Road surface defect segmentation using the U-Net deep learning model

Abstract

The rapid expansion of transportation infrastructure in Vietnam has introduced new challenges in road maintenance, particularly in detecting and managing surface defects. Early identification and precise localization of these defects are essential to ensuring traffic safety and optimizing maintenance operations. To address these issues, this study proposes a U-Net–based deep learning model integrated with an internet of things (IoT) architecture for real-time road surface defect detection and monitoring. The proposed system operates on an embedded Raspberry Pi platform equipped with a camera, global positioning system (GPS) module, and auditory alert unit, enabling on-device inference with minimal latency. A mobile web application is also developed to visualize detected defects, record their geolocations, and support road authorities in timely maintenance planning. Experimental results on the pothole dataset demonstrate a training accuracy of 97.7%, while the model achieves a test accuracy of 89.49% on unseen data, confirming the robustness of the model
under various environmental conditions. Unlike previous studies focusing solely on image-based detection, this work introduces a fully integrated IoT-driven ecosystem that bridges deep learning, embedded computing, and cloud-based visualization—enhancing automation, scalability, and real-time responsiveness in intelligent transportation systems.