Deployment-oriented benchmarking of YOLO detectors for multiclass pavement distress inspection using street-view imagery

Network-level pavement inspection requires reliable detection and classification of surface distress to support condition assessment and maintenance decisions. Although YOLO detectors show promise for automated pavement distress detection, their comparative performance under identical street-view inspection conditions and evaluation protocols remains largely uncharacterised. This study presents a systematic benchmarking and targeted improvement study on the public Street View Image Dataset for Automated Road Damage Detection (SVRDD). Twenty-five models from five YOLO families were evaluated under seven-fold cross-validation, providing a statistically reliable performance reference across model scales and generations. YOLO12 attains the highest detection accuracy, while YOLO11 offers the best balance of accuracy, computational efficiency, and cross-fold stability for network-level inspection. Further analysis of training configurations highlights the importance of consistent experimental settings for fair comparison across detector architectures. Building on the benchmarking results, morphology-aware adaptations applied to YOLO11n —a P2 detection branch and a DCNv2-based head—improve detection of compact and geometrically irregular distress, achieving 0.754 precision and 0.700 mAP@0.5. These findings provide model selection guidance for agencies deploying automated pavement inspection systems under varying computational and accuracy requirements.