Scrap inspection device, scrap inspection method and program

Abstract

To provide a foreign matter detection technique capable of detecting a foreign matter in a scrap by using a machine learning model even when there are dust and coating, etc. in the foreign matter to be detected.SOLUTION: One embodiment of the present disclosure relates to a scrap inspection device comprising: an acquisition unit which acquires an image obtained by imaging a scrap; a detection unit which acquires a detection result showing a ratio of an existence area of the foreign matter to an existence area of the scrap from the image by using a machine learning model; and a selection unit which determines a content ratio index of the foreign matter in the scrap on the basis of the detection result and selects the scrap according to the content ratio index.SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] The present disclosure relates to a scrap inspection device, a scrap inspection method, and a program. [Background technology]

[0002] As global interest in reducing CO2 emissions grows, the steel industry is turning its attention to steel recycling, and an increase in domestic scrap distribution is expected. To use scrap in high-quality steel, it is necessary to reduce the amount of tramp elements (such as Cu) that induce cracks during hot processing.

[0003] As a method for removing non-ferrous metals such as Cu from scrap iron products, there has been a conventional shredding system that separates iron from other impurities by shredding scrap iron products such as discarded automobiles. For example, Non-Patent Document 1 discloses a shredding system that includes a crusher that crushes iron products by using the force generated by the rotation of hammers attached to the circumference of a cylindrical roller, a wind-powered separator that removes light impurities, and a magnetic separator that separates non-ferrous metals by magnetic force.

[0004] However, some of the iron scrap pieces shredded by such shredders contain entangled iron and copper (Cu). This causes the iron to be attracted to the drum due to its magnetic properties, resulting in the copper being mixed into the scrap and becoming a raw material for steelmaking. For this reason, a technology is needed to remove copper-rich iron scrap pieces. Patent Document 1 discloses a technology that captures a color image of a group of iron scraps, individually recognizes each scrap piece in the color image based on the image's brightness (I) signal, determines whether each pixel contains copper based on the image's hue angle (H) signal and saturation (S) signal, and calculates the copper area in each scrap, i.e., the number of pixels identified as copper. The technology then detects copper by identifying whether each scrap contains copper based on the ratio R = SCu / St of the copper area of ​​each scrap piece to the total area St of each scrap piece, and removes the detected copper.

[0005] Furthermore, Patent Document 2 discloses that copper detected from scrap being transported on a transport device is removed by, for example, a robot or air. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Publication No. 07-280747 [Patent Document 2] Japanese Patent Publication No. 2022-91278 [Non-patent literature]

[0007] [Non-Patent Document 1] “Ferrous and non-ferrous scrap recycling” [Retrieved December 27, 2023] (http: / / www.fuji-metal.com / scrap) Summary of the Invention [Problem to be solved by the invention]

[0008] However, many copper wires are covered with resin, and the resin colors vary, including white, black, red, blue, and green, so the detection technology described in Patent Document 1, which assumes that the metal is exposed and uses hue angle (H) and saturation (S) signals to detect copper wires, may not be able to detect them adequately. Furthermore, the color of the copper wire itself can become dull not only due to the resin but also due to dust and dirt mixed in with the scrap, which can reduce detection accuracy.

[0009] In view of the above problems, an object of the present disclosure is to provide a foreign object detection technology that can detect foreign objects in scrap using a machine learning model, even if the foreign object to be detected contains dust, coating, or the like. [Means for solving the problem]

[0010] One aspect of the present disclosure relates to a scrap inspection device having an acquisition unit that acquires an image of scrap, a detection unit that uses a machine learning model to acquire detection results for foreign matter in the scrap from the image, and a sorting unit that determines a content ratio index that indicates the ratio of the area where the foreign matter exists to the area where the scrap exists based on the detection results, and sorts the scrap according to the content ratio index. [Effects of the Invention]

