An Improved UNet++ Architecture for Road Crack Segmentation, Image-Driven Characterization, and Severity-Focused Classification

Automated road-crack analysis is essential for pavement condition assessment, yet many deep learning methods focus primarily on binary segmentation and provide limited information on crack characteristics and severity. This study proposes a modular, multi-stage UNet++ framework for crack segmentation, image-based characterization, and severity-oriented classification. Three independently trained stages were developed: a baseline binary segmentation model, an enhanced model incorporating domain-oriented preprocessing and geometric characterization, and a four-class model distinguishing background/intact pavement and three image-derived severity categories. Experiments were conducted on 2,350 asphalt pavement images, divided into 70% training, 10% validation, and 20% testing sets. The three stages achieved 98.98%, 99.06%, and 99.09% accuracy, with IoU values of 0.86, 0.89, and 0.91, respectively. The framework also provided pixel-based crack-size descriptors. These results demonstrate the potential of combining UNet++ segmentation with domain-oriented preprocessing and image-derived characterization for more informative crack analysis. However, the reported dimensions are image-based rather than physically calibrated, and severity categories are derived from geometric thresholds without independent engineering annotations.