JP2020095709A - Scrap grade determination system, scrap grade determination method, estimation device, learning device, learnt model generation method and program - Google Patents

Abstract

To provide a scrap grade determination system capable of improving determination accuracy for a grade ratio of iron scraps.SOLUTION: A scrap grade determination system comprises: a camera that captures a set of iron scraps, in which iron scraps of multiple grades according to size are mixed; an inference unit that estimates a ratio of each grade included in the set of iron scraps from the image generated by the camera, using a model that has been built in advance by machine learning with an image of the set of iron scraps as input data and a ratio of each grade as teacher data; and an output unit that outputs the estimation result by the inference unit.SELECTED DRAWING: Figure 1

Description

The present invention relates to a scrap grade determination system, a scrap grade determination method, an estimation device, a learning device, a learned model generation method, and a program.

Generally, at a scrap acceptance site of an electric furnace maker, a total of two persons, an inspector and a crane operator, visually check iron scraps lifted by a scrap loading trailer and a lifting magnet to determine the ratio for each product type and grade.

JP-A-7-286969

However, since the inspector visually confirms the iron scrap from a distant place for safety, the grade ratio determination accuracy may not be sufficient. In addition, the judgment criteria may vary depending on the experience of the inspector.

In particular, for iron scrap called heavy scrap, whose grade is determined according to size, it is difficult to determine the grade ratio because iron scraps of various sizes are mixed.

The present invention has been made in view of the above problems, and its main object is to provide a scrap grade determination system, a scrap grade determination method, an estimation device, and learning that can improve the determination accuracy of the grade ratio of iron scrap. An object is to provide an apparatus, a learned model generation method, and a program.

In order to solve the above problems, a scrap grade determination system according to one aspect of the present invention is a mixture of a plurality of grades of iron scrap according to size, a camera for capturing a set of iron scraps, and a set of iron scraps. Using an image as input data and using a learned model pre-built by machine learning with the proportion of each grade as teacher data, estimate the proportion of each grade contained in the set of iron scrap from the image generated by the camera. An inference unit and an output unit that outputs the estimation result by the inference unit are provided.

A scrap grade determination method according to another aspect of the present invention is a method of capturing a set of iron scraps in which a plurality of grades of iron scraps corresponding to sizes are mixed by a camera, and using an image of the set of iron scraps as input data. , Using a learned model pre-built by machine learning with the proportion of each grade as teacher data, estimating the proportion of each grade included in the set of iron scrap from the image generated by the camera, and outputting the estimation result To do.

In addition, the estimation device according to another aspect of the present invention includes an acquisition unit that acquires an image obtained by capturing a set of iron scraps by a camera, in which iron scraps of a plurality of grades according to sizes are mixed, and an iron scrap. Image of the set as input data, using the learned model pre-built by machine learning with the ratio of each grade as teacher data, the ratio of each grade included in the set of iron scrap from the image generated by the camera And an inference unit for estimating.

Further, a program of another aspect of the present invention is an acquisition unit for acquiring an image obtained by photographing a set of iron scraps by a camera, in which iron scraps of a plurality of grades according to sizes are mixed, and an iron scrap. Image of the set as input data, using the learned model pre-built by machine learning with the ratio of each grade as teacher data, the ratio of each grade included in the set of iron scrap from the image generated by the camera The computer functions as an inference unit, which estimates

Further, a learning device according to another aspect of the present invention includes an image obtained by photographing a set of iron scraps by a camera in which iron scraps of a plurality of grades corresponding to sizes are mixed, and the set of the iron scraps. A learning model for estimating the proportion of each grade contained in the set of iron scrap from the image, using the acquisition unit for obtaining the proportion of each grade as the input data, and the proportion of each grade as the teacher data. And a learning unit configured by machine learning.

A program according to another aspect of the present invention is included in an image obtained by photographing a set of iron scraps with a camera, in which a plurality of grades of iron scraps corresponding to sizes are mixed, and the set of iron scraps. An acquisition unit that acquires the ratio of each grade, and the learned image for estimating the ratio of each grade included in the set of iron scraps from the image using the image as input data and the ratio of each grade as teacher data. The computer is made to function as a learning unit that is constructed by machine learning.

A method for generating a learned model according to another aspect of the present invention is an image obtained by photographing a set of iron scraps mixed with a plurality of grades of iron scraps according to size by a camera, and the iron scraps. The ratio of each grade included in the set is acquired, the image is used as input data, the ratio of each grade is used as teacher data, and learning is performed to estimate the ratio of each grade included in the set of iron scrap from the image. Build the model by machine learning.

According to the present invention, it is possible to improve the accuracy of determining the grade ratio of iron scrap.

It is a figure which shows the structural example of the scrap grade determination system which concerns on embodiment. It is a figure which shows the grade and requirements of heavy waste. It is a top view of a scrap acceptance site. It is a side view of a scrap acceptance site. It is a block diagram which shows the structural example of the determination apparatus which concerns on 1st Embodiment. It is a figure which shows the example of an association with an image and a label. It is a figure which shows the example of an image. It is a flowchart which shows the example of a procedure of a learning phase. It is a figure which shows the construction example of the learned model. It is a flow figure showing an example of a procedure of an inference phase. It is a figure which shows the output example of an inference result. It is a block diagram which shows the structural example of the determination apparatus which concerns on 2nd Embodiment. It is a figure which shows the example of an association with an image and a label. It is a flow figure showing an example of a procedure of an inference phase. It is a figure which shows the output example of an inference result. FIG. 11 is a diagram showing an example of associating an image with a label according to the third embodiment. It is a figure which shows the output example of an inference result. It is a block diagram which shows the structural example of the determination apparatus which concerns on 4th Embodiment. It is a figure which shows the example of an association with an image and a label. It is a figure which shows the example of an image. It is a figure which shows the example of image processing. It is a block diagram which shows the structural example of the determination apparatus which concerns on 5th Embodiment. It is a flow figure showing an example of a procedure of an inference phase. It is a figure for explaining processing. It is a figure for explaining processing.

Embodiments of the present invention will be described below with reference to the drawings.

[System overview]
FIG. 1 is a diagram showing a configuration example of a scrap grade determination system 100 according to an embodiment. The scrap grade determination system 100 includes a determination device 1, a camera 2, a weight sensor 3, output units 6 and 7, and a crane 9. The determination device 1 can communicate with the camera 2 and the output units 6 and 7 via a communication network such as the Internet or a LAN.

The scrap grade determination system 100 determines the grade ratio of the iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y by the crane 9 at the scrap receiving site.

