Subsurface damage detection of a steel bridge using deep learning and uncooled micro-bolometer

A new deep learning-based method is proposed to detect subsurface damage of steel members in a steel truss bridge using infrared thermography (IRT). To reduce computation costs, the original deep inception neural network (DINN) is modified for transfer learning. The proposed method provides bounding boxes for detecting and localizing subsurface damage such as corrosion and debonding between paint with coating and steel surface. Robustness and accuracy were tested on 200 thermal images (640 x 480 pixels), and 96% accuracy and 97.79% specificity was achieved. The results were validated with ultrasonic pulse velocity (UPV) tests. (C) 2019 Elsevier Ltd. All rights reserved.