An attribute-based few-shot learning framework for structural damage classification

Performances of learning algorithms for structural damage identification in the complicated real-world situations are faced with significant barriers when training a classifier from only a handful of labeled examples. Real structural local damages always contain a variety of different classes in view of material categories, surface roughness situations and structural components where they emerge. However, only limited image collections are formed and annotated with manual labels. Meta-learning can obtain a parameterized model by training with a small labeled training set and its corresponding test set. Inspired by the perceptions of object attributes transferring and few-shot learning, this study proposed an attribute-based few-shot learning approach for structural damage identification with knowledge transfer. The proposed method is expected to achieve good classification performances by few-shot meta learning and avoid underfitting encountered in the conventional supervised learning using only a few training samples. Moreover, the proposed method exploits an attribute-based transfer learning procedure based on prior knowledge extraction from source categories. The proposed framework consists of two loops, which denotes the external few-shot learning paradigm and the internal attribute-based regression NLP model, respectively. Model updating procedures are then established by using samples in support and query sets and iterations of meta-batches.