TY - JOUR N2 - High-resolution (HR) crack images have proven valuable for bridge inspection using unmanned aerial vehicles (UAVs), offering fine details crucial for accurate segmentation. Traditional deep learning (DL) struggles with HR images due to downsampling issues and limited computational resources. To address this, we propose Cascade-FcaHRNet, a HR representation learning-based multiscale architecture. It incorporates a frequency-channel attention mechanism to capture tiny crack features, a two-stage cascade operation for global and local refinement, and a region-sensitive loss to avoid ambiguous predictions. Ablation studies confirm the effectiveness of these modifications. Robustness experiments show improvements in performance metrics for crack segmentation. In a field test, the Cascade-FcaHRNet accurately segments bridge cracks wider than 0.5 mm from 4 K resolution images, enhancing safety and efficiency in UAV-based bridge inspection. The approach holds potential for developing scientifically sound maintenance and management strategies. VL - 61 KW - Deep learning KW - Crack segmentation KW - Cascade operation KW - Attention mechanism KW - High-resolution crack images KW - UAV crack inspection SN - 1474-0346 UR - http://dx.doi.org/10.1016/j.aei.2024.102508 A1 - Chu, H A1 - Chen, W A1 - Deng, L AV - restricted ID - discovery10192868 N1 - This version is the author accepted manuscript. For information on re-use, please refer to the publisher?s terms and conditions. Y1 - 2024/08/01/ PB - Elsevier BV TI - Cascade operation-enhanced high-resolution representation learning for meticulous segmentation of bridge cracks JF - Advanced Engineering Informatics ER -