The resilience of civil infrastructures decreases the risk of collapse and guarantees the durability of structures under permanent loads. The maintenance, management, and inspection optimize and improve structural integrity and resilience, ensuring public safety and uninterrupted operation. Crack detection and its three-dimensional (3D) reconstruction in the early steps are essential for the long-term sustainability of infrastructure. In this study, a novel low-cost and time-efficient method based on deep learning and smartphone images is proposed for crack segmentation and 3D reconstruction. Data augmentation, Squeeze-and-Excitation (S & E) channel attention module, and Atrous Spatial Pyramid Pooling (ASPP) feature fusion module are added to DeepLabv3+ to increase the accuracy and generalization of the hybrid method. The patches are classified with an overall accuracy of 98% and an Intersection over Union (IoU) of 81%. Considering our custom dataset, the proposed hybrid method improved the precision by 2% and raised it to 92%. Next, segmented images are imported into the Structure-from-Motion (SfM) procedure to generate a sparse point cloud. Information of ground control points that are measured during a land survey is considered to solve the 3D transformation, and to perform accuracy evaluation. Less than 1 pixel in re-projection error and 1.3mm in the reconstruction error are achieved. In addition, the reconstruction error of 620 µm is achieved, which is validated with a check scale bar. In the end, the ground truth validation results demonstrate that the width and length of cracks are measurable with the accuracy of two-tenths of a millimeter.
Majidi et al. (Mon,) studied this question.