US-12620076-B2 - Scrap discrimination system and scrap discrimination method

US12620076B2US 12620076 B2US12620076 B2US 12620076B2US-12620076-B2

Abstract

A scrap discrimination system and a scrap discrimination method that can improve scrap discrimination technology are provided. A scrap discrimination system includes a scrap part extraction model ( 221 ) that extracts, based on a camera image, a scrap part located in a central portion included in the camera image with reference to a window ( 107 ) defined in advance in an image, a scrap discrimination model ( 222 ), generated by teacher data including training images, that sorts grades of scrap and a ratio of each grade from a scrap image extracted by the scrap part extraction model ( 221 ), and an output interface ( 24 ) that outputs information on the grades of scrap and the ratio of each grade as discriminated based on the scrap image using the scrap discrimination model ( 222 ).

Inventors

Futoshi Ogasawara
Yasuo Kishimoto
Yasuyoshi TAKAO

Assignees

JFE STEEL CORPORATION

Dates

Publication Date: 20260505
Application Date: 20210728
Priority Date: 20200814

Claims (9)

1 . A scrap discrimination system comprising: an acquisition interface configured to acquire a camera image that includes scrap; a scrap part extraction model configured to extract, based on the camera image, a scrap part located in a central portion included in the camera image with reference to a window defined in advance in the camera image; a scrap discrimination model, generated by teacher data including training images, configured to sort grades of scrap and a ratio of each grade from a scrap image extracted by the scrap part extraction model; and an output interface configured to output information on the grades of scrap and the ratio of each grade as discriminated based on the scrap image using the scrap discrimination model, wherein the window identifies a range occupying ¼ of an entire image based on a center of the camera image and extracts a scrap part by semantic segmentation starting from scrap located in the window, the semantic segmentation is a method of categorizing each pixel based on information of surrounding pixels with respect to the camera image, executed by the scrap part extraction model, and the semantic segmentation includes: a step of extracting a scrap part starting from the scrap in the window; and a step of treating scrap that does not start from the scrap in the window as background.
2 . The scrap discrimination system according to claim 1 , wherein the training images are images of a single grade of iron scrap, and when the scrap discrimination model is used to discriminate the grades of scrap included in the scrap image and the ratio of each grade, the ratio is discriminated based on an area ratio of each grade of scrap in the scrap image.
3 . The scrap discrimination system according to claim 1 , wherein the training images are images of mixed grade iron scrap.
4 . The scrap discrimination system according to claim 1 , wherein the scrap discrimination model comprises a first scrap discrimination model, generated by teacher data including first training images, configured to discriminate the grades of scrap included in the camera image and the ratio of each grade based on the scrap image; a second scrap discrimination model, generated by teacher data including second training images different from the first training images, configured to discriminate the grades of scrap included in the camera image and the ratio of each grade based on the scrap image; and a selection model configured to determine whether to use the first scrap discrimination model or the second scrap discrimination model based on the scrap image, and the output interface is configured to output information on the grades of scrap and the ratio of each grade discriminated based on the scrap image using a model selected by the selection model from between the first scrap discrimination model and the second scrap discrimination model.
5 . The scrap discrimination system according to claim 4 , wherein the first training images are images of a single grade of iron scrap, and when the first scrap discrimination model is used to discriminate the grades of scrap included in the scrap image and the ratio of each grade, the ratio is discriminated based on an area ratio of each grade of scrap in the scrap image.
6 . The scrap discrimination system according to claim 4 , wherein the second training images are images of mixed grade iron scrap.
7 . The scrap discrimination system according to claim 4 , wherein the first training images, the second training images, and the scrap image are normalized based on zoom information corresponding to each image.
8 . The scrap discrimination system according to claim 4 , wherein the first scrap discrimination model, the second scrap discrimination model, and/or the selection model is retrained based on the scrap image and the information outputted by the output interface.
9 . A scrap discrimination method using a scrap part extraction model configured to extract, based on a camera image that includes scrap, a scrap part located in a central portion included in the camera image with reference to a window defined in advance in the camera image and a scrap discrimination model, generated by teacher data including training images, configured to discriminate grades of scrap and a ratio of each grade, the scrap discrimination method comprising: acquiring a camera image that includes the scrap; extracting a scrap image based on the camera image using the scrap part extraction model; and outputting information on the grades of scrap and the ratio of each grade as discriminated based on the scrap image using the scrap discrimination model, wherein the window identifies a range occupying ¼ of an entire image based on a center of the camera image and extracts a scrap part by semantic segmentation starting from scrap located in the window, the semantic segmentation is a method of categorizing each pixel based on information of surrounding pixels with respect to the camera image, executed by the scrap part extraction model, and the semantic segmentation includes: a step of extracting a scrap part starting from the scrap in the window; and a step of treating scrap that does not start from the scrap in the window as background.

