a system in this specification. The configuration of a system 100 according to a first embodiment of the present disclosure will be described with reference to FIG. 1. FIG. 1 is a block diagram showing the functional configuration of the system 100. As shown in FIG. 1, the system 100 includes a camera 10, a processor 20, a storage unit 30, and a display unit 40. The camera 10, the processor 20, the storage unit 30, and the display unit 40 of the system 100 may be interconnected via a bus (not shown).

[0019] The camera 10 is an imaging device that a seller uses to capture an image of a collection of scrap metals to be sold.

[0020] The processor 20 is configured to execute the acquisition step, estimation step, and the like, described below, using a learning model stored in the memory unit 30. Specifically, as shown in FIG. 1, the processor 20 includes an acquisition unit 21 and an estimation unit 22. The acquisition unit 21 acquires captured images of a collection of scrap to be sold from the camera 10. The estimation unit 22 estimates, from the captured images, judgment results of each of multiple potential buyers regarding the collection of scrap, including the percentage of scrap belonging to each of multiple grades, using one or more learning models constructed through machine learning. The processor 20 may be, for example, a central processing unit (CPU), a graphic processing unit (GPU), a microprocessor, a digital signal processor, a microcontroller, an application-specific integrated circuit (ASIC) such as a tensor processing unit (TPU), or a combination thereof. In the example shown in FIG. 1, the acquisition unit 21, the estimation unit 22, and the integration unit 27 are included in a single processor 20. However, the estimation unit 22 and the integration unit 27 may be included in a processor separate from the processor that includes the acquisition unit 21. Moreover, the integration unit 27 may be omitted.

[0021] The storage unit 30 is configured to store a learning model constructed by machine learning. As the storage unit 30, for example, a random access memory, a flash memory, a hard disk drive, or the like can be used.

[0022] The display unit 40 is a display terminal such as a display, and displays an image indicating the judgment result estimated by the estimation unit 22 of the processor 20 on the display screen.

[0023] (Learning models and machine learning devices) In this disclosure, machine learning generally refers to a process of automatically constructing an estimation algorithm based on data. While deep learning is a typical example of machine learning, the machine learning in this disclosure is not limited to deep learning. For example, the machine learning in this disclosure may be supervised learning other than deep learning, unsupervised learning, or reinforcement learning.

[0024] In addition, in this disclosure, a learning model generally refers to an algorithm for inference constructed by machine learning. A typical trained model is a convolutional neural network (CNN), but the learning model in this disclosure is not limited to a CNN. The learning model in this disclosure may be, for example, a neural network other than a CNN, a logistic regression model, a support vector machine, a random forest, or a multiple regression model.

[0025] For example, the learning model according to the present disclosure can be generated (manufactured) by supervised learning using training data including image data representing an image of a collection of scraps taken by a seller and labels representing the proportion of scraps in the collection that belong to each grade, as determined by a buyer for the collection of scraps. Furthermore, the machine learning for constructing the learning model is supervised learning using training data including image data including a collection of scraps taken by a seller as a subject and training labels representing the judgment results of each of multiple potential buyers regarding the proportion of scraps in the collection that belong to each grade. The training labels may be judgment results based on past purchasing history. In other words, the training labels may represent the judgment results of each of multiple potential buyers with purchasing history regarding the proportion of scraps in the collection that belong to each grade.

[0026] The learning model and the machine learning device for constructing the learning model according to the present disclosure will be described in detail below.

[0027] 2 is a block diagram showing the configuration of a machine learning device 50. The machine learning device 50 is a device that implements a learning model creation method for creating a learning model according to the present disclosure. In other words, the machine learning device 50 is a device that implements a machine learning method for constructing a learning model according to the present disclosure.

[0028] As shown in FIG. 2, the machine learning device 50 includes a storage 51, a processor 52, and a memory 53. The storage 51, the processor 52, and the memory 53 are connected to one another via a bus (not shown). An input / output interface (not shown) and a communication interface (not shown) may also be connected to this bus. The input / output interface is used, for example, to input training data from an external device (e.g., a sensor and a keyboard) to the machine learning device 50. The communication interface is also used, for example, to provide a learning model to an external device (e.g., the above-mentioned system 100).

[0029] The storage 51 is configured to store a training dataset. The training dataset is a set of training data in which image data of a collection of scrap is labeled with a label indicating the scrap ratio of the collection of scrap. Here, the image data of the collection of scrap may be image data captured within the seller's company and including the collection of scrap before shipping as a subject. The determination of the scrap ratio of the collection of scrap may be the determination of the proportion of scrap belonging to each grade in the collection of scrap, determined by transporting the captured collection of scrap to the buyer and having the buyer's inspector visually inspect the collection of scrap.

