A physic-guided YOLO framework for pavement deformation distress detection

This research develops an automated methodology for detecting and assessing deformation pavement distress such as rutting and corrugation, using deep learning algorithms. Utilizing a dataset of road images, various YOLO algorithms were initially explored and YOLOv5 achieved the best results. The initial YOLOv5 model demonstrated strong performance in rutting detection, achieving a 96.3% mAP for high severity, and 81.9% mAP for low severity. Due to limited corrugation instances, a dedicated YOLOv5 model was developed using an auto-annotation method, which achieved a 66.5% mAP. To enhance prediction reliability, a Physics-Informed Neural Network (PINN) approach was integrated into the YOLO framework by introducing a physics-informed penalty into its loss function. The hybrid PINN-YOLO model statistically improved bounding box accuracy and reduced false positives for both rutting and corrugation, yielding superior mAP results compared to the baseline YOLOv5 models. These results highlight the substantial potential of hybrid PINN-YOLO, offering a robust and accurate solution for pavement distress evaluation.