The camera 2 photographs the set of iron scraps S loaded on the scrap loading vehicle T and the set of iron scraps S lifted by the lifting magnet 91, and sends the generated image to the determination device 1. The arrangement of the camera 2 will be described in detail later.

The determination device 1 estimates the grade ratio of the iron scrap S from the image generated by the camera 2 and sends the estimation result to the output units 6 and 7.

The output units 6 and 7 output the estimation result from the determination device 1. The output unit 6 is a terminal such as a tablet computer carried by the inspector A, for example. The output unit 7 is, for example, a printer for issuing an inspection document.

The weight sensor 3 is, for example, a tread sensor embedded in the entrance of the scrap receiving site, and detects the load applied from the wheels of the scrap loading vehicle T.

The iron scraps S loaded on the scrap loading vehicle T are mixed with iron scraps S of various sizes. The iron scrap S is, for example, heavy scrap. As shown in FIG. 2, the grades HS to H4 according to the size (specifically, thickness, width or height, length) and weight are defined for heavy waste.

[Scrap receiving site]
3 and 4 are a plan view and a side view schematically showing a scrap receiving site. The scrap receiving site has a scrap yard Y and a vehicle approach area P adjacent to the scrap yard Y. The crane 9 installed at the scrap receiving site moves the lifting magnet 91 above the scrap yard Y and the vehicle approach area P.

The crane 9 is, for example, a trolley type overhead crane, and includes a pair of runway girders 92, a pair of crane girders 93 arranged over the pair of runway girders 92, and a club trolley 94 arranged over the pair of crane girders 93. Is equipped with. The runway girder 92 is supported by a building pillar 97.

The crane girder 93 is movable along the longitudinal direction of the runway girder 92, and the club trolley 94 is movable along the longitudinal direction of the crane girder 93. A hook block 98 is suspended from the club trolley 94, and a lifting magnet 91 is suspended from the hook block 98.

As shown in FIG. 3, some of the cameras 21 are installed so as to capture an image of the loading platform of the scrap loading vehicle T stopped in the vehicle entry area P (hereinafter, the camera 21 capturing the loading level of the scrap loading vehicle T). Is referred to as "vehicle camera 21"). A plurality of vehicle photographing cameras 21 are installed, for example, in the corridor 95 and the inspection room 96 at the scrap receiving site, and photograph the entire loading platform of the scrap loading vehicle T from above.

The corridor 95 and the inspection room 96 are places where the inspector has conventionally observed the loading platform of the scrap-laden vehicle T. By installing the vehicle photographing camera 21 in such a place, the field of view of the conventional inspector is reproduced. be able to.

When the scrap loading vehicle T enters the vehicle entry area P, the vehicle shooting camera 21 captures an image of the iron scrap S that is fully loaded on the bed of the scrap loading vehicle T. Subsequently, the vehicle shooting camera 21 shoots the iron scrap S left on the loading platform every time the lifting magnet 91 lifts a part of the iron scrap S from the iron scrap S stacked on the loading platform of the scrap loading vehicle T. The vehicle photographing camera 21 repeats this photographing until the iron scrap S disappears from the bed of the scrap loading vehicle T.

Here, of the iron scraps S left on the platform, the iron scraps S in the surface layer imaged by the vehicle imaging camera 21 are almost the same as the iron scraps S that are next lifted by the lifting magnet 91.

As shown in FIG. 4, the other camera 22 is installed so as to photograph the iron scrap S lifted by the lifting magnet 91 (hereinafter, the camera 22 that photographs the iron scrap S lifted by the lifting magnet 91). Is referred to as the "Rifmag camera 22". A plurality of the lift mag photographing cameras 22 are installed, for example, in the vicinity of the ground of the inspection room 96, the building pillar 97 and the scrap yard Y, and photograph the iron scrap S lifted by the lifting magnet 91 from the side or the bottom.

The crane 9 moves the lifting magnet 91 from the scrap loading vehicle T to the scrap yard Y so that the lifting magnet 91 passes a predetermined position. The lift-mag shooting camera 22 is installed so as to shoot toward the fixed position, and shoots the iron scrap S lifted by the lifting magnet 91 at the timing when the lifting magnet 91 passes through the fixed position.

The inspection room 96 is also a place where the inspector has conventionally observed the iron scrap S lifted by the lifting magnet 91. By installing the riff mag photographing camera 22 in such a place, the conventional inspector's view is reproduced. can do.

Further, the lift mag photographing camera 22 is installed on the opposite side of the inspection room 96 or near the ground of the scrap yard Y, and the iron scrap S lifted by the lifting magnet 91 is entirely taken from various angles. As a result, it is possible to use the image of the visual field, which was not obtained by the conventional inspection staff, for learning and inference.

The lifting magnet photographing camera 22 photographs the iron scrap S lifted by the lifting magnet 91 every time the lifting magnet 91 lifts a part of the iron scrap S from the iron scrap S stacked on the bed of the scrap loading vehicle T. The riffmag photographing camera 22 repeats this photographing until the iron scrap S disappears from the bed of the scrap loading vehicle T.

[First Embodiment]
FIG. 5 is a block diagram showing a configuration example of the determination device 1 according to the first embodiment. The determination device 1 is both a learning device and an estimation device. The determination device 1 includes a control unit 10 including a CPU, a RAM, a ROM, a non-volatile memory, an input/output interface, and the like. The determination device 1 is composed of, for example, one or a plurality of server computers.

The control unit 10 can access the database 20. The database 20 may be provided inside the determination device 1 or may be provided outside the determination device 1 and may be accessed via a communication network such as the Internet or a LAN.

The control unit 10 includes an image acquisition unit 11, a learning unit 13, an inference unit 14, a partial ratio determination unit 15, and an overall ratio determination unit 16. These functional units are realized by the CPU of the control unit 10 executing information processing according to a program loaded from the ROM or the nonvolatile memory to the RAM.

The program may be supplied via an information storage medium such as an optical disk or a memory card, or may be supplied via a communication network such as the Internet or a LAN.

[Learning phase]
Among the functional units realized by the control unit 10 shown in FIG. 5, the image acquisition unit 11 and the learning unit 13 are functional units for executing the learning phase and correspond to the learning device. The learning phase corresponds to the method of generating a learned model. The functional unit for executing the learning phase may be realized by a device different from the determination device 1.

The image acquisition unit 11 acquires an image of the set of iron scraps S generated by the camera 2. The acquired image is stored in the database 20 in association with the label including the grade ratio.

The learning unit 13 reads an image and a label associated with the image from the database 20, uses the image as input data, and uses the label as teacher data to generate a learned model for estimating the grade ratio of the set of the iron scrap S from the image. Build by machine learning.

