Pages: 179-187
Ahmad Khalid Hussain , Fati Oiza Ochepa
This study presents a deep learning solution for detecting road anomalies via a hybrid architecture consisting of a Masked Autoencoder (MAE) and a Vision Transformer (ViT) model. It presented a framework for dual road classification, namely an intact road (good) and defected road (bad) where defected roads are characterized by anomalies such as potholes or cracks. The target road anomaly classification model was trained and tested using publicly available datasets of road condition images. The model demonstrated good feature extraction as well as good generalization with a training accuracy of 99.79% and a test accuracy of 98.29%. Furthermore, we integrated the road anomaly detection model into a web-application providing real-time road anomaly detection, exemplifying the possible benefits of applying computer vision and machine learning algorithms to improve road maintenance in Nigeria.
Anomaly detection, Deep learning, Image processing, Masked autoencoders, Transport safety, Vision Transformers
© kvscsjournal.org . All Rights Reserved.