Self-supervised Learning for Robust Surface Defect Detection

In this study, we discuss about the use of Self-Supervised Learning to improve robustness of Surface Defect Detection (SDD) models. We show how different state-of-the-art SDD methods are already implementing some sort of self-supervision in their learning procedure, and we discuss how more advanced techniques inspired to Confident Learning can be used in a generic pipeline. We also propose One-Shot Removal strategy, a baseline approach that can be applied to any SDD model to improve its robustness. Our method employs a three-step training pipeline: initial training on the entire dataset, followed by removal of anomalous samples, and fine-tuning on the refined dataset. Experiments conducted on the challenging Kolektor SDD2 dataset show how this process enhances the representation of 'normal' data and mitigates overfitting risks.