FIG. 6 is a diagram showing an example of associating an image with a label. The association between the image and the label is managed by the table shown in FIG. The “image ID” is identification information of the image.

The "label" includes, as the grade ratio, the percentage of each grade of heavy scrap contained in the set of iron scrap S. The illustrated example includes grades HS, H1, H2, H3, and H4.

The “label” may be given a value determined by the inspector, or may be given an actually measured value.

The ratio is, for example, a weight ratio, but is not limited to this, and may be an area ratio or a volume ratio. Since the iron scrap S is made of the same material, the area ratio and the volume ratio are almost similar to the weight ratio.

FIGS. 7A and 7B are diagrams showing examples of images of a set of iron scraps S in which a plurality of grades of iron scrap S are mixed.

FIG. 7A is an example of an image generated by the vehicle photographing camera 21 photographing the loading platform of the scrap-loaded vehicle T. This image corresponds to the entry whose image ID is "aaaa" in the table shown in FIG.

When a plurality of vehicle photographing cameras 21 simultaneously photograph the loading platform of the scrap loading vehicle T from different directions, the same label is given to the generated images. Alternatively, a set of a plurality of generated images may be treated as one input data.

FIG. 7( b) is an example of an image generated by the rifmag photographing camera 22 photographing the scrap iron S lifted by the lifting magnet 91. This image corresponds to the entry whose image ID is "bbbb" in the table shown in FIG.

When the iron scraps S lifted by the lifting magnet 91 from different directions at the same time are photographed by the plurality of riff mag photographing cameras 22, the same label is given to the generated images. Alternatively, a set of a plurality of generated images may be treated as one input data.

The image used for machine learning is not necessarily limited to the image captured by the vehicle photographing camera 21 or the riffmag photographing camera 22, and may be, for example, an iron scrap image photographed in another place.

Further, an image of a set of iron scraps S for each grade (that is, an image including only the iron scraps S of the same grade) as shown in FIG. 7C may be used for machine learning. This image corresponds to the entry whose image ID is "cccc" in the table shown in FIG.

Further, a partial image cut out from the image of the set of iron scraps S in accordance with a user's operation input may be used for machine learning. For example, it is possible to expect an improvement in estimation accuracy by cutting out a portion of interest from an experienced inspector from the image and using it for machine learning.

FIG. 8 is a flowchart showing a procedure example of a learning phase as a method of generating a learned model. The control unit 10 functions as the image acquisition unit 11 and the learning unit 13 by executing the information processing shown in FIG.

First, the control unit 10 acquires an image and a label associated with the image (S11, processing as an acquisition unit). In the learning phase, the image and the label associated with the image are stored in the database 20 in advance and read from the database 20 as needed.

Next, the control unit 10 extracts a part of a set of prepared images and labels associated with the images as training data for machine learning (S12), and executes machine learning using the extracted training data. (S13, processing as learning unit). Machine learning is performed using images as input data and labels as teacher data. Thereby, a learned model for estimating the grade ratio of the set of the iron scrap S from the image is constructed.

Next, the control unit 10 extracts a part other than the training data as test data from a set of a plurality of prepared images and labels associated with the images (S14), and has learned using the extracted test data. The model is evaluated (S15). After that, the control unit 10 stores the learned model whose evaluation is equal to or higher than the predetermined value in the database 20 (S16), and ends the learning phase.

FIG. 9 is a diagram showing an example of constructing a learned model. The trained model is, for example, a convolutional neural network, and includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Particularly, a deep neural network in which neurons are combined in multiple stages is suitable.

The output layer is provided with a number of elements corresponding to the labels (see FIG. 6). That is, a plurality of elements corresponding to the proportion of each grade of heavy scrap contained in the set of iron scrap S are provided.

A softmax function, for example, is used for the element relating to the grade ratio of the output layer (the element relating to the ratio of each grade of heavy waste). In this case, since the output value is obtained as a real number between 0 and 1, the output value can be interpreted as a ratio.

The illustrated convolutional neural network is merely an example, and the layer structure is not limited to this, and the number of layers of the convolutional layer, the pooling layer, and the total connection layer may be different. In addition, machine learning other than neural networks such as support vector machines, Gaussian processes, and decision trees may be used.

[Inference phase]
Of the functional units realized by the control unit 10 shown in FIG. 5, the image acquisition unit 11, the inference unit 14, the partial ratio determination unit 15, and the overall ratio determination unit 16 are functional units for executing the inference phase, It corresponds to an estimation device. The inference phase corresponds to the scrap grade determination method.

The image acquisition unit 11 acquires an image of the set of iron scraps S generated by the camera 2 and outputs it to the inference unit 14. In the inference phase, the image acquired by the image acquisition unit 11 is directly input to the inference unit 14. Not limited to this, the image may be stored in the database 20 and then read.

The inference unit 14 uses the image acquired by the image acquisition unit 11 as input data, and uses the learned model constructed in the learning phase to estimate the grade ratio of the set of the iron scrap S from the image. The inference unit 14 outputs the estimation result to the partial ratio determination unit 15.

Every time the lifting magnet 91 lifts the iron scrap S, the partial ratio determination unit 15 estimates the grade ratio estimated from the image generated by the vehicle photographing camera 21 immediately before the lifting, and the image generated by the lift mag photographing camera 22 at the time of lifting. The grade ratio of the lifted iron scrap S is determined based on the grade ratio estimated from the above.

The overall ratio determination unit 16 is unloaded from the scrap loading vehicle T to the scrap yard Y based on the grade ratio of the iron scrap S determined by the partial ratio determination unit 15 every time the lifting magnet 91 lifts the iron scrap S. The grade ratio of the entire iron scrap S is determined.

FIG. 10 is a flowchart showing a procedure example of an inference phase as a scrap grade determination method. The control unit 10 functions as the image acquisition unit 11, the inference unit 14, the partial ratio determination unit 15, and the overall ratio determination unit 16 by executing the information processing shown in FIG.

First, the control unit 10 acquires an image captured by the vehicle imaging camera 21 (see FIG. 3) (S21, processing as an acquisition unit). The image photographed by the vehicle photographing camera 21 is an image of the iron scrap S stacked on the bed of the scrap-carrying vehicle T before being lifted by the lifting magnet 91.

Next, the control unit 10 uses the acquired image as input data and estimates the grade ratio (rate of each grade of heavy scraps) contained in the iron scrap S using the learned model created in the learning phase ( S22, processing as an inference unit). When acquiring a plurality of images simultaneously photographed by a plurality of vehicle photographing cameras 21, the grade ratios estimated from the respective images are averaged.