Description

TECHNICAL FIELD The present disclosure relates to a scrap discrimination system and a scrap discrimination method. BACKGROUND To effectively utilize resources, demand has increased in recent years for reuse of waste materials, such as scrap, as recyclable resources. Recyclable resources need to be discriminated in order to reuse waste materials. A processing method to discriminate waste materials from camera images without human labor has already been proposed. In this method, images of waste materials are manually inputted in advance, and a learned model constructed by machine learning with information on waste materials as teacher data is simultaneously used. For example, see Patent Literature (PTL) 1. CITATION LIST Patent Literature PTL 1: JP 2017-109197 APTL 2: JP 2020-95709 A SUMMARY Technical Problem However, the technology of PTL 1 targets demolished houses and waste materials, such as disaster debris, for discrimination. No consideration is made for how to efficiently discriminate scrap metal, for example. Iron scrap, for example, circulates on the market as a reusable resource related to iron and is recycled into iron using an electric heating furnace or the like. Conventionally, the grade of scrap metal is discriminated visually by workers at the iron scrap processing site. This is because scrap metal pieces after crushing are of various scales, and the shape of each scrap piece differs. It is thus necessary to visually inspect the entire piece to determine the grade, making it difficult to automate the process. On the other hand, visual discrimination by workers leads to a problem of inconsistent discrimination results due to the skill level of the workers. Aging workers and the need to secure personnel are also problematic. PTL 2 discloses technology for determining the weight grade of iron scrap. In this method, scrap is suspended by a magnetic crane, and images captured onsite are used to estimate the percentage of scrap grade for the part fished out. This estimation process is repeated multiple times to arrive ultimately at an overall determination. However, the technology of PTL 2 is limited to the case of measuring the scrap grade while fishing scrap out with a magnetic crane. In actual scrap processing, there are many cases where scrap is brought in without using a crane, and such a method is difficult to use in these cases. In addition, the method of PTL 2 requires a long time to determine all of the scrap, since determinations are made sequentially while the scrap is suspended with a magnet. In this way, technology for recognizing the overall shape of scrap from a captured image and for discriminating the grade of scrap has room for improvement. In view of these circumstances, it is an aim of the present disclosure to provide a scrap discrimination system and a scrap discrimination method that can improve scrap discrimination technology. Solution to Problem A scrap discrimination system according to an embodiment of the present disclosure includes: an acquisition interface configured to acquire a camera image that includes scrap;a scrap part extraction model configured to extract, based on the camera image, a scrap part located in a central portion included in the camera image with reference to a window defined in advance in an image;a scrap discrimination model, generated by teacher data including training images, configured to sort grades of scrap and a ratio of each grade from a scrap image extracted by the scrap part extraction model; andan output interface configured to output information on the grades of scrap and the ratio of each grade as discriminated based on the scrap image using the scrap discrimination model. A scrap discrimination method according to an embodiment of the present disclosure is a scrap discrimination method using a scrap part extraction model configured to extract, based on a camera image that includes scrap, a scrap part located in a central portion included in the camera image and a scrap discrimination model, generated by teacher data including training images, configured to discriminate grades of scrap and a ratio of each grade, the scrap discrimination method including: acquiring a camera image that includes the scrap;extracting a scrap image based on the camera image using the scrap part extraction model; andoutputting information on the grades of scrap and the ratio of each grade as discriminated based on the scrap image using the scrap discrimination model. A scrap discrimination system according to an embodiment of the present disclosure includes: an acquisition interface configured to acquire a camera image that includes scrap;a scrap part extraction model configured to extract, based on the camera image, a scrap part located in a central portion included in the camera image with reference to a window defined in advance in an image;a foreign object discrimination model, generated by teacher data including training images, configured to sort foreign