A Review of Steel Surface Defect Detection for Industrial Applications: Evaluation Frameworks, Data Challenges, and Model Mechanisms

Steel surface defect detection is a critical component of industrial quality inspection based on artificial intelligence (AI). In recent years, deep learning has substantially improved the level of automation in this task. However, model performance in real production lines remains jointly affected by evaluation metrics, data organization strategies, inference and post-processing settings, and deployment conditions, resulting in limited comparability across studies, poor reusability of conclusions, and difficulties in engineering transfer. Focusing on AI-based industrial inspection scenarios, this paper presents a problem-driven narrative review of steel surface defect detection research. It provides a systematic analysis centered on industrial-grade evaluation metrics, public datasets and data risks, key model mechanisms, and their industrial applicability. Particular attention is paid to recall evaluation under fixed false-positive constraints, tiling- and batch-level leakage, long-tail distributions and annotation noise, as well as issues related to input fidelity, representation under complex backgrounds, end-stage decision-making, and lightweight deployment. This review aims to provide a problem-oriented reference framework for result interpretation, method comparison, and industrial deployment evaluation in steel surface defect detection research.