Two-Stage Detection of Road Surface Anomalies Using Faster R-CNN with CNN and Transformer Backbones

Road-surface anomalies such as potholes and speed bumps remain a significant concern for traffic safety and vehicle maintenance. Reliable early detection is essential for proactive road management and for supporting Intelligent Transportation Systems (ITS) and Advanced Driver Assistance Systems (ADAS). This work investigates a two-stage detection framework using Faster R-CNN with multiple backbone options, including a convolutional ResNet-50-FPN, an Effi-cientNet-B3-FPN, and a transformer-based Swin-Tiny-FPN. All models were trained and evaluated on a custom COCO-formatted dataset comprising over 4,700 annotated road images captured under varied illumination conditions, including daytime, nighttime, and shadow variations. Among the evaluated backbones, Swin-FRCNN achieved superior detection performance, attaining a mean IoU of 84.75% evaluated at IoU@0.5 and a mean Average Precision of 0.638 (AP@[0.5–0.95]), compared with ResNet-FRCNN (78.25% mIoU, 0.561 mAP). EfficientNet-B3-FRCNN exhibited lower detection accuracy despite reduced backbone FLOPs. Performance gains were particularly pronounced for small-scale anomalies, as reflected in improvements in both Average Precision and Average Recall. All reported results are averaged over three independent runs, and statistical significance is assessed using paired t-tests at a 95% confidence level. While Swin-FRCNN provides higher accuracy and stronger multi-scale representations, it incurs increased computational cost in terms of FLOPs, inference latency, and GPU memory usage, whereas ResNet-FRCNN remains more suitable for real-time or resource-constrained deployment. For additional context, a lightweight single-stage detector (YOLOv8n), fine-tuned from pretrained weights, was evaluated as a reference baseline. Overall, the study highlights clear accuracy–efficiency trade-offs and provides a comprehensive comparison of CNN and trans-former backbones for deployable road-surface anomaly detection in intelligent transportation and smart-city systems.