[0011] According to the present disclosure, it is possible to provide a foreign object detection technology that can detect foreign objects in scrap using a machine learning model, even if the foreign object to be detected contains dust, coating, or the like. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a schematic diagram illustrating a foreign object detection model according to one embodiment of the present disclosure. [Figure 2] FIG. 2 is a block diagram illustrating a hardware configuration of a scrap inspection device according to an embodiment of the present disclosure. [Figure 3] FIG. 3 is a block diagram illustrating a functional configuration of a scrap inspection device according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a schematic diagram illustrating foreign object detection using a peripheral image region according to one embodiment of the present disclosure. [Figure 5A] FIG. 5A is a schematic diagram illustrating foreign object detection using attention according to one embodiment of the present disclosure. [Figure 5B] FIG. 5B is a schematic diagram illustrating foreign object detection using attention according to one embodiment of the present disclosure. [Figure 5C] FIG. 5C is a schematic diagram illustrating detection of areas where scrap is present according to one embodiment of the present disclosure. [Figure 6] FIG. 6 is a schematic diagram illustrating a scrap sorting process according to one embodiment of the present disclosure. [Figure 7] FIG. 7 is a flowchart illustrating a foreign object detection process according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0014] In the following embodiment, a scrap inspection device for detecting foreign matter such as coated copper wires in scrap is disclosed.

[0015] [Summary of the Disclosure] In foreign matter detection according to an embodiment of the present disclosure, the scrap inspection device 100 uses a machine learning model that receives as input scrap images captured from scrap and outputs detection results indicating the presence or absence of foreign matter in the scrap. Such a machine learning model can be realized as a trained foreign matter detection model that is trained using a training dataset that includes training images of scrap and annotations that indicate areas in the training images where foreign matter is present.

[0016] For example, as shown in FIG. 1, the trained foreign object detection model 20 receives a scrap image as input and outputs a detection result indicating an image area, such as a pixel, where a foreign object exists. Accurate detection of dust, coated copper wire, and the like requires detecting the area where a foreign object exists based on the shape of the copper wire within the scrap, rather than relying primarily on the color of the copper wire. As a machine learning model that does not rely solely on the color of the foreign object, the foreign object detection model 20 can be realized, for example, by a machine learning model that detects the area where a foreign object exists by taking into account not only the image area of ​​interest but also the surrounding area of ​​the image area of ​​interest, or a machine learning model that utilizes a self-attention mechanism.

[0017] When the detection result is obtained from the foreign matter detection model 20, the scrap inspection device 100 determines a content ratio index that indicates, for example, the ratio (e.g., area ratio, volume ratio, etc.) of the area where foreign matter exists to the area where scrap exists, based on the detection result, and removes scrap whose content ratio index is equal to or greater than a predetermined threshold. In other words, if the detected foreign matter could become a tramp element that induces cracks during hot working, the scrap containing the foreign matter can be removed.

[0018] Here, the scrap inspection device 100 may be realized by a computing device such as a server or a personal computer (PC), and may have a hardware configuration such as that shown in Fig. 2. That is, the scrap inspection device 100 has a storage device 101, a processor 102, an interface device 103, and a communication device 104, which are interconnected via a bus B.

[0019] The programs or instructions that realize the various functions and processes described below in the scrap inspection device 100 may be downloaded from any external device via a network, or may be provided from a removable storage medium such as a CD-ROM (Compact Disk-Read Only Memory) or flash memory.

[0020] Storage device 101 may be implemented by random access memory, flash memory, a hard disk drive, or the like, and stores installed programs or instructions as well as files, data, etc. used in executing the programs or instructions. Storage device 101 may also include a non-transitory storage medium.

[0021] The processor 102 may be realized by one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), processing circuitry, etc., which may be composed of one or more processor cores, and performs various functions and processes of the scrap inspection device 100 described below in accordance with programs, instructions, data such as parameters required to execute the programs or instructions, etc. stored in the memory device 101.

