Transfer Learning with ResNet-18 for Surface Defect Inspection in Hot-Rolled Steel

Surface inspection in hot-rolled steel manufacturing requires decisions about material moving through a production line. A model that distinguishes photographed defects must also handle normal surfaces, changing acquisition conditions and uncertain predictions before its output can support coil release. This paper develops a research framework for adapting ResNet-18 to this setting with limited factory annotations. The proposed pipeline combines line-scan image acquisition, patch classification and a decision policy that assigns cases to pass, human review or hold. The NEU surface defect database supplies source-domain examples, factory images supply a normal class and the target-domain evidence required for deployment. The proposed experiment compares training budgets of 10, 25 and 50 images per class, with separate coils allocated to training, validation and testing. Comparisons distinguish six-class defect recognition from seven-class factory inspection, preventing misleading comparisons between models with different label spaces. Evaluation covers class performance, defect escapes, unnecessary holds, review workload and processing deadlines. The analysis identifies the conditions under which the experiment could support deployment, particularly representative normal data and independent validation of release decisions. No model training or factory trial is reported. The contribution is an experimentally testable inspection framework connecting transfer learning to operational quality control.