Dynamic Adaptation of Classification and Segmentation Models for Evolving Steel Scrap Classes Using Contrastive Learning

Abstract

With the rise of Electric Arc Furnaces (EAFs) in steel production, the industry is shifting toward more sustainable and flexible recycling-based steel production processes. In this context, automated classification and segmentation of steel scrap are essential, but can be challenging, due to class definitions changing over time, or new scrap types being introduced. This thesis explores whether Contrastive Learning (CL) can help Computer Vision (CV) systems adapt to such evolving class structures. A multi-stage segmentation pipeline is developed and evaluated, comparing a CL-based approach using the SelfOrganizing Features (SOF) loss against a standard non-contrastive baseline trained with Categorical Cross-Entropy (CCE) loss. To simulate dynamic class scenarios, one scrap class is excluded during initial training and only introduced later during fine-tuning. While both models reach similar segmentation performance after the last training stage, the contrastive model achieves noticeably better results after the first segmentation training stage. This indicates that CL can enable more robust and generalizable feature learning in early phases, where only partial class knowledge is available. These findings highlight the potential of CL as a foundation for adaptive, future-proof CV systems in dynamic industrial environments such as EAF-based steel recycling.

Date of Award	2025
Original language	English
Supervisor	Stephan Dreiseitl (Supervisor) & Raphael Pisoni (Supervisor)