[0022] The interface device 103 realizes an interface with a user of the scrap inspection device 100. For example, the user operates a GUI (Graphical User Interface) displayed on a display or touch panel using a keyboard, mouse, etc., to send and receive various information, data, instructions, etc. to and from the scrap inspection device 100 via the interface device 103.

[0023] The communication device 104 is realized by various communication circuits that execute communication processes with external devices, the Internet, a communication network such as a LAN (Local Area Network), and the like.

[0024] However, the above-described hardware configuration is merely an example, and the scrap inspection device 100 according to the present disclosure may be realized by any other appropriate hardware configuration.

[0025] [Scrap inspection device] Next, a description will be given of the scrap inspection device 100 according to an embodiment of the present disclosure. Fig. 3 is a block diagram showing the functional configuration of the scrap inspection device 100 according to an embodiment of the present disclosure.

[0026] 3, the scrap inspection device 100 includes an acquisition unit 110, a detection unit 120, and a sorting unit 130. Each functional unit of the acquisition unit 110, the detection unit 120, and the sorting unit 130 may be realized by the processor 102 executing a computer program stored in the storage device 101 of the scrap inspection device 100.

[0027] The acquisition unit 110 acquires images of scrap. For example, when scrap is transported by a transport device, the acquisition unit 110 acquires scrap images of the scrap captured by a camera installed above the transport device. For example, the camera may capture images of the entire width of the transport device, perpendicular to the transport direction, so that the position of the scrap captured in the scrap image in the width direction can be identified. In addition, it is desirable to set the capture cycle according to the camera's field of view and transport speed so that all scrap transported on the transport device can be captured. The acquisition unit 110 transmits the acquired scrap images to the detection unit 120.

[0028] The detection unit 120 uses a machine learning model to obtain a detection result of foreign matter in the scrap from the image. Specifically, the detection unit 120 inputs the scrap image to the foreign matter detection model 20 and obtains a detection result indicating whether the captured scrap contains a foreign matter from the foreign matter detection model 20.

[0029] For example, the foreign object may be a coated copper wire, a copper wire discolored by dust, or the like, but the present disclosure is not necessarily limited thereto and may be other metals that may be mixed into scrap. The foreign object to be detected may vary depending on the training data used to train the foreign object detection model 20. For example, if the training data is composed of scrap images and scrap images with annotations indicating the presence area of ​​the detection object contained in the captured scrap, the object will be the detection target of the foreign object detection model 20.

[0030] For example, a foreign object detection model 20 for detecting coated copper wires or discolored copper wires is trained using a training dataset consisting of scrap images containing coated copper wires or discolored copper wires and scrap images with annotations indicating the areas where the copper wires are present. Specifically, when the foreign object detection model 20 to be trained is realized as a deep learning model, scrap images are input to the foreign object detection model 20 to be trained, and parameters of the foreign object detection model 20 to be trained are adjusted based on the error between the processing results from the foreign object detection model 20 to be trained and the annotated scrap images. When a predetermined termination condition is met (e.g., the completion of training processing using a predetermined number of training data), the finally acquired foreign object detection model 20 to be trained can be acquired as the trained foreign object detection model 20.

[0031] In one embodiment, the foreign object detection model 20 may be a machine learning model that determines the presence or absence of a foreign object for each predetermined image region obtained by dividing an image, based on the image region and its surrounding region. For example, as shown in Fig. 4, when determining the presence or absence of a foreign object in image region A in a scrap image, the foreign object detection model 20 according to this embodiment may determine the detection result of a foreign object in image region A by taking into account the image content of surrounding region B of image region A and, more broadly, surrounding region C.

[0032] To achieve this, for example, the concept of receptive field in deep learning models may be utilized. The size of the receptive field is a quantity that indicates how far away information from pixels is ultimately used when determining whether a pixel is a feature region. In other words, taking the example of a case where the image region to be determined is each pixel, taking into account the image content of the image region as well as the surrounding image region of the image region when determining the image region to be determined means that the receptive field is greater than 1.

