Towards robust ferrous scrap material classification with deep learning and conformal prediction

The classification of ferrous scrap materials is a well-explored problem in the literature, recognized for its significance in the steel production industry. While deep learning models are effective for this task, their deployment in industrial settings requires addressing model uncertainties and ensuring proper calibration. This study proposes adapting split conformal prediction to quantify uncertainties and facilitate model calibration. The results indicate that the Hierarchical Vision Transformer using Shifted Windows (Swin) models, particularly Swin V2, serves as the most reliable backbone for this task. Although the performance of Swin models is comparable to other evaluated models, Swin V2 demonstrates superior confidence, achieving 95.51% accuracy and the lowest conformal prediction threshold. The method is rigorously evaluated on a real-world dataset comprising 8,147 images across nine classes of ferrous scrap widely used in the Brazilian steel industry. Explainability methods corroborate the results of conformal prediction, enhancing transparency and trust in model predictions, and thereby facilitating industrial adoption. This approach bridges the gap between advanced deep learning and practical application in ferrous scrap classification, underscoring the importance of model calibration in industrial deployment.