2021 Volume 2 Issue 2 Pages 46-53
This research aims to compare the capabilities on pixel-level damage detection between semantic segmentation and instance segmentation, mainly in terms of speed and accuracy. Thus, the typical networks of these two segmentation methods, Fully Convolutional Network (FCN) and Mask R-CNN are trained and tested. In this research, 300 high resolution images of corrosions on steel bridges are collected and labeled to train and evaluate several FCN and Mask R-CNN models. To compare these two Deep Learning methods, the time of training and predicting processes are recorded and the predicted results are processed and calculated through different evaluation methods. Besides, one UAV-taken 4K image of severely corroded bridge is adopted to test the capability of the Deep Learning models under more complex environmental conditions.