[0030] In this case, the appearance of the collection of scraps may change depending on the state when photographed by the seller and the state when judged by the buyer as a result of transportation from the seller to the buyer. In other words, the appearance of the collection of scraps seen in the image data as learning data may be different from the appearance of the collection of scraps seen by the buyer's inspector for labeling to create the correct answer data.

[0031] In the past, in machine learning of learning models used to estimate the results of grading from images of scrap, labeling was performed based on the image data of the scrap itself. For example, the learning models disclosed in the above-mentioned Patent Documents 1 to 5 are constructed by machine learning in which an inspector of a buyer labels image data captured in the buyer's company. One feature of the present disclosure is that the labeling is not performed based on the image data used as learning data.

[0032] By using image data acquired by the seller and label data representing the judgment results made by the buyer as a learning dataset, a learning model can be obtained that estimates the buyer's judgment results from the image data acquired by the seller.

[0033] Attaching a label indicating the determination result to image data refers to associating the image data with the determination result in any manner, not limited to embedding the determination result in the image data (for example, as metadata). Methods for associating the image data with the determination result include, for example, creating a table indicating the correspondence between the image data and the determination result, or storing the image data in a directory corresponding to the determination result. Note that the storage 51 can be, for example, a flash memory, an HDD, an SSD, or a combination thereof.

[0034] The processor 52 is configured to load the learning dataset stored in the storage 51 onto the memory 53 and execute a construction process to construct a learning model through machine learning using this learning dataset. As described above, the learning model is an algorithm that estimates, from captured images of a collection of scrap, the judgment results of multiple potential buyers for the collection, including the judgment results of the proportion of scrap belonging to each of multiple grades. The processor 52 can be, for example, a CPU, a GPU, a microprocessor, a digital signal processor, a microcontroller, an ASIC such as a TPU, or a combination of these.

[0035] The memory 53 is configured to store the learning model obtained by executing the construction process by the processor 52. For example, a semiconductor RAM can be used as the memory 53. The learning model stored in the memory 53 may be stored (non-volatilely saved) in the storage 51 described above.

[0036] 3 is a flow diagram showing the flow of a method for manufacturing a learning model according to the present disclosure. The method shown in FIG. 3 can be regarded as a learning method for a learning model using a machine learning device 50, or as a manufacturing method (generation method) for a learning model using a machine learning device 50.

[0037] First, the processor 52 acquires training data including a plurality of pairs of image data of an image of a collection of scraps and a correct label representing the result of the judgment result of the scrap ratio in the collection of scraps (S31).

[0038] Next, the acquired training data is used to train a model (S32). In step S32, the model parameters may be corrected and updated so that the output when the image data of the training data is input to the training model approaches the correct label corresponding to each image data. Then, it is determined whether the model training has been completed (S33).

[0039] If the learning of the model has not been completed (S33, No), the process returns to step S31 and repeats the subsequent processes. If the learning of the model has been completed (S33, Yes), the learned model is output (S34).

[0040] Although the machine learning method including the construction process for constructing a learning model has been described above as being implemented by a single processor 52 installed in a single computer, the present invention is not limited to this. That is, it is also possible to adopt a configuration in which the machine learning method is executed cooperatively by multiple processors installed in a single computer or distributed across multiple computers.

[0041] Although the configuration in which the training dataset is stored in a single storage 51 provided in a single computer has been described here, the present invention is not limited to this. That is, it is also possible to adopt a configuration in which the training dataset is stored in a single computer or in multiple storages provided in multiple computers in a distributed manner. Furthermore, the dataset does not need to be stored in the storage 51 built into the computer together with the processor 52 and memory 53, but may be stored in a cloud server configured to be able to communicate with the computer via a network.

[0042] Although the configuration in which the learning model is stored in a single memory 53 provided in a single computer has been described here, the present invention is not limited to this. That is, it is also possible to adopt a configuration in which the learning model is stored in a single computer or in multiple memories provided in multiple computers in a distributed manner.

[0043] The program for causing the processor 52 to execute the machine learning method is recorded, for example, on a computer-readable, non-transitory, tangible recording medium. This recording medium may be the storage 51, the memory 53, or another recording medium. For example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit may be used as the other recording medium.

[0044] The machine learning device 50 can construct a learning model to be used by the system 100 described above.

