The exterior facades of towers in cable-supported bridges, with considerable height and limited accessibility, constitute a weak link in bridge inspection. Conventional approaches, long-distance telescope observation and manual gondola inspection, suffer limited coverage, high missed detections, low efficiency, and elevated operational risks. This study proposes a UAV-based fixed-route rapid screening and precision inspection method for detecting surface defects on bridge towers, aiming to balance inspection efficiency and accuracy and enable fully unmanned inspection. High-quality facade images are acquired by a UAV on a fixed-route inspection with a wide field of view. A YOLOv8-based rapid defect focus method identifies and localizes abnormal regions in large-scale complex images. For these focused areas, a DeepLabV3+-based pixel-level defect segmentation method achieves fine-grained extraction. The model integrates a composite loss combining class-weighted cross-entropy, Tversky loss, and Sobel boundary consistency loss, with lightweight morphological post-processing, to enhance boundary adherence for elongated cracks. The ground sampling distance (GSD) is computed from camera intrinsic parameters and the UAV-to-surface distance. Using skeletonization and shortest-path graph algorithms, the true physical dimensions of cracks are quantitatively measured. A case study on a 70 m tower shows preliminary screening within 40 s at 10 m standoff, with abnormal-region detection accuracy greater than 90%. Fine inspection attains mIoU greater than 0.80, mDice greater than 0.90, and F1 greater than 0.98, and crack-length relative error within 3%. The approach maintains a low false-detection rate and identifies most recorded defects. • Propose a UAV-based method for rapid focus and fine identification of tower defects. • Develop a coarse point cloud based routine inspection method to quantify defects. • Provide fast tower anomaly area focus method based on YOLOv8, full 70m tower in 40s. • Develop a DeepLabV3+ based defect segmentation method with crack length error below 3%. • Demonstrate the method’s practicality and outperformance in real bridge inspection.
Xiong et al. (Sun,) studied this question.