Data Augmentation for Improving Deep Learning Models in Building Inspections or Postdisaster Evaluation

Abstract

Current building evaluations, whether for occupant safety or insurance appraisal, are conducted primarily via visual inspections performed by certified individuals. These inspections, which often can number in the hundreds of thousands when performed following a disaster, can take weeks to conduct. This time can significantly affect the economic and societal resilience of a community. This paper proposes a framework for the development of unmanned aerial systems (UAS)-driven object detection algorithms for use in automating visual structural inspections. In this framework, domain-specific data augmentation methods are developed and utilized by image-based deep learning models for building inspections. A large, labeled, posthailstorm building evaluation database was developed to train and validate these models. Three data augmentation methods were developed and implemented: background cropping, high-resolution image cropping, and vent cropping. A unique combination of algorithm, novel data augmentations, and ensembling techniques was investigated to increase the performance of the framework. The results demonstrated that the framework can be applied to structural inspections to increase the efficiency and reliability of these assessments while minimizing the risk to human life.