Advancing Robust Few-shot Surface Defect Detection through Meta-learning

The primary limitation of deep learning (DL) techniques in detecting manufacturing defects is the requirement for a substantial volume of labeled data. Nevertheless, obtaining precise annotations is an arduous process due to the diverse range of defect categories. To tackle this issue, we propose MetaFormer, a transformer-based few-shot surface defect detection that learns the meta-features from a large-scale dataset of base classes. This method enables generalization to a novel class with only a few training examples. We demonstrate the effectiveness of our model on the steel surface defect dataset, surpassing the performance of the baseline model despite using only a limited number of annotated examples.