The integration of Machine Learning (ML) techniques, particularly Deep Learning (DL) models, within Structural Health Monitoring (SHM) has expanded exponentially in the past couple of decades. Though significant advancements have been made concerning the adaptability of these technologies to SHM, DL-driven methods still face current, long-standing challenges within this domain. The availability and size of existing training datasets for SHM applications remain relatively restrictive creating difficulties in training new DL models which require a significant number of samples. Furthermore, these training datasets remain largely unbalanced, with the majority class belonging to undamaged structures, resulting in models that are biased toward the dominant class of damaged structures. As such, to address the data scarcity, and class imbalance issues of the SHM domain, few-shot learning (FSL) models have been explored in this study. Contrasting traditional ML methods, these models attempt to optimize the data versus accuracy problem by developing models that have significantly high accuracy while using a significantly smaller training sample size. In this paper, a prototypical network is proposed for the classification of structural damages of concrete and asphalt surfaces using limited data. The performance of each model is explored for low-sample environments, including 1, 2, 5, 10, and 20 images per class. Additionally, the performance of this FSL model is explored for different image transformation techniques, including histogram equalization, logarithmic transform, power transform, and phase stretch transform. Finally, the impact of the aforementioned image transformations is investigated to reduce the overfitting of inter-material datasets (datasets originating from different material types).
