CBAM-Optimized Automatic Segmentation and Reconstruction System for Monocular Images With Asphalt Pavement Potholes

Rapid and accurate three-dimensional (3D) detection of asphalt pavement pothole damage is crucial for pavement performance and quality evaluation. However, existing 3D reconstruction methods based on lasers and depth cameras are expensive and lack operability. In addition, ‘false potholes’ are formed through gaps between the asphalt aggregate particles and they interfere with the reconstruction of real potholes. In order to solve the above problems, a new CBAM (convolutional block attention module) optimized pothole segmentation and reconstruction system for monocular images is proposed. First, a network called CBAM-Seg-CapsNet is developed to accurately segment pothole areas from two-dimensional monocular images of pavement. The influence of ‘false potholes’ on the reconstruction of real potholes is avoided. Secondly, an unsupervised monomural depth estimation intelligent network, called ‘CBAM-Recon-Depth’, is developed to realize effective reconstruction of the segmented pothole area. The expensive problem of 3D reconstruction methods, based on lasers and depth cameras, is solved. Compared with some famous segmentation and reconstruction models, the segmentation and reconstruction accuracy of the proposed system are 96.62% and 91.32%, respectively, and the F1-score is 93.89% and 90.76%, respectively. The Dice value of segmentation is 92.27%. The reconstructed RMES Log and Abs Rel are as low as 0.1436 and 0.1252, respectively. On-site experiments are carried out and the reconstruction rate reached 96.05%. The results show that the CBAM optimization system can accurately extract the pothole area and realize the high-precision reconstruction of pothole damage after extraction.