Next, the control unit 10 acquires the image captured by the Rifumag camera 22 (see FIG. 4) (S23, processing as an acquisition unit). The image photographed by the lift mag photographing camera 22 is an image of a part of the iron scrap S lifted by the lifting magnet 91 from the iron scrap S stacked on the bed of the scrap loading vehicle T.

Next, the control unit 10 uses the acquired image as input data and estimates the grade ratio (the percentage of each grade of heavy scrap) contained in the iron scrap S using the learned model created in the learning phase ( S24, processing as an inference unit). When acquiring a plurality of images simultaneously photographed by a plurality of Rifmag photographing cameras 22, the grade ratios estimated from the respective images are averaged.

Next, the control unit 10 averages the grade ratios estimated in S22 and S24 and determines the averaged grade ratio as the grade ratio of the iron scrap S lifted by the lifting magnet 91. (S25, processing as a partial ratio determination unit). For the averaging, an arithmetic average may be used, or, for example, a weighted average weighted such that one is a main and the other is a sub may be used.

Here, of the iron scraps S stacked on the bed of the scrap-loaded vehicle T before being lifted by the lifting magnet 91, the iron scrap S of the surface layer part taken by the vehicle shooting camera 21 is taken by the riff mag shooting camera 22. It is almost the same as the iron scrap S lifted by the lifting magnet 91. That is, it can be said that the image acquired in S21 and the image acquired in S23 represent almost the same iron scrap S from different viewpoints.

Therefore, in S25, the grade ratio of the iron scrap S lifted by the lifting magnet 91 is determined by averaging the grade ratios estimated in S22 and S24.

The control unit 10 repeats the processing of S21 to S25 each time the lifting magnet 91 lifts the iron scrap S until all the iron scrap S is unloaded from the scrap loading vehicle T (S26).

When all the iron scraps S are unloaded from the scrap loading vehicle T (S26: YES), the control unit 10 totals the grade ratios determined in S25 every time the lifting magnet 91 lifts the iron scrap S, and averages the grade ratios. By calculating the value, the grade ratio of the entire iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y is determined (S27, processing as an overall ratio determination unit). For the averaging, an arithmetic average may be used, or for example, a weighted average obtained by multiplying a weight according to weight may be used.

After that, the control unit 10 transmits the grade ratio of the entire iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y determined in S27 to the output units 6 and 7 as an estimation result (S28). The output units 6 and 7 output the received estimation result by, for example, displaying or printing.

FIG. 11 is a diagram showing an output example of the inference result. The estimation result includes the grade ratio of the entire iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y (the ratio of each grade of heavy scrap).

According to the above-described embodiment, the grade ratio of the set of the iron scrap S is estimated from the image by using the learned model created in advance by machine learning, so that the estimation accuracy can be improved. ..

Further, according to the embodiment, it is possible to eliminate variations in grade determination levels among inspectors and level the grade determination levels.

In addition, according to the embodiment, it is possible to further improve the estimation accuracy by using the images obtained by shooting the set of the iron scrap S from various angles by the plurality of cameras 2.

Further, the camera 2 should be installed in a place where the inspector could not be placed so far, for example, by taking a picture of the iron scrap S lifted by the lifting magnet 91 from below with the riffmag camera 22. As a result, it is possible to use the images of the set of the iron scrap S captured from various angles for learning and inference while ensuring the safety of the inspectors.

Further, according to the embodiment, since the grade ratio estimation is automated, the inspection work can be performed by one driver of the crane 9, and the labor of the inspection staff can be saved. Furthermore, by combining the automation of the operation of the crane 9, it is possible to receive iron scrap, unload it, and unattend the inspection work.

[Second Embodiment]
The second embodiment will be described below. The same numbers are assigned to the same configurations or procedures as those in the above-described embodiment, and detailed description thereof will be omitted.

FIG. 12 is a block diagram showing a configuration example of the determination device 1 according to the second embodiment. The control unit 10 includes, in addition to the image acquisition unit 11, the learning unit 13, the inference unit 14, the partial ratio determination unit 15, and the overall ratio determination unit 16, a weight measurement unit 12, a weight calculation unit 17, a weight integration unit 18, and The weight correction unit 19 is included.

[Learning phase]
The image of the set of the iron scraps S acquired by the image acquisition unit 11 is stored in the database 20 in association with the label including the grade ratio and the weight.

The learning unit 13 reads out the image and the label associated with the image from the database 20, uses the image as input data, and uses the label as teacher data to estimate the grade ratio and weight of the set of the iron scrap S from the image. Build the model by machine learning.

FIG. 13 is a diagram showing an example of associating an image with a label. The “label” is the ratio of each grade (HS, H1, H2, H3, and H4 in this example) of heavy scraps included in the set of iron scrap S as a grade ratio, as well as the set of iron scrap S. Including the weight of.

Note that the weight label is not attached to the image (see FIG. 7A) generated by the vehicle photographing camera 21 photographing the loading platform of the scrap loading vehicle T, which corresponds to the entry with the image ID “aaaa”. ..

[Inference phase]
Of the functional units realized by the control unit 10 shown in FIG. 12, the image acquisition unit 11, the inference unit 14, the partial ratio determination unit 15, the overall ratio determination unit 16, the weight calculation unit 17, the weight integration unit 18, and the weight correction unit. Reference numeral 19 is a functional unit for executing the inference phase, which corresponds to an estimation device.

The inference unit 14 uses the image acquired by the image acquisition unit 11 as input data and estimates the grade ratio and weight of the set of the iron scrap S from the image using the learned model constructed in the learning phase. The inference unit 14 outputs the estimation result related to the grade ratio to the partial ratio determination unit 15, and outputs the estimation result related to the weight to the weight calculation unit 17.

The weight measuring unit 12 obtains the weight of the scrap loading vehicle T when the scrap loading vehicle T enters and the weight of the scrap loading vehicle T when the scrap loading vehicle T exits from the weight sensor 3 provided at the entrance of the scrap receiving site, and calculates the difference between them to obtain the scrap loading vehicle. The total weight (measured weight) of the iron scrap S unloaded from T to the scrap yard Y is calculated.

The weight calculation unit 17 acquires the weight of the lifted iron scrap S estimated by the inference unit 14 every time the lifting magnet 91 lifts the iron scrap S.

Further, the weight calculation unit 17 is lifted based on the grade ratio of the lifted iron scrap S determined by the partial ratio determination unit 15 and the weight of the lifted iron scrap S estimated by the inference unit 14. The weight of each grade contained in the iron scrap S is calculated.

The weight accumulating unit 18 accumulates the weight of the iron scrap S estimated by the inference unit 14 every time the lifting magnet 91 lifts the iron scrap S, so that the iron scraps unloaded from the scrap loading vehicle T to the scrap yard Y are unloaded. The total weight of S (estimated weight) is calculated.

