Accurate quantification of concrete bridge defects, including category classification, spatial localization, and geometric measurement, is fundamental to structural health monitoring, since condition assessment relies directly on the localization and dimensional accuracy of the detection system. Unmanned aerial vehicle (UAV) inspection offers a promising acquisition modality, but deploying deep-learning detectors on edge devices faces three challenges: defect morphological diversity, limited resolution of small-scale damage, and constrained onboard computation. This paper proposes BR-DETR, a lightweight variant of RT-DETR-R18 with three task-specific components. The C2f-DIMB backbone employs dynamic multi-branch depthwise convolutions with heterogeneous kernel geometries to adapt the receptive field to diverse defect morphologies while reducing backbone parameters. The Multi-Scale Small Defect Enhanced Pyramid (MSDEP) injects fine-grained P2-level information into the feature pyramid to improve small-defect sensitivity without adding a detection head. The AIFI-HiLo encoder decomposes self-attention into frequency-specialized branches for high-frequency boundaries and low-frequency structural context. On a five-class dataset of 8,924 images, BR-DETR attains 72.51% mAP50 and 54.36% mAP50-95, improving over RT-DETR-R18 by 4.67 and 5.67 percentage points while reducing parameters by 31.9% to 13.54 M and GFLOPs by 8.6%. Considered as a measurement instrument, it reduces the mean center localization error from 6.83 px to 4.92 px (−28.0%) and raises the share of high-quality detections (IoU ≥ 0.75) from 52.3% to 61.8%. Deployed on an NVIDIA Jetson Xavier NX, it achieves 46 FPS with 21.7 ± 1.2 ms latency and a memory footprint below 700 MB, providing a practical tool for real-time, quantitative UAV-based bridge condition assessment.
扫码关注我们
求助内容:
应助结果提醒方式:

