FasterGeo-YOLO: Multi-Scale Geometric Adaptive Feature Enhancement for Lightweight Road Damage Detection

Automatic road damage detection is essential for intelligent transportation infrastructure inspection and road health monitoring. However, road damage targets often exhibit small scales, weak boundaries, irregular geometries, low texture contrast, and large inter-class appearance variations, which make accurate and efficient detection challenging for lightweight detectors. To address these issues, this paper proposes FasterGeo-YOLO, a lightweight multi-scale geometric adaptive enhancement framework based on the YOLO26n baseline. FasterNet is introduced as an efficiency-oriented lightweight backbone, while C2f_DCN deformable convolution modules are embedded at the P3 and P4 feature levels to provide complementary geometric adaptation at different spatial resolutions. The original YOLO26n baseline is DFL-free, whereas FasterGeo-YOLO intentionally reintroduces Distribution Focal Loss (DFL) in its modified detection formulation to strengthen distribution-based bounding-box regression. For fair baseline comparison, all detectors in the main comparison are evaluated at 640 × 640 input resolution; the subsequent 800 × 800 stage is treated separately as a high-resolution refinement strategy. At 640 × 640, FasterGeo-YOLO achieves an mAP@0.5 of 0.567 and an mAP@0.5:0.95 of 0.286. After high-resolution refinement, the final model reaches a Precision of 0.632, Recall of 0.538, mAP@0.5 of 0.580, and mAP@0.5:0.95 of 0.293. Three-seed controlled ablation further confirms the reproducibility of the sequential design. The results demonstrate that FasterGeo-YOLO improves road damage detection while maintaining a lightweight and practically efficient model scale.