[0045] The learning model according to the present disclosure may be, like learning model M shown in FIG. 1, a learning model that takes image data containing a collection of scrap for sale as a subject and outputs data including judgment results of the proportion of scrap belonging to each grade for each of a plurality of potential buyers in the collection. The learning dataset for constructing learning model M is a collection of training data in which image data of a collection of scrap is labeled with labels indicating the judgment results of each of a plurality of potential buyers regarding the scrap ratio in the collection of scrap. Learning model M can estimate and output a plurality of judgment results corresponding to each potential buyer in response to input image data.

[0046] Alternatively, the learning model according to the present disclosure may be a learning model corresponding to each potential buyer, as shown in FIG. 8. FIG. 8 will be described later. The memory unit 30 in FIG. 8 stores learning model A corresponding to buyer A, learning model B corresponding to buyer B, and learning model C corresponding to buyer C. In other words, the memory unit 30 stores multiple learning models corresponding to multiple buyers. For example, the learning model may be a learning model corresponding to buyer A, which receives image data containing a collection of scrap for sale as a subject and outputs data including a judgment result regarding the proportion of scrap belonging to each grade of buyer A in the collection. In a similar manner, the above-mentioned learning model M outputs data including a plurality of judgment results corresponding to multiple buyers A, B, and C from a single learning model in response to input image data.

[0047] The learning datasets for constructing learning models A, B, and C are sets of training data in which image data of a collection of scrap is labeled with labels representing the judgment results of the corresponding potential buyer regarding the scrap ratio in that collection of scrap. For example, learning model A can estimate and output the judgment results of buyer A in response to input image data.

[0048] Here, when a seller inputs image data captured in the seller's company into a learning model constructed using a set of training data labeled by the buyer for image data captured in the buyer's company as a learning dataset and estimates the buyer's judgment results, the optimal image format for each learning model may differ for each buyer. In this case, the image data input to the learning model may be data obtained by converting the captured image captured by the seller into input data for the learning model. Specifically, the image data input to the learning model may be image data that has been preprocessed, such as brightness adjustment, resolution change, distortion adjustment, and specific area extraction, for the captured image.

[0049] (Flow of scrap sales support method (1)) An exemplary flow of the scrap sales support method executed by the system 100 will be described with reference to Figures 4 to 7. Figure 4 is a flowchart showing the flow of the scrap sales support method according to the first embodiment.

[0050] In step S11, the acquisition unit 21 of the processor 20 acquires an image of a collection of scraps for sale captured by the camera 10 as image data (acquisition step). FIG. 5 is an example of an image of a collection of scraps captured by the camera 10 (image G). In this specification, the image of a collection of scraps for sale captured by the camera 10 is referred to as image G. In other words, image G is an image that includes the collection of scraps for sale as a subject.

[0051] The image of the collection of scraps by the camera 10 may be captured by the seller. The collection of scraps to be sold may be, for example, a collection of scraps stored in the back of a truck that has entered the seller's facility. Alternatively, it may be a collection of scraps stored in a storage container installed at the seller's facility. The entire collection of scraps may be included in the image G so that the subsequent estimation by the estimation unit 22 can be performed appropriately.

[0052] In the following step S12, the estimation unit 22 of the processor 20 uses the learning model M constructed by machine learning to estimate the judgment results of each of the multiple potential buyers from the image G acquired by the acquisition unit 21 in step S11 (estimation step).

[0053] In the following step S13, the processor 20 causes the display unit 40 to display the determination result estimated by the estimation unit 22.

[0054] Fig. 6 is an example of an image showing the judgment results displayed on the display unit 40. Reference numeral 601 in Fig. 6 is a table showing the judgment results estimated by the estimation unit 22 when buyer company A performs a grading judgment on the collection of scrap included in image G. Similarly, reference numerals 602 and 603 are tables showing the judgment results of buyer companies B and C, respectively, estimated by the estimation unit 22.

[0055] The judgment results of each potential buyer estimated by the estimation unit 22 may include information on the percentage of scrap belonging to each grade in the collection of scrap, as shown in Fig. 6. Note that in Fig. 6, the judgment results of each company are shown in a separate table for each buyer, but the present invention is not limited to this example, and the judgment results of each company may be shown together in a single table.

[0056] FIG. 7 is another example image showing the judgment results estimated by the estimation unit 22 and displayed on the display unit 40. The input to the learning model M may be multiple images G of a collection of scrap for sale photographed from various angles or positions. The output from the learning model M may be the judgment results, determined for each image, of the proportion of scrap belonging to each grade for each of multiple potential buyers. Reference numeral 701 in FIG. 7 is a table showing the judgment results of buyer company A estimated by the estimation unit 22. Similarly, reference numerals 702 and 703 are tables showing the judgment results of buyer companies B and C, respectively, estimated by the estimation unit 22. While FIG. 7 shows the judgment results for each image in a separate table for each buyer company, this example is not limiting.

