BCGW-YOLO: A lightweight network for road damage detection using enhanced feature fusion and dynamically adjusted gradient loss

Abstract

Achieving accurate road damage detection while maintaining low computational cost remains a key challenge, particularly for deployment on the edge devices with limited resources. To address this, we propose BCGW-YOLO, a lightweight yet practical detection framework based on YOLOv8s, tailored for road damage detection in diverse and dynamic environments. We propose the Bidirectional Ghost-based Hierarchical Feature Fusion Network (BG-HFFN), which effectively aggregates features across shallow to deep layers (P2 to P5). By leveraging lightweight Ghost convolutions, the network preserves fine-grained spatial information while significantly reducing computational overhead. In addition, a Content-Shape Feature Enhancement (CSFE) module is specifically designed to improve the model’s ability to extract and fuse shape-specific and contextual features, thereby enhancing the recognition of various crack types. To further improve robustness and convergence, we design a Weighted Focal IoU (WFIoU) loss function that integrates WIoU and Focaler-IoU to address class imbalance and anchor quality issues in complex scenarios. Extensive experiments on the RoadDamageDataset, RDD2020, and RDD2022 validate the effectiveness of the proposed framework, demonstrating a 4.9 % increase in precision and a 2.0 % improvement in [email protected] over the YOLOv8s baseline, along with a 33.9 % reduction in parameters, 9.9 % lower computation, and a 33.3 % smaller model size. These results indicate that BCGW-YOLO provides a practical and efficient solution for real-time road damage inspection in real-world applications.