In addition, the weight accumulating unit 18 accumulates the weights of the respective grades included in the lifted iron scrap S calculated by the weight calculating unit 17 every time the lifting magnet 91 lifts the iron scrap S, and thus the scrap loading vehicle. The weight (estimated weight) of each grade contained in the entire iron scrap S unloaded from T to the scrap yard Y is also calculated.

The weight correcting unit 19 corrects the estimated weight of the entire iron scrap S calculated by the weight integrating unit 18, based on the measured weight of the entire iron scrap S calculated by the weight measuring unit 12.

The weight correction unit 19 also calculates the weight of the entire iron scrap S calculated by the weight measuring unit 12 and the estimated weight of the entire iron scrap S calculated by the weight integrating unit 18, based on the ratio. The estimated weight of each grade contained in the entire iron scrap S calculated by 18 is also corrected.

FIG. 14 is a flowchart showing a procedure example of an inference phase as a scrap grade determination method. The control unit 10 executes the information processing shown in the figure according to a program to obtain an image acquisition unit 11, an inference unit 14, a partial ratio determination unit 15, an overall ratio determination unit 16, a weight calculation unit 17, a weight integration unit 18, And functions as the weight correction unit 19.

The control unit 10 uses the image of the Rifumag photography camera 22 (see FIG. 4) acquired in S23 as input data, and uses the learned model created in the learning phase to change the iron scrap S lifted by the lifting magnet 91. The weight is estimated together with the included grade ratio (S31, processing as an inference unit).

In S22, the estimation using the learned model is performed by using the image of the vehicle photographing camera 21 (see FIG. 3) acquired in S21 as input data. Here, the weight of the iron scrap S is estimated by the learned model. Even if you are, don't use it.

Next, the control unit 10 multiplies the lifting magnet 91 determined or estimated in S25 and S31 by the grade ratio of the iron scrap S lifted and its weight to include the iron scrap S lifted by the lifting magnet 91 in the weight ratio. The weight of each grade is calculated (S32, processing as a weight calculation unit).

Next, the control unit 10 integrates the weight of the iron scrap S lifted by the lifting magnet 91 estimated in S31 (S33, processing as a weight integrating unit). Specifically, the control unit 10 updates the accumulated weight by adding the weight estimated in S31 to the accumulated weight stored in the memory up to that point. Further, the control unit 10 similarly integrates the weight of each grade calculated in S32.

Next, the control unit 10 determines the weight accumulated in S33 each time the lifting magnet 91 lifts the iron scrap S as the total weight of the iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y ( S34). Further, the control unit 10 similarly determines the weight of each grade accumulated in S33 as the weight of each grade included in the entire iron scrap S.

Next, when the total weight (estimated weight) of the iron scrap S determined in S34 is different from the total weight (measured weight) of the iron scrap S measured by the weight sensor 3 (S35: YES). The estimated weight is corrected based on the measured weight (S36). For example, the measured weight is set to a correct value, and the estimated weight is replaced with the measured weight.

Further, the control unit 10 multiplies the weight of each grade contained in the entire iron scrap S determined in S34 by the ratio of the measured weight to the estimated weight (measured weight/estimated weight), to thereby obtain the entire iron scrap S. Correct the weight of each grade included.

After that, the control unit 10 determines the grade ratio and the weight of the entire iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y, which are determined in S27 and S34 (or S36), as the estimation result, and outputs the output unit 6, 6. 7 (S28). The output units 6 and 7 output the received estimation result by, for example, displaying or printing.

FIG. 15 is a diagram showing an output example of the inference result. The estimation result includes the grade ratio of the entire iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y (proportion of each grade of heavy scrap), the total weight of the iron scrap S, and the total iron scrap S. The weight of each grade is included.

The weight of each grade contained in the entire iron scrap S may be calculated as follows. Conventionally, the inspector observes the bed of the scrap loading vehicle T loaded with the entire iron scrap S that has just entered the scrap receiving site, and subtracts the vehicle weight from the weight of the scrap loading vehicle T at the time of entry. Considering the weight of the entire iron scrap S, the weight of each grade included in the entire iron scrap S is roughly determined.

In order to obtain an outline similar to this, from the image of the vehicle photographing camera 21 that photographs the bed of the scrap loading vehicle T on which the entire iron scrap S that has just entered the scrap receiving site is loaded, the images of each grade included in the iron scrap S are obtained. The weight of each grade included in the entire iron scrap S may be calculated by estimating the ratio and multiplying it by the weight of the entire iron scrap S measured by the weight sensor 3.

[Third Embodiment]
The third embodiment will be described below. The same numbers are assigned to the same configurations or procedures as those in the above-described embodiment, and detailed description thereof will be omitted. The configuration example and the procedure example of the determination device 1 according to the third embodiment are the same as those of the determination device according to the first embodiment shown in FIGS. 5, 8 and 10 above.

At the scrap receiving site, the iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y by the crane 9 contains, in addition to heavy scraps, shredder scraps, new cuttings, or steel dull powder, water and dust, etc. Sometimes.

Therefore, in the third embodiment, not only the proportion of each grade of heavy scrap contained in the set of iron scrap S, but also the proportion of shredder scrap, new cutting, steel Daray powder, and moisture/dust is learned and estimated.

That is, as shown in FIG. 16, the “label” used in the learning phase is, in addition to the proportion of each grade (HS, H1, H2, H3, and H4) of heavy scraps included in the set of iron scraps S, It also includes shredder scraps, new cuttings, steel Dalai powder, and water/dust ratios.

The learning unit 13 (see FIG. 5) uses such a label as teacher data, together with the ratio of each grade of heavy scraps included in the set of iron scraps S from the image, shredder scraps, new cutting, steel dull powder, and water content. Construct a trained model for estimating the dust ratio by machine learning.

Then, the inference unit 14 uses the learned model constructed by the learning unit 13 and the ratio of each grade of the heavy scraps included in the set of the iron scraps S from the image acquired by the image acquisition unit 11 along with the shredder scraps and the new scraps. Estimate the proportion of cutoff, steel dwarf powder, and water/dust.

The partial ratio determination unit 15 determines the ratio of each grade of heavy scraps included in the set of iron scraps S lifted by the lifting magnet 91, as well as the ratio of shredder scraps, new cuttings, steel Daray powder, and moisture/dust.

The overall ratio determining unit 16 determines the ratio of each grade of heavy scrap contained in the entire iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y, together with shredder scrap, new cutting, steel dull powder, and moisture/dust. Determine the proportion.