[0057] (Another aspect of the processing by the estimation unit 22) (Another aspect 1) The estimation unit 22 may determine the proportion of scrap belonging to each grade among scraps present in each predetermined image region (e.g., pixel by pixel) obtained by dividing the image G acquired by the acquisition unit 21. In other words, the output from the learning model may be primary data representing the determination result of the proportion of scrap belonging to each grade within the range included in the divided image data obtained by dividing the captured image of the collection of scraps into small pieces. After determining the scrap ratio, the estimation unit 22 may further estimate the determination result of the proportion of scrap belonging to each grade in the collection by integrating multiple pieces of the primary data. Such integration processing will be described below. For simplicity, the "proportion of scrap belonging to each grade" may also be simply referred to as the "scrap ratio."

[0058] In this case, the estimation unit 22 may determine the scrap ratio of the scrap present in the divided image area to be determined by using information on the divided image area to be determined for the scrap ratio and information on the surrounding image areas surrounding the divided image area to be determined in the image G acquired by the acquisition unit 21.

[0059] The processor 20 may include an integration unit 27. The integration unit 27 tally up the scrap ratio determination results determined for each divided image area of ​​the image, and determine the scrap ratio for the entire image based on the tallying result. Specifically, when the scrap ratio determination results for each divided image area of ​​the image G are acquired, the integration unit 27 may tally up the scrap ratio determination results determined for each divided image area of ​​the image G, and determine the scrap ratio for the entire image G based on the tallying result.

[0060] For example, the integration unit 27 calculates an averaged certainty vector by arithmetically averaging the certainty vectors of the scrap rates determined for each divided image region of image G by the estimation unit 22 for all image regions included in image G. Then, the integration unit 27 may determine the averaged certainty vector as the scrap rate of image G. This makes it possible to derive a scrap rate determination result for the entire image based on the scrap rate determination results for each divided image region.

[0061] (Another aspect 2) Alternatively, the estimation unit 22 may determine the scrap ratio of the entire collection of scrap to be determined from a plurality of images of the collection of scrap to be determined. Specifically, when the acquisition unit 21 acquires a plurality of images of the collection of scrap to be sold, the estimation unit 22 may determine the scrap ratio for each image, and then integrate the results estimated from the plurality of captured images to estimate the determination result of the proportion of scrap belonging to each grade in the collection of scrap.

[0062] In this case, the acquisition unit 21 acquires multiple images G1 to GI captured of the collection of scraps to be determined, and the estimation unit 22 determines the scrap ratio for each image Gi (i = 1, 2, . . . , I) of the acquired multiple images G1 to GI. The integration unit 27 may then aggregate the determination results of the scrap ratio for each image Gi across the multiple images G1 to GI and determine the scrap ratio of the collection of scraps based on the aggregation results. For example, as described above, when the determination results for each image Gi are acquired as a confidence vector, the integration unit 27 may arithmetically average the confidence vectors of each image Gi for all images G1 to GI and determine the averaged confidence vector as the determination result of the scrap ratio of the collection of scraps to be determined.

[0063] In the above description of (Another Aspect 2), the estimation unit 22 determines the scrap ratio for each image and then integrates the results estimated from the multiple captured images. As another aspect, as disclosed in Patent Document 5, the estimation unit 22 may directly estimate the scrap ratio corresponding to the entire group of multiple images from multiple images of a collection of scrap to be determined. This corresponds to the estimation unit 22 including a function for integrating the results for the multiple captured images.

[0064] (Alternative Aspect 3) Furthermore, the above-described (Another embodiment 1) and (Another embodiment 2) may be combined. Specifically, in the system 100 shown in FIG. 1, the estimation unit 22 may input each image Gi of the captured multiple images G1 to GI into a learning model and acquire a confidence vector of the scrap ratio determination result for each divided image region Gj of each divided image Gi. Note that the learning model may be, for example, the above-described learning model M or learning models A, B, and C. The integration unit 27 then arithmetically averages the confidence vectors of each image region or pixel Gj of image Gi for all divided image regions, and acquires the averaged confidence vector as the scrap ratio determination result for image Gi. Furthermore, the integration unit 27 may perform this process for all images G1 to GI, arithmetically average the confidence vectors of each image Gi for all images G1 to GI, and determine the averaged confidence vector as the scrap ratio determination result for the set of scraps being determined.

