Automatic Detection of Pavement Marking Defects in Road Inspection Images Using Deep Learning

Pavement markings serve to convey information to drivers, regulate driving behavior, and effectively mitigate traffic congestion and reduce accidents. Nonetheless, due to traffic exposure and temperature stress, pavement markings may develop defects to diverse degrees. Consequently, the inspection and maintenance of pavement markings has been paid high attention. Traditional manual detection methods prove time-consuming, subjective, and present security risks. Therefore, we employed four object detection models [You Only Look Once version 5 (YOLOv5), YOLOv7, faster region convolutional neural networks (Faster R-CNN) with visual geometry group laboratory (VGG), and Faster R-CNN with residual network (ResNet)] to achieve intelligent recognition of pavement marking defects through deep learning. Each model underwent 1,000 epochs of training and utilized 2,000 annotated road inspection images. Through data augmentation, module optimization, and anchor redesign, these models can locate pavement markings and classify their defects. The accuracy and efficiency of the model were evaluated by mean average precision (mAP) and frames per second. In addition, we introduced evaluation indicators that focused on defect types to assist in selecting models with high applicability in detecting markings. Among these models, the optimized Faster R-CNN with VGG as the backbone network has an mAP of 93.96% and can detect over 28 images per second, which meets the engineering requirements.