Bridge-RF-DETR: an edge-guided transformer framework for multi-type bridge surface damage detection
Abstract
Bridge surface damage detection is essential for intelligent infrastructure monitoring, yet remains challenging due to large-scale variations, weak boundaries, complex backgrounds, and imbalanced damage categories. This study develops Bridge-RF-DETR, a task-oriented extension of RF-DETR for multi-type bridge surface damage detection. The framework introduces three targeted modifications. First, a progressive Asymptotic Feature Pyramid Network (AFPN) is integrated to improve multi-scale feature representation for damages of different sizes. Second, an Edge Branch (EB) and an Edge Guidance Module (EGM) are used to incorporate boundary cues for weak-texture and slender damage patterns such as cracks and leakage. Third, a class-balanced sampling strategy is adopted to reduce the effect of category imbalance. Experiments on a public bridge damage dataset show that Bridge-RF-DETR achieves an overall F 1 -score of 0.871 and mAP@0.5 of 0.862, providing moderate but consistent improvements over RF-DETR and other representative detectors. Alternative-component comparison, per-class analysis, three-seed variability analysis, and evaluation on DACL10K provide additional evidence for the observed performance trends. These results indicate that Bridge-RF-DETR provides a practical visual detection pipeline for bridge surface damage inspection, while broader validation under diverse field conditions remains necessary.