[0065] For example, suppose that the estimation unit 22 acquires a certainty Pijk that the scrap captured in the divided image area Gj has the determined scrap ratio k as a result of determining the scrap ratio for an image Gi of the images G1 to GI capturing a collection of scrap. Also, the weight index determined for each grade class is αk, and the weight ratio of each grade class k in the collection of scrap is Rk.

number

[0066] The weight index is a correction coefficient that takes into account the fact that scrap of a higher grade, such as HS, has a larger weight even if they occupy the same image area, and can be set as a constant in advance. Of course, this correction can be omitted and an equal αk, such as αk = 1, can be used for all grade classes k in equation (1). This is equivalent to performing a simple arithmetic average process.

[0067] In addition, in the formula (1), the confidence vector of each image region Gj is used, but it is also possible to assign a confidence of 1.0 to the class with the highest confidence in each image region. This is done by changing Pijk in the formula (1) as follows:

number

[0068]

number

[0069] This configuration allows sellers to more accurately determine buyers who are advantageous for sales.

[0070] [Embodiment 2] Other embodiments of the present disclosure will be described below. For ease of explanation, the same reference numerals will be used to designate components having the same functions as those described in the above embodiment, and the description thereof will not be repeated.

[0071] FIG. 8 is a block diagram showing the functional configuration of a system 100A according to a second embodiment. The system 100A differs from the system 100 according to the first embodiment in that the processor 20 includes a calculation unit 23, a comparison / determination unit 24, and a table creation unit 25, and the memory unit 30 includes multiple learning models (learning models A, B, and C). In the second embodiment, an example is described in which the memory unit 30 stores learning model A, which has learned the judgment results of the inspector of buyer A; learning model B, which has learned the judgment results of the inspector of buyer B; and learning model C, which has learned the judgment results of the inspector of buyer C. The learning model stored in the memory unit 30 according to the second embodiment is the aforementioned learning model M, and learning model M may output results for three buyers: A, B, and C. Furthermore, the number of buyers is not limited to three, and may be any number greater than or equal to two.

[0072] The calculation unit 23 calculates the expected purchase price of the collection of scrap for each of the multiple potential buyers based on the scrap ratio determination results for each of the multiple potential buyers estimated by the estimation unit 22 and the purchase price (transaction unit price) of scrap determined for each grade by each of the multiple potential buyers. The purchase price of scrap may be the transaction unit price of scrap (yen / ton).

[0073] The comparison and determination unit 24 compares the expected purchase prices of the multiple potential buyers calculated by the calculation unit 23 and ranks them.

[0074] The table creation unit 25 creates a table showing the determination results based on the determination results of the comparison and determination unit. The table creation unit 25 may further include the estimation results of the estimation unit 22 and / or the calculation results of the calculation unit 23 in the table that it creates.

[0075] (Flow of scrap sales support method (2)) An exemplary flow of the scrap sales support method executed by the system 100A will be described with reference to Fig. 9 to Fig. 15. Fig. 9 is a flowchart showing the flow of the scrap sales support method according to the second embodiment.

[0076] In step S21, the acquisition unit 21 of the processor 20 acquires an image of the collection of scraps to be sold, captured by the camera 10, as image data.

[0077] In the following step S22, the estimation unit 22 of the processor 20 uses the learning model A·B·C constructed by machine learning to estimate the judgment result of the buyer agents A·B·C from the image G acquired by the acquisition unit 21 in step S11.

[0078] In the next step S23, the calculation unit 23 of the processor 20 calculates the expected purchase price of the collection of scrap for each of the multiple potential buyers based on the judgment result estimated by the estimation unit 22 in step S22 and the scrap trading unit price determined by each of the multiple potential buyers for each grade. FIG. 10 is an exemplary price list showing the scrap trading unit price for each grade. Such price lists for buyers A, B, and C referenced by the calculation unit 23 may be stored in the memory unit 30. Furthermore, the price lists may be updated periodically or at times provided by the buyers.

[0079] Specifically, the calculation unit 23 may calculate the expected purchase price of the collection of scrap as follows: That is, the expected purchase price of the collection of scrap is calculated by multiplying the total weight of the scrap by w (tons) and the proportion of grade i in the collection of scrap by r i , the unit price of grade i is u i (yen / ton), it can be calculated using the following formula (3). The total weight of the scrap can be calculated, for example, by subtracting the total weight of the truck before loading the scrap from the total weight of the truck after loading the scrap. The total weight of the truck can be measured using a weighing machine.

[0080]

number

[0081]

number

[0082]

number

[0083] In the following step S25, the display unit 40 displays the results based on the calculation results calculated by the calculation unit 23. FIGS. 11 to 15 are example images showing tables displayed on the display unit 40. As shown in FIG. 11, the display unit 40 may display information indicating potential buyers that the comparison / determination unit 24 has determined to have the highest expected purchase price for the collection of scrap. By displaying the information on the display unit 40 as the names of potential buyers estimated to have the highest expected purchase price, the seller can easily identify the most promising buyers.

