Efficient semi-supervised surface crack segmentation with small datasets based on consistency regularisation and pseudo-labelling

Despite promising results in vision-based surface crack detection, data-driven approaches still suffer from the scarcity of rich labelled datasets. Such a limitation has hindered a wider practical application of detection models. To address this issue, a semi-supervised framework is proposed, capable of learning from a substantial amount of unlabelled data and achieving high accuracy, even when the available labelled datasets are of limited size. The framework is designed by tailoring supervised training, semi-supervised consistency regularisation, and self-training with certainty-based pseudo-labelling, resulting in a simple yet effective approach. Despite using only 2% of the total labelled Concrete and Asphalt datasets, the resulting mIoU was only 2.6% and 4.7%, respectively, lower than the best performance of the model trained on 100% of the labelled data. Remarkably, the designed framework assisted the model in approaching and even exceeding saturation levels with as little as 20% and 25% of the Concrete and Asphalt datasets.