FIG. 17 is a diagram showing an output example of the inference result. The estimation result shows the ratio of each grade of heavy scrap contained in the entire iron scrap S unloaded from the scrap loading vehicle T to the scrap yard Y, as well as the ratio of shredder scrap, new cutting, steel Daray powder, and moisture/dust. include.

It should be noted that the ratio of each grade may be learned and estimated for shredder scraps, new cuttings, or steel dwarf powder.

The third embodiment described above may be further combined with the weight learning/estimation described in the second embodiment. That is, it is possible to learn and estimate the ratio and weight of each grade of heavy scraps, shredder scraps, new cuttings, steel Daray powder, and moisture/dust contained in the set of iron scraps S.

[Fourth Embodiment]
The fourth embodiment will be described below. The same numbers are assigned to the same configurations or procedures as those in the above-described embodiment, and detailed description thereof will be omitted.

At the scrap receiving site, the iron scrap S that is unloaded from the scrap loading vehicle T to the scrap yard Y by the crane 9 includes heavy scraps, non-ferrous metal such as copper or lead, hermetically sealed objects, and those having a specified size or weight or more. Or, exempted products such as dango-shaped reinforcing bars may be mixed.

Therefore, in the fourth embodiment, together with the proportion of each grade of heavy scrap contained in the set of iron scrap S, the presence/absence of rejected products is also learned/estimated.

FIG. 18 is a block diagram showing a configuration example of the determination device 1 according to the fourth embodiment. The control unit 10 includes an alarm determination unit 81 in addition to the image acquisition unit 11, the learning unit 13, the inference unit 14, the partial ratio determination unit 15, and the overall ratio determination unit 16.

The alarm output unit 8 includes, for example, a revolving light and a siren, and outputs an alarm when it receives an output command from the determination device 1 (alarm determination unit 81).

As shown in FIG. 19, the “label” used in the learning phase includes the ratio of each grade (HS, H1, H2, H3, and H4) of the heavy scraps included in the set of the iron scraps S, as well as the exclusion of acceptance. It also includes the presence or absence of goods.

Not limited to this, the “label” is provided with items for each type of non-ferrous metal such as copper or lead, hermetically sealed objects, those having a specified size or weight or exempted goods such as dango-shaped reinforcing bars. Preferably.

The learning unit 13 constructs a learned model for estimating the presence or absence of rejected products from the image by machine learning by using such labels as teacher data and the ratio of each grade of heavy scrap contained in the set of iron scrap S from the image. To do.

When a convolutional neural network (see FIG. 9) is used for the learned model, it is preferable to use, for example, a sigmoid function that outputs a value of 0 or more and 1 or less as an element related to the presence or absence of the exclusion product in the output layer.

Then, the inference unit 14 uses the learned model constructed by the learning unit 13 and is included in the set of iron scraps S from the image acquired by the image acquisition unit 11 every time the lifting magnet 91 lifts the iron scrap S. Estimate the presence or absence of exempted goods together with the ratio of each grade of heavy waste. The inference unit 14 outputs the estimation result regarding the presence or absence of the excluded product to the alarm determination unit 81.

The alarm determination unit 81 determines the presence/absence of an exclusion product based on the estimation result from the inference unit 14, and when it determines that there is an exclusion product, sends an output command to the alarm output unit 8 to issue an alarm. Output. For example, when the value indicating the presence or absence of the acceptance excluded product (for example, the value of 0 or more and 1 or less) is equal to or more than the threshold value, it is determined that there is the acceptance excluded product.

Regarding non-ferrous metals such as copper and lead, there is little difference in appearance from the iron scrap S, and the accuracy of identifying non-ferrous metals may not be sufficient. Therefore, for example, metals using laser, X-rays, or γ-rays A component analyzer or the like may be used supplementarily.

The weight learning/estimation described in the second embodiment may be further combined with the fourth embodiment described above. That is, the proportion and weight of each grade of heavy scraps included in the set of iron scraps S may be learned and estimated, and the presence or absence of rejected products may be learned and estimated.

Further, the fourth embodiment may be further combined with the learning/estimation of the ratio of shredder dust, new cutting, steel Daray powder, and moisture/dust described in the third embodiment. That is, by learning and estimating the grades of heavy scraps, shredder scraps, new cuttings, steel Daray powder, and water/dust ratios contained in the set of iron scraps S, and learning/estimating the presence or absence of rejected products. Good.

Further, in the fourth embodiment, the weight learning/estimation described in the second embodiment, and the ratio of the shredder waste, the new cutting, the steel Daray powder, and the moisture/dust described in the third embodiment are learned. -Both estimations may be combined. That is, while learning and estimating the grades and weights of heavy scraps included in the set of iron scraps S, shredder scraps, new cuttings, steel Darai powder, and moisture/dust, and whether or not there are rejected products. You may.

[Fifth Embodiment]
The fifth embodiment will be described below. The same numbers are assigned to the same configurations or procedures as those in the above-described embodiment, and detailed description thereof will be omitted.

[Learning phase]
FIG. 20 is a diagram showing an example of the original image L before being processed into the learning image in the present embodiment. The original image L is an image of a set of the iron scrap S lifted by the lifting magnet 91 taken by a camera such as the Rifmag shooting camera 22 (see FIG. 4).

The set of the iron scraps S lifted by the lifting magnet 91 usually has a substantially inverted conical shape (a substantially inverted triangular shape in a side view) that gradually narrows downward from the lifting magnet 91.

The lifting magnet 91 is an example of a hanging tool. Not limited to this, a grab bucket or the like may be used as the hanging tool.

FIG. 21 is a diagram showing an example of image processing performed on the original image L. First, as shown in FIG. 21A, the range N of the lifting magnet 91 in the original image L is specified. The range N of the lifting magnet 91 is determined and designated by the user, for example.

Next, as shown in FIG. 21B, the partial image LO determined based on the range N of the lifting magnet 91 is cut out from the original image L. The partial image LO is defined so as to include at least the range below the lifting magnet 91.

Specifically, the left and right ends of the partial image LO are located slightly outside the left and right ends of the lifting magnet 91. The horizontal width of the partial image LO is set to be slightly wider (for example, about 1.1 to 1.5 times) than the horizontal width of the lifting magnet 91.

The upper end of the partial image LO is located at the upper end of the lifting magnet 91. The lower end of the partial image LO is located below the lower end of the lifting magnet 91 by a predetermined height.

Specifically, the lower end of the partial image LO has the same horizontal width as the partial image LO, has a base overlapping with the lower end of the hanging portion 91, and is a range of an inverted triangle (for example, an inverted right triangle) that gradually narrows downward. The position is the lower end (vertex) of the NN.

