Learning to detect important visual changes for structural inspections using physicsbased graphics models

Periodic structural inspections of large, difficult-To-Access infrastructure like dams and bridges are often time-consuming, laborious and unsafe. In the united states, state and federal agencies responsible for managing such infrastructure assets have recently begun investigating the use of unmanned aerial vehicles (UAV) to allow for remote data acquisition. However, processing the large amounts of data acquired by the UAV is very challenging. One method to reduce the amount of information for inspectors to evaluate, and isolate regions of interest for structural inspections, is to identify changes from a baseline state. While the identification of changes can be beneficial, not all changes have structural significance and changes can represent a myriad of environmental variations such as lighting changes or growth of vegetation that have little or no impact on the overall health of a structure. We propose the use of deep semantic segmentation to identify important changes on structures. A challenge with the use deep semantic segmentation is the requirements for large sets of training data for the algorithm to perform well. To overcome this, in we use physics-based graphics models to generate synthetic data. Graphics model for defects such as cracks and corrosion are included, while accommodating environmental variations such lighting, vegetation growth, and dirt. The proposed methodology is tested on a virtual environment of inland navigation infrastructure including miter gates and tainter gate dams.