Published in Eleventh U.S. National Conference on Earthquake Engineering, June 25, 2018.
The amount of structural damage image data produced in the aftermath of an earthquake can be staggering. It is challenging for a few human volunteers to efficiently filter and tag these images with meaningful damage information. The proposed solution is to automate post-earthquake reconnaissance image tagging activities by training a computer algorithm to classify each occurrence of damage per building material and structural member type. The approach is based on deep learning (DL), a subset of machine learning loosely based on the operation of a biologic neural system, which aims to learn and extract accurate representations from large data sets. DL algorithms are data driven; improving with increased training data. Thanks to the vast amount of data available and advances in computer architectures, DL has become one of the most popular image classification algorithms producing results comparable to and in some cases superior to human experts. The authors implemented a DL algorithm to automatically identify multiple damage types and associated structural members in a single image by adapting a pre-trained deep residual network. The algorithm was tested as follows: (i) binning building images as damage-no damage (88% accuracy), (ii) drawing a bounding box around damage in buildings (85% accuracy) and short/captive reinforced concrete columns with shear damage (77% accuracy). The lower accuracy of correctly identifying a target region in an image (test ii) compared to simple binning (test i) is anticipated since it is a more complex problem and there is a more limited number of expertly tagged training images (200 count) for shear damage-short column condition being studied. The research team expects algorithm accuracy will improve with training on additional images tagged for certain damage-structure pairs by a diverse set of experts.
Copyright © YEAR Earthquake Engineering Research Institute.
Number of Pages