[0084] 12, the display unit 40 may further display the expected purchase price of the potential buyer who has the highest expected purchase price for the collection of scrap based on the judgment result of the comparison / judgment unit 24 and the calculation result of the calculation unit 23. By displaying the expected purchase price instead of just the buyer's name, the expected purchase price of the most promising buyer can also be confirmed.

[0085] Furthermore, in step S25, the display unit 40 may display a list created by the table creation unit 25 based on the calculation results of the calculation unit 23 and the determination results of the comparison and determination unit 24. Specifically, as shown in FIG. 13, the display unit 40 may display a list including the expected purchase prices of the collection of scrap from each of a plurality of potential buyers based on the calculation results of the calculation unit 23. By displaying the expected purchase prices from each of a plurality of potential buyers in a list, the seller can compare the purchase prices with his or her own eyes. In this case, the processor 20 does not include the comparison and determination unit 24, and the above-mentioned step S24 may be omitted.

[0086] Furthermore, the display unit 40 may also display the list together with the determination result of the comparison and determination unit 24. For example, as shown in Fig. 13, a display indicating which buyer has the highest expected purchase price may be added to the list.

[0087] 14, the display unit 40 may display a list including the determination results of the proportion of scrap belonging to each grade and the expected purchase price for each of a plurality of potential buyers based on the estimation results of the estimation unit 22 and the calculation results of the calculation unit 23. By adding the determination results of the scrap proportion, information that serves as the basis for deriving the expected purchase price for each potential buyer can be obtained, improving interpretability.

[0088] The display unit 40 may also display a list including the transaction unit prices determined by each class for each potential buyer, as shown in Fig. 15. Note that, although the examples shown in Figs. 13 to 15 display the estimated results for all buyers A, B, and C, it is also possible to display only the result for the buyer with the highest expected purchase price.

[0089] (Installation example) An example of the installation of each component of the scrap sales support system described in the above embodiment will be described below.

[0090] 16, 17, and 18 are diagrams showing examples of the installation of each component of the scrap sales support system. The installation location and the target for managing each component of the scrap sales support system may be changed depending on the desired application.

[0091] For example, as shown in Figure 16, the camera 10 and the display unit 40 may be installed in a seller area managed by the seller, and the processor 20 and the storage unit 30 may be installed in a service provider area managed by a third-party service provider. In other words, the processor 20 and the storage unit 30 may be built into a computer installed in the service provider area. With this configuration, even if a learning model constructed by machine learning in which image data captured in the buyer's company is labeled by the buyer's inspectors is used, the buyer can realize a scrap sales support system without providing the learning model to the seller.

[0092] 17, the camera 10, processor 20, storage unit 30, and display unit 40 may all be installed in an area managed by the seller. In other words, the processor 20 and storage unit 30 may be built into a computer installed in the seller's area.

[0093] Alternatively, as shown in FIG. 18, a camera 10, a display unit 40, and a processor 20X may be installed in the seller's area, and processors 20A, 20B, and 20C and memories 30A, 20B, and 30C for storing the learning models of each buyer may be installed in the buyer's area for each of buyers A, B, and C, respectively. Specifically, the processor 20X may be built into a computer installed in the seller's area. The processor 20A and the memory 30A for storing the learning model A may be built into a computer installed in the buyer's area. The processor 20B and the memory 30B for storing the learning model B may be built into a computer installed in the buyer's area. The processor 20C and the memory 30C for storing the learning model C may be built into a computer installed in the buyer's area.

[0094] The processor 20X installed in the seller area may have an image transmission unit 26, a comparison / determination unit 24, and a table creation unit 25. The image transmission unit 26 transmits images captured by the camera 10. The processors 20A, 20B, and 20C installed in the buyer areas may each have an acquisition unit 21, an estimation unit 22, and a calculation unit 23.

[0095] With this configuration, each buyer can manage and own the learning model that has learned its own judgment results.

[0096] [Software implementation example] The functions of the scrap sales support system (hereinafter referred to as the "system") can be realized by a program that causes a computer to function as the system, and a program that causes a computer to function as each control block of the system (particularly each part included in processor 20).

[0097] In this case, the system includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0098] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0099] In addition, some or all of the functions of each of the control blocks can be realized by logic circuits. For example, integrated circuits in which logic circuits that function as each of the control blocks are formed are also included in the scope of the present disclosure. In addition, the functions of each of the control blocks can also be realized by, for example, a quantum computer.

[0100] The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present disclosure.

