Pixel-Level Recognition of Pavement Distresses Based on U-Net

This study develops and tests an automatic pixel-level image recognition model to reduce the amount of manual labor required collect data for road maintenance. Firstly, images six kinds pavement distresses, namely, transverse cracks, longitudinal alligator block potholes, patches, are collected from four asphalt highways in three provinces China build a labeled dataset containing 10,097 images. Secondly, U-net model, one most advanced deep neural networks segmentation, is combined with ResNet network as basic classification recognize distressed areas Data augmentation, batch normalization, momentum, transfer learning, discriminative learning rates used train model. Thirdly, trained models validated on test dataset, results experiments show following: if types distresses not distinguished, pixel accuracy (PA) values using ResNet-34 ResNet-50 97.336% 95.772%, respectively, validation set. When PA two 66.103% 44.953%, respectively. For ResNet-34, category (CPA) intersection over union (IoU) identification no distress 99.276% 99.059%, featuring images, CPA IoU highest at 82.774% 73.778%, lowest 14.077% 12.581%,