<p>This study proposes a framework integrating unmanned aerial vehicles and deep learning for the detection, classification, and quantification of surface cracks—a critical damage type in concrete bridges—while addressing limitations of traditional manual inspection methods; it classifies cracks into two categories: slightly tortuous normal linear cracks with measurable width and length, and crazing cracks, the latter with a disordered, reticular morphology precluding effective width/length quantification, so only normal linear cracks are quantified. Considering massive high-resolution image data from unmanned aerial vehicles and high computational demands of deep learning models for semantic segmentation tasks involving millions of pixels, a high-performance computing cluster accelerates image preprocessing, model training, and inference—meeting timeliness requirements of real-time bridge health monitoring—with unmanned aerial vehicles capturing high-resolution images to cover large bridge surfaces, split via an adaptive segmentation technique to enhance processing efficiency and accuracy, while a bridge classification model filters out irrelevant backgrounds to focus crack detection on the bridge structure. Subsequently, a deep learning-based multi-type crack recognition and damage detection model is deployed: a U-Net network performs crack segmentation, a custom algorithm conducts measurement, and integrating these technologies into a unified system significantly improves bridge crack identification efficiency, providing data-driven support for maintenance and management; experimental results show the method substantially reduces annotation costs, mitigates challenges in data collection, and enhances model generalization—delivering considerable value for concrete bridge health monitoring applications.</p>