Steel manufacturing requires high-throughput and high-reliability surface inspection to minimize safety risks, scrap rates, and downstream quality reductions. Conventional rule-based vision and manual inspection are often impeded in real production environments by variable illumination, complex textures, subtle defect morphology, and stringent latency constraints imposed by production-line operation. Deep learning (DL) has become a dominant paradigm for the detection and classification of defects when inspecting steel, but many previous studies have performed broad architectural overviews without explicitly connecting model and pipeline choices to deployment-critical factors such as processing speed, hardware availability, annotation cost, and robustness during domain shift. This review synthesizes ten representative case studies on DL-based defect inspection in steel manufacturing and closely related industrial settings, spanning classification, object detection, and segmentation workflows. To enable structured comparisons, we harmonized practical considerations across the ten studies, including task formulation, backbone design, training strategy (e.g., transfer learning), data augmentation, reported throughput, and commonly used performance indicators such as accuracy, precision/recall, mean average precision, and processing speed. Due to the heterogeneity in datasets, metrics, and hardware configurations across studies, we further introduce a transparent, review-oriented figure of merit as a heuristic summary of reported benefit–cost parameters (performance accuracy and processing speed vs. training burden and model complexity), while explicitly addressing the limitations associated with missing values and avoiding claims of statistically definitive ranking. Based on recurring patterns across the selected studies, we propose a conceptual hybrid framework blueprint—derived from the reviewed literature rather than newly experimentally validated results—that integrates transfer learning, real-time detection, and end-to-end learning principles as an engineering template for practical deployment. We conclude by providing actionable guidance for applications and outline future directions in label-efficient learning, cross-domain robustness, and standardized benchmarking and reporting to improve the reproducibility and industrial relevance of our approach.
Noé et al. (Thu,) studied this question.