An efficient feature aggregation network for urban road damage detection using street-view images

Road damage detection is essential for proactive road maintenance and the long-term sustainability of urban transportation infrastructure. Existing inspection methods, which rely on specialized equipment or frequent surveys, are often inefficient and costly, especially for large-scale urban road networks. To address these limitations, this paper proposes an Efficient Feature Aggregation Network (EFANet) for urban road damage detection using street-view images. EFANet is specifically designed to cope with the challenges posed by street-view road damage detection, including the small-scale appearance, weak texture representation, and complex background interference. To this end, a cascaded dilation space pyramid module is proposed to enhance multi-scale feature extraction by progressively expanding the receptive field. In addition, a cross-layer feature interaction attention module is introduced to model semantic discrepancies between multi-level and multi-scale features, promoting cross-scale feature fusion and contextual information exchange. Furthermore, an auxiliary feature enhancement branch is designed to incorporate edge and texture information from low-level features into the prediction pathway, which improves the sensitivity toward fine-grained and small-scale damage patterns. Experiments conducted on two public street-view road damage datasets demonstrate the effectiveness and generalization capability of EFANet. On the SVRDD dataset, EFANet achieves 71.7% mAP@50 and 45.4% mAP@50:95, while obtaining 63.8% mAP@50 and 35.4% mAP@50:95 on the USRDD dataset. Moreover, EFANet maintains real-time inference performance at 253.2 FPS with 11.87 M parameters and 39.4 GFLOPs, showing a favorable balance between accuracy and computational efficiency. Overall, by leveraging widely available street-view images, EFANet provides a scalable, cost-effective, and practical solution for intelligent road infrastructure monitoring and maintenance systems.