Comparative Analysis of Deep Learning Models for Road Damage Classification

Abstract

Road damage, particularly cracks and potholes, poses significant risks to transportation safety and infrastructure sustainability. Traditional manual inspection methods are time-consuming, costly, and often unreliable, creating the need for automated solutions. This study proposes a deep learning-based approach for road damage classification using a dataset of 1,026 Bangladeshi road images, consisting of 490 crack and 536 pothole samples. Six pretrained convolutional neural network (CNN) models—MobileNetV2, ResNet50, InceptionV3, DenseNet121, VGG16, and VGG19—were implemented and evaluated. All models were trained for 15 epochs with a learning rate of 0.001. Among them, MobileNetV2 achieved the highest performance with an accuracy of 90%, precision of 0.91, recall of 0.90, and F1-score of 0.90, demonstrating its suitability for resource-efficient real-world deployment. The proposed system offers an effective solution for timely road maintenance, contributing to improved road safety, reduced vehicle damage, and sustainable infrastructure management.