Semantic segmentation with small training datasets: A case study for corrosion detection on the surface of industrial objects

In this research, we investigate possibilities to train convolutional neural networks with a small dataset for semantic segmentation, while achieving the best possible model generalization. In particular, we want to segment corrosion on the surface of industrial objects. In order to achieve model generalization, we utilize a selection of established and advanced strategies, i.e. Self-Supervised-Learning. Besides radiometric- and geometric-based data augmentation, we focus on model complexity regarding encoder and decoder, as well as optimal pretraining. Finally, we evaluate the best performing model against a pixel-wise random forest classification. As a result, we achieve an f1-score of 0.79 for the best performing model regarding the segmentation of corrosion.