Mobile Vision Transformer for Surface Defect Classification from A Tiny Dataset

Surface defect classification is one pivotal step during the last product inspection before selling it to the market. Indeed, we want to serve our product to the customer in good condition. Speaking of smart manufacturing, existing studies on hot-rolled steel strip surface defect classification used deep learning models, which are very costly in terms of hardware, parameters, and dataset size. In this work, we propose a finetuned lightweight model from the Mobile Vision Transformer (MobileViT) to efficiently classify the surface defect of hot-rolled steel strips from a tiny data set. In this study, we experimented on a public benchmark dataset for surface defects of hot-rolled steel strip classification. The results showed that the proposed model can predict the defect accurately with fewer parameters and a smaller dataset than the SOTA ones.