DOES - A multimodal dataset for supervised and unsupervised analysis of steel scrap

Area of Science:

• Materials Science
• Computer Science
• Industrial Engineering

Background:

• Scrap is a vital industrial raw material, with its importance set to increase due to the green steel transition.
• Visual inspection of scrap is critical, but its appearance changes due to oxidation, posing challenges for automated systems.
• Existing scrap datasets often lack order, hindering the training of machine learning models.

Purpose of the Study:

• To introduce the Dataset of European scrap classes (DOES), a free resource for non-alloyed European scrap.
• To address the challenges of scrap oxidation and lack of intrinsic order in images for AI training.
• To facilitate the development of industrial applications and research in scrap classification algorithms.

Main Methods:

• The DOES dataset was curated to include scrap images exhibiting varying degrees of corrosion.
• A novel technique was employed to extract numerous overlapping rectangular patches from raw scrap images.
• The dataset was validated using expert assessments and machine learning model performance.

Main Results:

• DOES provides a comprehensive collection of European scrap images, accounting for oxidation effects.
• The image processing technique enables effective training of deep learning algorithms on unordered scrap data.
• Validation confirmed the dataset's suitability for developing and testing scrap classification models.

Conclusions:

• The DOES dataset is a valuable, free resource for advancing AI in scrap material classification.
• It supports research and development for industrial applications in the context of sustainable steel production.
• The dataset's design effectively mitigates common challenges in scrap image analysis.