[0101] 〔summary〕 The scrap sales support system according to aspect 1 of the present disclosure is a scrap sales support system that supports sellers in selling scrap, and includes an acquisition unit that acquires an image of a collection of scrap to be sold, and an estimation unit that uses one or more learning models constructed by machine learning to estimate, from the image, the judgment results of each of a plurality of potential buyers regarding the collection, including the proportion of scrap belonging to each of a plurality of grades.

[0102] The scrap sales support system according to a second aspect of the present disclosure is the system of the first aspect, further comprising a display unit that displays the determination result estimated by the estimation unit.

[0103] The scrap sales support system according to aspect 3 of the present disclosure further includes a calculation unit that calculates the expected purchase price of the set for each of the plurality of potential buyers based on the judgment result estimated by the estimation unit in aspect 1 or 2 above and the purchase price of scrap set by each of the plurality of potential buyers for each grade.

[0104] In the scrap sales support system according to aspect 4 of the present disclosure, in the above-mentioned aspect 3, the calculation unit further calculates an expected purchase price of the collection based on data regarding the weight of foreign matter and the like contained in the collection.

[0105] The scrap sales support system of aspect 5 of the present disclosure, in aspect 3 above, further includes a display unit that displays information indicating the buyer with the highest expected purchase price for the collection based on the calculation results calculated by the calculation unit.

[0106] A scrap sales support system according to a sixth aspect of the present disclosure is the above-mentioned fifth aspect, wherein the display unit further displays the expected purchase price of the potential buyer with the highest expected purchase price for the collection.

[0107] The scrap sales support system of aspect 7 of the present disclosure, in aspect 3 above, further includes a display unit that displays a list of the expected purchase prices of the set for each of the multiple potential buyers based on the calculation results calculated by the calculation unit.

[0108] A scrap sales support system according to aspect 8 of the present disclosure is one in which, in any of aspects 1 to 7 above, the input to the learning model is image data containing the set as a subject, and the output from the learning model is data including a judgment result of the proportion of scrap belonging to each grade for each of a plurality of potential buyers for the set.

[0109] A scrap sales support system according to aspect 9 of the present disclosure is one in which, in any of aspects 1 to 8 above, the estimation unit uses a plurality of the learning models corresponding to a plurality of potential buyers, the input to each of the learning models is image data representing an image including the set as a subject, and the output from the learning model is data representing the judgment result of the corresponding potential buyer regarding the proportion of scrap belonging to each grade in the set.

[0110] A scrap sales support system according to a tenth aspect of the present disclosure is, in the eighth or ninth aspect, wherein the image data is data obtained by converting the captured image into input data for the learning model.

[0111] A scrap sales support system according to aspect 11 of the present disclosure is one in which, in any of aspects 1 to 10 above, the output from the learning model is primary data representing the judgment result of the proportion of scrap belonging to each grade within the range contained in the divided image data obtained by dividing the captured image in the set into small pieces, and the estimation unit estimates the judgment result of the proportion of scrap belonging to each grade in the set by integrating multiple pieces of primary data.

[0112] A scrap sales support system according to aspect 12 of the present disclosure is one in which, in any of aspects 1 to 11 above, the acquisition unit acquires multiple images of the collection of scrap to be sold, and the estimation unit estimates the judgment result of the proportion of scrap belonging to each grade in the collection by integrating the results estimated from each of the multiple images.

[0113] A scrap sales support system according to aspect 13 of the present disclosure is one in which, in any of aspects 1 to 12 above, the machine learning is supervised learning using training data including image data containing a collection of scrap photographed by a seller as a subject, and labels representing the judgment results of each of multiple buyers regarding the proportion of scrap belonging to each grade in the collection.

[0114] A scrap sales support method according to aspect 14 of the present disclosure includes an acquisition step of acquiring an image of a collection of scrap to be sold, and an estimation step of estimating, from the image, the judgment results of each of a plurality of potential buyers regarding the proportion of scrap belonging to each grade in the collection, using one or more learning models constructed by machine learning.

[0115] A scrap sales support program according to aspect 15 of the present disclosure is a scrap sales support program that supports sellers in selling scrap, and causes a computer to perform an acquisition process of acquiring an image of a collection of scrap to be sold, and an estimation process of estimating, from the image, the judgment results of each of a plurality of potential buyers regarding the proportion of scrap belonging to each grade in the collection, using one or more learning models constructed by machine learning.

