Semantic Segmentation of Cracks Using DeepLabv3+

Detecting cracks of structures is a crucial role in the structural health monitoring. Destructive techniques and non-destructive approaches have been widely used to evaluate the structural health. Recently, a deep learning-based approach is developed for noncontact inspections. The goal of this paper is to suggest an efficient backbone of Resnet family in terms of crack detections using DeepLabv3+ architecture for the structural health monitoring. Five kinds of backbones, namely Resnet-18, Restnet-34, Resnet-50, Resnet-101, and Resnet-152 were implemented in this study. Adaptive moment estimation (Adam) optimizer and dice loss function were applied to train the models. In addition, the mean intersection over union (IoU) was employed to investigate the accuracy of proposed models. The study results show that all backbones effectively detected the concrete cracks with over 90% IoU. The Resnet-50 presents the best performance of 93.5% IoU for DeepLabv3+ architecture. The findings highlighted the feasibility of proposed method in terms of structural crack detections.