An Edge-Oriented Lightweight Network for Efficient Pavement Damage Detection

Abstract

Pavement damage detection plays a fundamental role in smart and sustainable road infrastructure by enabling timely maintenance and extending pavement service life. For large-scale and continuous road monitoring, detection models are expected to deliver reliable accuracy under practical computational constraints, particularly for edge based deployment. However, many existing approaches involve substantial computational overhead when higher accuracy is pursued. This study proposes an efficient and lightweight pavement damage detection framework based on the you only look once version 8. The framework improves the accuracy-efficiency trade-off through three targeted modifications: a lightweight EfficientNetV2-based backbone for compact feature extraction, an improved bidirectional feature pyramid network for multi-scale feature fusion, and a wise intersection over union-inspired loss for robust localisation of irregular pavement damage. The proposed method is evaluated on the public RDD2022 dataset. Experimental results show that the model achieves an mAP@0.5 of 61.2% and an mAP@0.5:0.95 of 31.9%, outperforming state-of-the-art detectors while maintaining a compact model lightweighting and real-time inference capability. The results indicate that the proposed framework provides a deployment-oriented solution for edge-oriented pavement damage detection in smart and sustainable road infrastructure systems.