[0116] The method for manufacturing a learning model according to aspect 16 of the present disclosure is a method for manufacturing a learning model for estimating the judgment result of the proportion of scrap belonging to each grade in a collection of scrap for sale from an image of the collection of scrap, and is a manufacturing method for manufacturing a learning model using training data including image data representing an image of the collection of scrap taken by the seller and labels representing the proportion of scrap belonging to each grade in the collection of scrap judged by the buyer for the collection of scrap. [Explanation of symbols]

[0117] 100, 100A··· System (Scrap Sales Support System) 10. Camera 20, 20A·B·C, 20X··· Processor 21... Acquisition section 22...Estimation part 23 Calculation section 24...Comparison / judgment section 25. Table Creation Section 26 Image transmission unit 27. Integration Department 30, 30A·B·C···Storage section 40...Display section 50 Machine Learning Device 51 Storage 52 processors 53 Memory

Claims

1. A scrap sales support system that supports a seller in selling scrap, comprising: an acquisition unit that acquires a captured image of a collection of scraps to be sold; A scrap sales support system comprising an estimation unit that uses one or more learning models constructed by machine learning to estimate, from the captured image, the judgment results of each of a plurality of potential buyers regarding the set, which are the judgment results regarding the proportion of scrap belonging to each of a plurality of grades.

2. The scrap sales support system according to claim 1 , further comprising a display unit that displays the judgment result estimated by the estimation unit.

3. The scrap sales support system of claim 1 further comprises a calculation unit that calculates the expected purchase price of the set for each of the plurality of potential buyers based on the judgment result estimated by the estimation unit and the purchase price of scrap determined by each of the plurality of potential buyers for each grade.

4. The scrap sales support system according to claim 3 , wherein the calculation unit further calculates an expected purchase price of the collection based on data relating to the weight of foreign matter and the like contained in the collection.

5. 4. The scrap sales support system according to claim 3, further comprising a display unit that displays information indicating the buyer who has the highest expected purchase price for the collection based on the calculation result calculated by said calculation unit.

6. 6. The scrap sales support system according to claim 5, wherein said display unit further displays the expected purchase price of the potential buyer having the highest expected purchase price for said collection.

7. 4. The scrap sales support system according to claim 3, further comprising a display unit that displays a list of expected purchase prices of the set for each of the plurality of potential buyers based on the calculation results calculated by the calculation unit.

8. The scrap sales support system of claim 1, wherein the input to the learning model is image data including the set as a subject, and the output from the learning model is data including a judgment result of the proportion of scrap belonging to each grade for each of a plurality of potential buyers for the set.

9. the estimation unit uses a plurality of the learning models corresponding to a plurality of potential buyers, The scrap sales support system of claim 1, wherein the input to each of the learning models is image data representing an image including the set as a subject, and the output from the learning model is data representing the corresponding potential buyer's judgment result regarding the proportion of scrap belonging to each grade in the set.

10. 10. The scrap sales support system according to claim 8, wherein the image data is data obtained by converting the captured image into input data for the learning model.

11. The output from the learning model is primary data representing a determination result of the proportion of scrap belonging to each grade within a range included in divided image data obtained by dividing the captured image in the set into small pieces, The scrap sales support system according to claim 1 , wherein the estimation unit estimates the determination result of the proportion of scrap belonging to each grade in the set by integrating a plurality of the primary data.

12. The acquisition unit acquires a plurality of captured images of the collection of scraps to be sold, The scrap sales support system according to claim 1 , wherein the estimation unit estimates a judgment result of the proportion of scrap belonging to each grade in the collection by integrating the results estimated from each of the plurality of captured images.

13. The scrap sales support system of claim 1, wherein the machine learning is supervised learning using training data including image data of a collection of scrap photographed by a seller as a subject, and learning labels representing the judgment results of each of a plurality of buyers regarding the proportion of scrap belonging to each grade in the collection.

14. An acquisition step of acquiring a captured image of a collection of scraps to be sold; and an estimation step of estimating, from the captured images, the judgment results of each of a plurality of potential buyers regarding the proportion of scrap belonging to each grade in the set, using one or more learning models constructed by machine learning.

15. A scrap sales support program that supports sellers in selling scrap, comprising: On the computer, An acquisition process for acquiring an image of a collection of scraps to be sold; A scrap sales support program for executing an estimation process that uses one or more learning models constructed by machine learning to estimate, from the captured images, the judgment results of each of multiple potential buyers regarding the proportion of scrap belonging to each grade in the set.

16. A manufacturing method for manufacturing a learning model for estimating a judgment result of a proportion of scrap belonging to each grade in a collection of scrap to be sold from a captured image of the collection of scrap, comprising: A method for manufacturing a learning model, which uses training data including image data representing an image of a collection of scrap taken by a seller and labels representing the proportion of scrap belonging to each grade in the collection of scrap as determined by a buyer for the collection of scrap.

Images