[0033] As shown in the figure, the foreign object detection model 20 may be a deep learning model capable of determining the presence or absence of a foreign object in an image area A of an input scrap image by taking into account one or more image areas B surrounding the image area A in the scrap image. For example, such a foreign object detection model 20 may be realized by a convolutional neural network. This allows foreign object detection for a certain image area or pixel to be performed by taking into account scrap not only in the image area or pixel but also in neighboring image areas or pixels, thereby enabling more accurate foreign object detection.

[0034] That is, compared to conventional techniques that subdivide an image of a scrap to be detected and perform foreign object detection independently for each of the subdivided images, performing foreign object detection for a certain image region or pixel by considering not only the image region or pixel but also the scrap in its neighboring image regions or pixels enables consistency between the detection results determined for adjacent images that are vertically and horizontally adjacent. This may enable more accurate foreign object detection, even when the foreign object being detected contains dust or coating. Note that, to ensure consistency between the detection results determined for adjacent images that are vertically and horizontally adjacent, it is desirable for image region B to include eight image regions of the same size as the image regions adjacent to image region A. In other words, it is preferable that the size of image region B is at least eight times larger than that of image region A. Furthermore, a larger image region C may also be considered. However, the foreign object detection model 20 according to the present disclosure is not necessarily limited to this and may be realized by other types of deep learning models, or may be a mathematical model, a statistical learning model, or the like.

[0035] In another embodiment, the foreign object detection model 20 may be a machine learning model including a self-attention mechanism. By utilizing such a self-attention mechanism, when considering surrounding image areas, instead of equally considering each surrounding image area, it is possible to weight and consider image areas that are important for determining whether or not a foreign object exists. In other words, depending on the shape of the foreign object to be detected captured in the scrap image, the presence or absence of a foreign object in the image area is determined by placing emphasis on one of the surrounding image areas of the image area to be determined. To achieve this, the foreign object detection model 20 may be configured to include a self-attention mechanism.

[0036] For example, as shown in FIG. 5A , if the foreign object detection model 20 has a self-attention mechanism, when determining whether or not a foreign object is present in image region A in a scrap image, the foreign object detection result for image region A can be determined by taking into account the image content of attention region D for image region A and the broader attention region E based on known knowledge of the foreign object's shape. As shown in the figure, the attention region is determined based on the shape of the coated copper wire. If the foreign object detection model 20 does not include a self-attention mechanism, when foreign object detection for a certain image region or pixel is performed by considering scrap in not only the image region or pixel but also the surrounding image regions or pixels, the neighboring image regions or pixels to be considered for a certain image region or pixel are constant (fixed) regardless of the input image. On the other hand, if the foreign object detection model 20 has a self-attention mechanism, as can be seen by comparing FIGS. 5A and 5B , when foreign object detection for a certain image region or pixel in an input image is performed, the surrounding image regions to be emphasized dynamically change depending on what is depicted where in each input image. That is, the foreign object detection model 20 may be configured to determine the surrounding area depending on the image content in the image area.

[0037] When the detection result of the foreign matter in the scrap image is obtained in this way, the detection unit 120 provides the detection result to the selection unit 130.

[0038] The sorting unit 130 determines a content ratio index of foreign matter in the scrap based on the detection results and sorts the scrap according to the content ratio index. Here, the content ratio index may be calculated as a ratio (e.g., area ratio) of the area where the foreign matter exists to the area where the scrap exists. For example, when scrap containing foreign matter is detected, the sorting unit 130 may calculate the number of areas X where the foreign matter exists based on the detection results from the foreign matter detection model 20, identify the number of areas Y where the scrap exists, and determine X / Y as the content ratio index.

