Published in Eleventh U.S. National Conference on Earthquake Engineering Proceedings, June 25, 2018.
A significant task in earthquake reconnaissance is to conduct rapid and accurate assessments of damage to built infrastructure. This can be accomplished, in part, by analyzing the large volumes of high-resolution image data collected after a seismic event. However, detailed image tagging remains a task for trained human volunteers, which is both time-intensive and error prone. The authors developed a software tool to simplify and standardize the process of assigning damage and structure pairs to sub-regions of images. The goal of the tool is to facilitate the tagging of thousands of images from historic and recent earthquakes to train a deep learning (DL) algorithm to automatically identify damage observed in civil infrastructure. DL is a subset of machine learning that can be used for image classification problems. This process requires thousands of expertly tagged images for robust and automatic visual recognition capabilities. In detecting specific structural damage after an earthquake, images must have explicit tags for the building material, damage and location, as well as the impacted structural members. To obtain such a descriptive set of images, there is a need for a task-specific tool that facilitated tagging of the most common post-earthquake structural damage types. The resulting software solution consists of a simple user interface that displays the most frequently used damage and structural member tags as pre-loaded radio buttons and includes the flexibility for users to customize tags when necessary. The program generates marked-up images that show location-specific damage and structural member labels, as well as output files in the PASCAL Visual Object Classes (VOC) format that are compatible with TensorFlow and most DL frameworks, such that tagged images are ready to be used for training a DL algorithm.
Copyright © YEAR Earthquake Engineering Research Institute.
Number of Pages