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

• DOI: 10.1177/1475921720921135
• Publication Date: 2020-05-27T12:04:49Z
• Authors: Yang Xu; Yuequan Bao; Yufeng Zhang; Hui Li
• Keywords: Identification; Transfer of learning; Feature vector; Feature (linguistics)

Image archives of multi-class structural damages can be collected by manual inspection and then used for damage identification. On one hand, conventional image-processing-based approaches rely on optimal designs hand-crafted feature detectors lack universal adaptability various application cases; the other regular supervised learning techniques require complete types sufficient training examples to establish a robust recognition model, which brings up time-labor-consuming image collection process. To solve these problems, this study proposes nested attribute-based few-shot meta paradigm First, an external module is established based different classification tasks named as meta-batches produce classifiers new types, in support query subsets including partial few are randomly sampled from original dataset. Second, embedded internal transfer model trained minimizing l 2 -norm angular losses attribute representation vectors end-to-end manner, where attributes act common inter-class knowledge transferred source space set target set. Finally, proposed approach validated real-world dataset, contains 1000 10 representative total. Results show produces overall accuracy 93.5% average area under ROC curve 0.96 types. The general equilibrium precision recall indicates that balanced both positive negative each type. Compared with directly classifying input images one-hot vector labels, generates higher better robustness. Parameter suggests enables train stable reliable perform well across series settings about ratio between subsets. Theoretical analysis also performed explain why surpasses learning.