[0039] Here, the areas where scrap exists may be identified by performing any known image processing on the scrap image. For example, as described in Patent Document 1, a color image of a group of iron scrap is captured, and individual scrap existence areas in the color image are detected from the brightness (I) signal of the captured color image. The sorting unit 130 calculates the number of areas or the number of pixels in the detected areas where scrap exists, and determines the ratio of the number of areas or the number of pixels in the areas where foreign matter exists detected by the foreign matter detection model 20 to the number of areas or the number of pixels in the areas where scrap exists, thereby determining the content ratio index.

[0040] Alternatively, the scrap detection model 30, which has been trained to detect areas where scrap exists from scrap images, may also be configured to detect areas where scrap exists. For example, when a scrap image is acquired, the scrap detection model 30 detects areas where scrap exists, as shown in FIG. 5C , and the sorting unit 130 calculates the number of areas or the number of pixels in the areas where the detected scrap exists, and determines the content ratio index by calculating the ratio of the number of areas or the number of pixels in the areas where the foreign matter exists detected by the foreign matter detection model 20 to the number of areas or the number of pixels in the areas where the scrap exists. Note that, similar to the foreign matter detection model 20 described above, the scrap detection model 30 can be acquired by training a machine learning model from images containing scrap and annotation information indicating the areas where the scrap exists in the captured image.

[0041] Specifically, regarding the sorting of scraps, the sorting unit 130 may decide to remove the scraps if the content ratio index is equal to or greater than a predetermined threshold. For example, as shown in FIG. 6, each scrap is placed on a conveyor of a transport device 50 from a feeder 40, and each scrap is transported by the movement of the conveyor of the transport device 50. The acquisition unit 110 acquires scrap images of each scrap taken by a camera 60 installed above the transport path. Then, the detection unit 120 detects foreign objects from the captured scrap images.

[0042] If the foreign matter content index of the captured scrap is equal to or greater than a predetermined threshold, the sorting unit 130 determines to remove the scrap and sends a control signal to the scrap removal device 70 to blow air onto the scrap to identify it as scrap to be removed. Note that scrap removal is not limited to air removal; it may also be performed by, for example, a robot. For example, the timing of blowing air may be delayed from the time the scrap to be removed passes the camera 60 by a time determined by the distance between the camera 60 and the scrap removal device 70 and the conveyor speed. Alternatively, multiple air outlets may be provided in the width direction, and air may be blown only from the outlet corresponding to the widthwise position where the scrap to be removed is located. In this case, the widthwise position of the scrap to be removed must be identified. This can be achieved by using the camera 60 to capture the entire width direction perpendicular to the conveyance direction of the conveyor, allowing the widthwise position of the scrap captured in the scrap image to be identified. On the other hand, if the foreign matter content ratio index of the photographed scrap is less than the predetermined threshold, the sorting unit 130 determines that the scrap is suitable as a raw material for iron making, and sorts the scrap as a raw material for iron making.

[0043] Furthermore, the sorting unit 130 may identify the position of the scrap on the conveying device 50 based on the detection results and operate the scrap removal device 70 in accordance with that position. That is, each scrap may be placed at various positions in the width direction on the conveyor of the conveying device 50. To efficiently perform removal by air or a robot using the scrap removal device 70, removal by air or a robot arm may be performed in accordance with the position of the scrap to be removed.