Next, as shown in FIG. 21C, a mask process is performed in which the range NN of the inverted triangle is the display area and the other range NF is the non-display area of the partial image LO. For example, the range NF adjacent to the left and right of the inverted triangular range NN is filled with black or the like or made transparent.

The partial image LO extracted and processed in this way is used as a learning image in the learning phase. According to this, in the learning image, the background portion that does not need attention is excluded, so that the learning effect can be improved.

In the learning phase, in addition to the learned model for estimating the grade ratio of the set of iron scrap S from the image, a learned model for estimating the range of the lifting magnet 91 in the image (hereinafter, range estimation (Learned model) is further built.

For example, the learned model for range estimation is constructed by machine learning using the original image L shown in FIG. 21A as input data and the range N of the lifting magnet 91 specified in the original image L as teacher data. ..

The learned model for range estimation is an object detection model such as YOLO (You Only Look Once), SSD (Single Shot MultiBox Detector), or Faster R-CNN.

[Inference phase]
FIG. 22 is a block diagram showing a configuration example of the determination device 1 according to the fifth embodiment. The control unit 10, in addition to the image acquisition unit 11, the learning unit 13, the inference unit 14, the partial ratio determination unit 15, the overall ratio determination unit 16, the weight calculation unit 17, and the weight correction unit 19, the range estimation unit 41 and the processing. The part 42 is included.

FIG. 23 is a flowchart showing an example of the procedure of the inference phase as the scrap grade determination method realized by the determination device 1. The control unit 10 of the determination device 1 executes the information processing shown in FIG.

First, when the control unit 10 acquires an image (hereinafter, referred to as a camera image) captured by the riff mug camera 22 (S23), the control unit 10 uses the range estimation learned model described above to detect the lifting magnet 91 in the camera image. The range is estimated (S49, processing as the range estimation unit 41).

Next, the control unit 10 performs processing on the camera image based on the estimated range of the lifting magnet 91 (S50, processing as the processing unit 42). The processing processing is the same as the processing processing in the learning phase described above (see FIG. 21).

Specifically, first, as shown in FIG. 24A, the range M of the lifting magnet 91 is estimated in the camera image C acquired from the lift mag photographing camera 22.

Next, as shown in FIG. 21B, the partial image CO determined based on the range M of the lifting magnet 91 is cut out from the camera image C. The partial image CO is defined so as to include at least the range below the lifting magnet 91.

Specifically, the left and right ends of the partial image CO are located slightly outside the left and right ends of the lifting magnet 91. The horizontal width of the partial image CO is set to be slightly wider (for example, about 1.1 to 1.5 times) than the horizontal width of the lifting magnet 91.

The upper end of the partial image CO is located at the upper end of the lifting magnet 91. The lower end of the partial image CO is located below the lower end of the lifting magnet 91 by a predetermined height.

Specifically, the lower end of the partial image CO has the same horizontal width as the partial image CO, has a base overlapping with the lower end of the hanging portion 91, and is a range of an inverted triangle (for example, an inverted right triangle) that gradually narrows downward. It is the position of the lower end (vertex) of MN.

Next, as shown in FIG. 24C, a mask process is performed in which the range MN of the inverted triangle is the display area and the other range MF is the non-display area of the partial image CO. For example, the range MF adjacent to the left and right of the range MN of the inverted triangle is filled with black or the like or made transparent.

After that, the control unit 10 uses the partial image CO extracted and processed in this way as input data, and uses the learned model for estimating the grade ratio and the weight to collect the iron scrap S reflected in the partial image CO. The grade ratio included in is estimated (S31-S33 of FIG. 23).

According to this, in the image used as the input data when estimating the grade ratio and the weight, as in the learning image, the background portion that does not need attention is excluded, so that the estimation accuracy of the grade ratio is improved. It is possible to plan.

The shape of the range MN used as the display area in the mask processing is not limited to the inverted triangle. This applies not only to the masking process in the inference phase, but also to the masking process in the learning phase.

For example, the shape of the range MN may be an inverted trapezoidal shape that gradually narrows downward as shown in FIGS. 25(c) and 25(d), or as shown in FIGS. 25(e) and 25(f). It may have a rectangular shape. Further, as shown in FIGS. 25(b), (d), and (f), the suspending tool 91 above the range MN and both sides thereof may be filled with black or the like as a non-display area.

[Other]
In the above embodiment, the weight is included in the “label” used as teacher data during learning (see FIG. 6), and the weight is estimated together with the grade ratio of the iron scrap S during inference (see FIG. 10). Conversely, a weight measuring unit for measuring the weight of the iron scrap S lifted by the lifting magnet 91 may be provided, and the measured weight may be used as input data for learning/estimation. That is, the weight may be used together with the image as the input data at the time of learning, and the grade ratio may be estimated from the image and the weight of the iron scrap S lifted at the time of inference.

Furthermore, a voice measurement unit that measures the voice when the lifting magnet 91 lifts the iron scrap S may be provided, and the measured voice may be used as input data for learning/estimation. That is, the voice may be used together with the image as the input data at the time of learning, and the grade ratio may be estimated from the image and the voice of the iron scrap S lifted at the time of inference.

Although the embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above, and it is needless to say that various modifications can be made by those skilled in the art.

1 determination device (learning device, estimation device), 2, 21, 22 camera, 3 weight sensor, 6, 7 output unit, 10 control unit, 11 image acquisition unit, 12 weight measurement unit, 13 learning unit, 14 inference unit, 15 partial ratio determination unit, 16 total ratio determination unit, 17 weight calculation unit, 18 weight integration unit, 19 weight correction unit, 20 database, 8 alarm output unit, 9 crane, 91 lifting magnet, 92 runway girder, 93 crane girder, 94 Club Trolley, 95 Walkway, 96 Inspection Room, 97 Building Pillar, 98 Hook Block, 100 Scrap Grade Judgment System, A Inspector, Y Scrap Yard, S Iron Scrap, T Scrap Loading Vehicle, P Vehicle Entry Area

Claims (24)

