Gravel road classification based on loose gravel using transfer learning

Road maintenance agencies subjectively assess loose gravel as one of the parameters for determining road conditions. This study aims to evaluate performance deep learning-based pre-trained networks in rating images according classical methods done by human experts. The dataset consists roads extracted from self-recorded videos and Google Street View. were labelled manually, referring standard ground truth defined Maintenance Agency Sweden (Trafikverket). was then partitioned a ratio 60:40 training testing. Various models computer vision tasks, namely Resnet18, Resnet50, Alexnet, DenseNet121, DenseNet201, VGG-16, used present study. last few layers these replaced accommodate new image categories our application. All performed well, with an accuracy over 92%. results reveal that VGG-16 transfer learning exhibited best terms F1-score compared other proposed models.