[0044] The predetermined threshold may be set based on the impact on the quality of the steel product being manufactured, the processing capacity and cost of recycling the removed scrap, or the type of steel product being manufactured. Alternatively, it may be set based on the size of the flame generated when the scrap is fed into a scrap melting furnace, such as an electric furnace or converter. In other words, it is desirable to set the threshold so that the flame generated when the scrap is fed is as small as possible to ensure safe furnace operation. For example, if scrap containing trace amounts of impurities that do not significantly affect the quality of the steel were removed, the recycling costs for the removed scrap would increase. For example, in recycling, to recover copper from the removed coated copper wire, the coated copper wire is separated and extracted from the scrap, and then a calcination process is performed in which the coating is removed by heating the wire in a furnace at a temperature above a predetermined temperature. Since the processing capacity of the calcination process is determined by the furnace capacity, attempting to calcinate even scrap containing trace amounts of impurities may exceed the processing capacity. Therefore, it is preferable to send only copper wire with a certain amount of coating to the calcination process. Many conventional impurity detection methods have involved binary classification of whether or not copper is present as an impurity, which can lead to the disposal of even scrap containing trace amounts of copper, placing a strain on the disposal process. However, according to an embodiment of the present invention, if the copper content is so small that it is unlikely to affect the quality of the iron products to be manufactured, the scrap can be used as is as a raw material for steelmaking without being disposed of through disposal, thereby avoiding this problem. The predetermined threshold value may be variable depending on the disposal capacity and cost, which may vary depending on facility management and operating conditions.

[0045] In one embodiment, the acquisition unit 110 may further acquire height information of the scrap, and the sorting unit 130 may determine a foreign matter content index (e.g., volume ratio) of the scrap based on the detection results of foreign matter in the scrap and the height information. Specifically, a height sensor may be installed near the camera 60, and the acquisition unit 110 may further acquire height information indicating the height H of the captured scrap. The sorting unit 130 may determine the content index by X / (Y×H). This allows the foreign matter content index to be estimated according to the height of scrap of different heights, enabling more appropriate scrap sorting. Examples of height sensors that can be used include a laser rangefinder, a stereo camera, and an optical cutting-type unevenness measurement device.

[0046] It is also possible to install multiple cameras 60. They may be installed so that they can photograph the scrap not only from above the conveyor of the transport device 50 but also from the side of the conveyor, for example. In this case, the height of the scrap can be calculated from the image taken from the side, so the installation of the height sensor described above can be omitted.

[0047] [Scrap inspection processing] Next, a scrap inspection process according to an embodiment of the present disclosure will be described. The scrap inspection process is performed by the scrap inspection device 100 described above, and more specifically, may be realized by one or more processors 102 of the scrap inspection device 100 executing one or more programs or instructions stored in one or more storage devices 101. The scrap inspection device 100 may also be realized by multiple computers, and the multiple computers may execute the scrap inspection process. Figure 7 is a flowchart showing the scrap inspection process according to an embodiment of the present disclosure.

[0048] 7, in step S101, the scrap inspection device 100 acquires an image of the scrap. For example, the scrap inspection device 100 may acquire a scrap image obtained by capturing an image of the scrap on the conveyor of the transport device 50 with a camera.

[0049] In step S102, the scrap inspection device 100 inputs scrap images to the foreign matter detection model 20. The foreign matter detection model 20 may be trained using training images of scrap and annotations indicating areas where foreign matters exist in the training images, and may be trained to detect foreign matters including, for example, coated copper wires and discolored copper wires.

[0050] Here, the foreign object detection model 20 may be a machine learning model that determines the presence or absence of a foreign object for each predetermined image region obtained by dividing an image based on the image region and its surrounding region. The foreign object detection model 20 may also be a machine learning model equipped with a self-attention mechanism.

[0051] In step S103, the scrap inspection device 100 acquires a detection result from the foreign matter detection model 20. For example, the detection result may indicate the presence or absence of one or more foreign matters in image regions or pixels obtained by dividing the input scrap image.

[0052] In step S104, the scrap inspection device 100 identifies the number (N) of detected foreign objects.

[0053] In step S105, the scrap inspection device 100 determines a content ratio index for each detected foreign matter based on the detection result. For example, the content ratio index may be determined as a ratio (e.g., area ratio, volume ratio, etc.) of the area where each foreign matter exists to the area where the scrap exists.