A camera that shoots a set of iron scrap, in which multiple grades of iron scrap according to size are mixed,
Using an image of a set of iron scraps as input data, using a learned model pre-built by machine learning with the ratio of each grade as teacher data, each grade included in the set of iron scraps from the image generated by the camera An inference unit that estimates the ratio of
An output unit for outputting the estimation result by the inference unit,
A scrap grade determination system equipped with. Further equipped with a crane that lifts some iron scrap from the piled iron scrap,
Every time the crane lifts a part of the iron scrap, the inference unit, from the image of the part of the iron scrap lifted by the crane taken by the camera, each grade included in the part of the iron scrap. Estimate the proportion of,
The scrap grade determination system according to claim 1. Further equipped with a crane that lifts some iron scrap from the piled iron scrap,
Every time the crane lifts a part of the iron scrap, the inference unit determines, from the image of the stacked iron scrap before the crane lifts the part of the iron scrap, the part of the iron scrap. Estimate the proportion of each grade included in the scrap,
The scrap grade determination system according to claim 1. Percentage of each grade estimated from images of some iron scrap lifted by the crane, and percentage of each grade estimated from images of the stacked iron scrap before the crane lifts some iron scrap Based on, and further comprising a partial proportion determination unit that determines the proportion of each grade contained in the iron scrap of the part,
The scrap grade determination system according to claim 2 or 3. Based on the proportion of each grade estimated each time the crane lifts some iron scrap, further comprises an overall proportion determination unit that determines the proportion of each grade included in the entire stacked iron scrap.
The scrap grade judging system according to any one of claims 2 to 4. The inference unit uses an image of a set of iron scrap as input data, and uses a learned model pre-constructed by machine learning with the ratio and weight of each grade as teacher data, and the iron scrap from the image generated by the camera. Estimating the proportion of each grade contained in the set and the weight of the set of iron scrap,
The scrap grade judging system according to any one of claims 1 to 5. Further comprising a weight calculation unit for calculating the weight of each grade included in the set of iron scraps, based on the ratio of each grade included in the set of iron scraps and the weight of the set of iron scraps,
The scrap grade determination system according to claim 6. Further equipped with a crane that lifts some iron scrap from the piled iron scrap,
Every time the crane lifts a part of the iron scrap, the inference unit determines the proportion of each grade included in the part of the iron scrap and the weight of the part of the iron scrap from the image captured by the camera. presume,
The scrap grade judging system according to claim 6 or 7. A weight integration unit that integrates the estimated weight each time the crane lifts some iron scrap,
A weight measuring unit for measuring the weight of the entire stacked iron scraps,
A weight correction unit that corrects the weight accumulated by the weight accumulating unit based on the weight measured by the weight measuring unit;
Further comprising,
The scrap grade determination system according to claim 8. The camera photographs a set of iron scraps, in which a plurality of grades of iron scraps according to the size and iron scraps of different types are mixed,
The inference unit uses an image of a set of iron scrap as input data, and uses a learned model pre-constructed by machine learning with the ratio of each grade and the ratio of different varieties as teacher data, and from the image generated by the camera. Estimating the proportion of each grade and the proportion of different varieties contained in the set of iron scrap,
The scrap grade determination system according to any one of claims 1 to 9. The camera captures a set of iron scraps in which a plurality of grades of iron scraps according to the size and excluding products are mixed,
The inference unit uses an image of a set of iron scrap as input data, and uses a learned model pre-built by machine learning as a teacher data for the ratio of each grade and the presence or absence of excluded products, and from the image generated by the camera. Estimating the ratio of each grade included in the set of iron scrap and the presence or absence of excluded products,
An alarm output unit is further provided that outputs an alarm when it is estimated that there are excluded products.
The scrap grade judging system according to any one of claims 1 to 10. A crane that lifts some iron scrap from piled up iron scraps with hoisting equipment,
A processing unit that extracts, from the image generated by the camera, a partial image that includes at least a range below the suspending device, which is defined based on the range of the suspending device.
Further equipped with,
The inference unit estimates the proportion of each grade included in the set of iron scraps from the partial image,
The scrap grade judging system according to any one of claims 1 to 11. The processing means performs a mask process in which a display area is a range in which the display area gradually narrows as it goes downward from the hanger, and the other area is a non-display area in the partial image.
The scrap grade determination system according to claim 12. Further comprising a range estimation unit that estimates the range of the hanging device in the image generated by the camera,
The scrap grade determination system according to claim 12 or 13. The learned model is a partial image extracted from an image of a set of iron scraps lifted by a crane sling, defined as a range of the sling, and including at least a range below the sling. As input data, pre-built by machine learning,
The scrap grade judging system according to any one of claims 12 to 14. A set of iron scraps mixed with multiple grades of iron scraps according to the size is taken with a camera,
Using an image of a set of iron scraps as input data, using a learned model pre-built by machine learning with the ratio of each grade as teacher data, each grade included in the set of iron scraps from the image generated by the camera Estimate the proportion of
Output the estimation result,
Scrap grade judgment method. An acquisition unit that acquires an image obtained by shooting a set of iron scraps with a camera, in which iron scraps of multiple grades according to the size are mixed,
Using an image of a set of iron scraps as input data, using a learned model pre-built by machine learning with the ratio of each grade as teacher data, each grade included in the set of iron scraps from the image generated by the camera An inference unit that estimates the ratio of
An estimation device comprising: An acquisition unit that acquires an image obtained by capturing a set of iron scraps with a camera, in which a plurality of grades of iron scraps according to the size are mixed, and
Using an image of a set of iron scraps as input data, using a learned model pre-built by machine learning with the ratio of each grade as teacher data, each grade included in the set of iron scraps from the image generated by the camera An inference unit that estimates the ratio of
A program for operating a computer as a computer. Mixed with a plurality of grades of iron scraps according to size, an image obtained by capturing a set of iron scraps with a camera, and an acquisition unit for acquiring the ratio of each grade included in the set of iron scraps,
With the image as input data, the ratio of each grade as teacher data, a learning unit that constructs a learned model for estimating the ratio of each grade contained in the set of iron scrap from the image by machine learning,
A learning device. An image of a set of iron scrap by grade is also used as input data to build the trained model.
The learning device according to claim 19. The image is a partial image extracted from an image of the set of iron scraps lifted by the crane hoist and including at least a range below the hoist, defined based on the range of the hoist. ,
The learning device according to claim 19. The partial image is a partial image subjected to a masking process in which a range that gradually narrows downward from the hanging device is a display area and another range is a non-display area.
The learning device according to claim 21. A mixture of a plurality of grades of iron scrap according to the size, an image obtained by shooting a set of iron scraps by a camera, and an acquisition unit that acquires the ratio of each grade included in the set of iron scraps, and
A learning unit that constructs a learned model by machine learning for estimating the proportion of each grade contained in the set of iron scrap from the image, using the image as input data and the proportion of each grade as teacher data,
A program for operating a computer as a computer. Multiple grades of iron scrap depending on the size are mixed, an image obtained by shooting a set of iron scraps by a camera, and the ratio of each grade included in the set of iron scraps is acquired,
Using the image as input data and the ratio of each grade as teacher data, a learned model for estimating the ratio of each grade included in a set of iron scrap from the image is constructed by machine learning.
How to generate a trained model.

Images