Image-Based Road Damage Detection Using Deep Learning Models

Abstract

Road infrastructure deteriorates continuously under traffic load and weather, and undetected damage such as potholes, surface cracks, and uneven asphalt increases accident risk and long-term repair cost. Traditional road surveys rely on manual visual inspection, which is slow, subjective, and difficult to scale across large networks. This study develops an image-based road damage detection system built on deep learning models that automatically locate and classify damage from road surface images. A curated dataset of 160 annotated road images covering four categories—Normal Road, Potholes, Road Cracks, and Severe Damage—was preprocessed through resizing and normalization and expanded with data augmentation including flipping, rotation, brightness shift, and noise injection. Three architectures were trained and compared: a baseline Convolutional Neural Network (CNN), a lightweight MobileNet classifier, and a YOLOv8 object detector. Models were evaluated using accuracy, precision, recall, F1-score, and mean Average Precision (mAP). YOLOv8 achieved the strongest results with 94.60% accuracy, 94.00% F1-score, and 92.80% mAP, outperforming MobileNet and the baseline CNN while still supporting near real-time inference. A browser-based application was implemented to upload road images, preview them, run a detection simulation, and display class labels, confidence scores, and bounding boxes over pothole regions. A moving-vehicle simulation further demonstrates real-time risk reporting, raising the warnings “Risk Detected,” “Vehicle Passing Damaged Road,” and “Maintenance Required” when the vehicle crosses a damaged zone. The contributions are a comparative deep learning study for road damage detection, an interactive detection-and-simulation prototype, and a practical workflow for automated road condition monitoring