[0054] In step S106, the scrap inspection device 100 determines whether the content ratio index is equal to or greater than a predetermined threshold. If the content ratio index is equal to or greater than the predetermined threshold (S106: YES), the scrap inspection device 100 determines that the scrap should be removed, and proceeds to step S107. On the other hand, if the content ratio index is less than the predetermined threshold (S106: NO), the scrap inspection device 100 determines that the foreign matter contained in the scrap does not significantly affect the quality of the steel to be produced and that the scrap can be used as a raw material for steelmaking, and proceeds to step S108.

[0055] In step S107, the scrap inspection device 100 removes the scrap selected as the target for removal. Specifically, the scrap inspection device 100 transmits a control signal to the scrap removing device 70 to cause the scrap to be removed.

[0056] In step S108, the scrap inspection device 100 determines whether scrap removal determination based on the content ratio index has been completed for all of the foreign matter identified in step S104. If processing of all N foreign matter has been completed (S108: YES), the scrap inspection device 100 may end the scrap removal determination for the scrap image acquired in step S101 and start processing the next scrap image. On the other hand, if processing of all N foreign matter has not been completed (S108: NO), the scrap inspection device 100 returns to step S105 and executes steps S105 to S107 for the next foreign matter.

[0057] According to the above-described embodiment, the scrap inspection device 100 can detect foreign matter in scrap by using a machine learning model, even if the foreign matter to be detected contains dust, coating, or the like.

[0058] Although the embodiments of the present disclosure have been described in detail above, the present disclosure is not limited to the specific embodiments described above, and various modifications and variations are possible within the scope of the gist of the present disclosure as set forth in the claims. [Explanation of symbols]

[0059] 20 Foreign object detection model 30 Scrap Detection Model 40 Feeder 50 Conveyor 60 cameras 70 Scrap removal device 100 Scrap inspection equipment 110 Acquisition Department 120 Detector 130 Sorting Department

Claims

1. an acquisition unit that acquires an image of the scrap; a detection unit that acquires a detection result of a foreign object in the scrap from the image using a machine learning model; a sorting unit that determines a content ratio index indicating a ratio of an area where the foreign matter exists to an area where the scrap exists based on the detection result, and sorts the scrap according to the content ratio index; A scrap inspection device having the above features.

2. The scrap inspection device according to claim 1 , wherein the sorting unit determines to remove the scrap when the content ratio index is equal to or greater than a predetermined threshold value.

3. The scrap inspection device according to claim 1 , wherein the sorting unit identifies a position of the scrap on a conveying device based on the detection result, and operates a scrap removal device in accordance with the position.

4. The scrap inspection device according to claim 1 , wherein the machine learning model determines the presence or absence of the foreign matter for each predetermined image region obtained by dividing the image based on the image region and a peripheral region of the image region.

5. The scrap inspection device of claim 4 , wherein the machine learning model includes a self-attention mechanism.

6. The scrap inspection device according to claim 5 , wherein the machine learning model determines the surrounding area according to image content in the image area.

7. The acquisition unit further acquires height information of the scrap, The scrap inspection device according to claim 1 , wherein the sorting unit determines an index of a content ratio of the foreign matter in the scrap based on a result of detection of the foreign matter in the scrap and the height information.

8. The scrap inspection device according to claim 1 , wherein the foreign matter includes a coated copper wire or a discolored copper wire.

9. Obtaining an image of the scrap; obtaining a detection result of foreign matter in the scrap from the image using a machine learning model; determining a content ratio index indicating a ratio of an area where the foreign matter exists to an area where the scrap exists based on the detection result, and sorting the scrap according to the content ratio index; A computer-implemented scrap inspection method comprising:

10. Obtaining an image of the scrap; obtaining a detection result of foreign matter in the scrap from the image using a machine learning model; determining a content ratio index indicating a ratio of an area where the foreign matter exists to an area where the scrap exists based on the detection result, and sorting the scrap according to the content ratio index; A program that causes a computer to execute the following.

Images