Abstract
Accurate inspection of Reinforced Concrete (RC) structures requires precise rebar counting. Although deep-learning object detectors can extract this information from drone imagery, their effectiveness depends on large, diverse, and well-labeled datasets. Image augmentation can increase data variability, yet its impact on Unmanned Aerial Vehicles (UAVs)-based rebar counting has been underexplored. This study systematically evaluates ten augmentation methods-brightness, contrast, perspective, rotation, scale, shearing, translation, blurring, a probabilistic sampling policy, and a sum of techniques composition-using Faster R-CNN and YOLOv10 across six backbones (ResNet-101, ResNet-152, MobileNetV3; ViT, PVT, Swin Transformer). Performance is reported using AP50, AP50:95, and exact-count accuracy. Results show that augmentation efficacy is both architecture and metric-dependent. The best test-set configuration is YOLOv10-PVT with shearing, which achieves AP50 = 87.71%, AP50:95 = 68.53%, and rebar-count accuracy = 86.27%-improvements of + 5.92, + 9.07, and + 5.99 percentage points, respectively, over the PVT original baseline. A probabilistic sampling policy provides consistent, policy-level gains over original data and approaches the best single transform (especially with a magnitude ramp), whereas indiscriminate a sum of techniques application does not reliably outperform the top single augmentation.