This study proposes and evaluates deep learning models for the segmentation of exposed steel rebar in concrete structures — a critical damage type often linked to concrete delamination. The motivation emerged from technical discussions with the Municipal Infrastructure Secretaria(SMI) of Rio de Janeiro, which identified this pathology as a priority in the inspection of urban bridges. The dataset used includes approximately 100 images, combining real UAV-captured frames and publicly available databases, with precise binary masks manually annotated. Three segmentation architectures were evaluated: U-Net with ResNet-34 encoder, lightweight U²-NetP, and a hybrid model combining Swin Transformer and DeepLabV3. Training employed k-fold cross-validation, data augmentation, and grid search with BCE+Dice+Focal loss functions. Only U-Net and Swin+DeepLabV3 were tested in real UAV inspection footage. Results show that Swin+DeepLabV3 achieved higher accuracy in detectingrebar under diverse lighting and textures, outperforming U-Net especially in multi-instance scenarios. The study highlights the potential of modern semantic segmentation models for field-ready structural diagnostics and contributes to closing a gap in the literature regarding real-world applications focused on rebar exposure.
