Attribute-based structural damage identification by few-shot meta learning with inter-class knowledge transfer

Abstract

Image archives of multi-class structural damages can be collected by manual inspection and then used for structural damage identification. On one hand, conventional image-processing-based approaches rely on optimal designs of hand-crafted feature detectors and lack universal adaptability for various application cases; on the other hand, regular supervised learning techniques require complete damage types and sufficient training examples to establish a robust damage recognition model, which brings up a time-labor-consuming image collection process. To solve these problems, this study proposes a nested attribute-based few-shot meta learning paradigm for structural damage identification. First, an external few-shot meta learning module is established based on different classification tasks named as meta-batches to produce robust classifiers for new damage types, in which support and query subsets including partial damage types and a few examples are randomly sampled from the original image dataset. Second, an embedded internal attribute-based transfer learning model is trained by minimizing the l2-norm and angular losses of attribute representation vectors in an end-to-end manner, where damage attributes act as the common inter-class knowledge and are transferred from the source damage space of support set to the target damage space of query set. Finally, the proposed approach is validated on a real-world structural damage image dataset, which contains 1000 examples of 10 representative damage types in total. Results show the proposed approach produces an overall accuracy of 93.5% and an average area under the ROC curve of 0.96 for 10 damage types. The general equilibrium of average precision and recall indicates that the trained model is balanced to both positive and negative examples for each damage type. Compared with a regular supervised learning model by directly classifying input images with one-hot vector labels, the proposed approach generates higher accuracy and better robustness. Parameter study suggests the proposed paradigm enables to train a stable and reliable meta learning classification model that can perform well across a series of settings about the ratio between support and query subsets. Theoretical analysis is also performed to explain why meta learning surpasses regular supervised learning.