Modeling automatic pavement crack object detection and pixel-level segmentation

Timely pavement crack detection can prevent further pavement deterioration. However, obtaining sufficient quantities of crack information at low cost remains a challenge. This study therefore proposed a lightweight pavement crack-detection model to realize the dual tasks of object detection and semantic segmentation. First, the modified YOLOv4-Tiny model was used to predict the bounding box wrapping cracks, and the threshold for segmentation was proposed. Moreover, an attention feature pyramid network was proposed to compensate for the loss of accuracy owing to the reduction in model parameters and structure scaling. The denoising auto-encoder network was provided to remove any background noise that could be recognized as cracks in the segmentation mask. The final number of model parameters was 6.33 M. The performance of the proposed model was compared with that of conventional models, indicating approximately equivalent evaluation index values even though four to five times fewer parameters were included than in the conventional models.