Gravel road classification based on loose gravel using transfer learning

Abstract

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