South KoreaChemicals and materials2020 – ongoingChosun Refractories
AI inspection of X-ray images on the KAMP platform
A refractory maker cut X-ray inspection from 1.5 to 0.5 minutes per product with a model built on the state AI platform
Problem
Chosun Refractories, a maker of refractory products founded in 1947, checked finished products for internal cracks and foreign matter by having workers look at X-ray images. As product shapes multiplied and inspectors aged, judgments became less reliable, and extra staff added to fatigue relief raised manufacturing costs.
Approach
Through the Korea AI Manufacturing Platform (KAMP), run by the Ministry of SMEs and Startups, the company turned X-ray video of finished products into still images and labelled each as normal or defective. AI and process experts on the platform helped build a model that classifies new images. The case was presented when KAMP opened fully in January 2021.
Results
Takeaway
The company cut time on an inspection step that already produced images with clear normal/defective labels. Without its own data team, the mid-sized manufacturer used a public platform's storage, computing and expert help without hiring staff.
Categories
- Industry
- Chemicals and materials
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Computer vision
- Adoption
- Government support program
- Evidence
- Company's own claim
Sources
Source summaries
Edaily reported on Chosun Refractories when the Ministry of SMEs and Startups fully launched its AI manufacturing platform. Workers had inspected X-ray images by eye, and the company labelled X-ray frames to train a classification model. According to the ministry, judgment reliability rose from 90% to 96% and inspection time fell from 1.5 to 0.5 minutes.
Chosun Refractories had inspected cracks and foreign matter manually, but faced more product shapes and ageing inspectors. FA Journal reported that judgment reliability rose from 90% to 96% and inspection time fell from 1.5 to 0.5 minutes. It also said the model handled previously unseen defect patterns better.
Full details are in the original sources.