Abstract
Conventional bridge inspection is usually performed by experienced engineers, trying to detect and document damage patterns manually. By in-creased number of built Bridges, there is a growing interest in automated dam-age detection methods. Therefore, the field of autonomous bridge inspection with the application of machine learning techniques on UAV-taken images is gaining popularity. Due to recent technological advancement, a large number of datasets can be collected, with a high rate of productivity and accuracy, to train convolutional neural networks (CNNs) leading us to automated Structural health monitoring (SHM). In this paper, a case study is chosen to scan two times with almost one year as a time interval. In the first scanning, dataset was gathered to train four different CNNs. Then, the performance of CNNs was compared for the purpose of autonomous crack detection in the second round of scanning. Models evaluated on a number of performance metrics, namely- (i) accuracy, (ii) loss, (iii) computation time, (iv) model size, and (v) architectural depth. Finally, the performance of studied CNNs is discussed, which can lead researchers in the Transfer-Learning approach to generate a model for damage detection with a limited number of datasets prepared in the first turn of bridge inspection.