Pavement Surface Type Classification Based on Deep Learning to the Automatic Pavement Evaluation System

Abstract

. Computer vision techniques, image processing, and machine learning became incorporated into an automatic pavement evaluation system with technological advances. However, in most research, the models developed to identify defects in the pavement assume that all the segments evaluated are paved and with one specific pavement surface type. Nevertheless, there is a wide variety of road surface types, especially in urban areas. The present work developed models based on a deep convolutional neural network to identify the pavement surface types considering five classes: asphalt, concrete, interlocking, cobblestone, and unpaved. Models based on ResNet50 architectures were developed; also, the Learning Rate (LR) optimization “one-cycle” training technique was applied. The models were trained using almost 50 thousand images from Brazil’s states highway dataset. model results are excellent, highlighting the model based on ResNet50, in which it obtained accuracy, precision, and recall values of almost 100%.