NSLNet: An improved deep learning model for steel surface defect classification utilizing small training datasets

Abstract

Manufacturing industries contemplate integrating computer vision and artificial intelligence into shop floor operations, such as steel surface defect identification, to realize smart manufacturing goals. However, inadequate annotated training datasets and reduced prediction abilities with image perturbations restrict the practical implementation. This paper introduces NSLNet framework utilizing ImageNet as a feature-extractor combined with adversarial training in the extracted feature space through Neural Structure Learning to address these barriers. The experiments on public (NEU) and synthetically generated datasets (ENEU) showed that the NSLNet could learn with few training samples maintaining resilience against image perturbations outperforming conventional models significantly and nearest deep learning competitors marginally.