Exploring the Potential of Super-Resolution for Crack Analysis in UAV Facade Orthomosaics of Small Bridges

Abstract. UAV-based bridge inspections offer significant advantages in efficiency and safety, yet they face a fundamental trade-off between achieving the low Ground Sample Distance (GSD) required for high-precision damage analysis and maintaining operational efficiency. Acquiring imagery fine enough to quantify fine cracks (e.g., < 0.3 mm width) necessitates close-range flights that increase flight time and data volume, thereby diminishing the core benefits of UAVs. This study proposes and validates a workflow that leverages Super-Resolution (SR) technology to enhance the accuracy of quantitative analysis from efficiently captured, low-resolution orthomosaics. To achieve this, we first conducted a comparative analysis of four representative SR models (FSRCNN, SRGAN, Real-ESRGAN, and SwinIR) to identify the optimal architecture for bridge crack restoration. Second, the selected model was applied to a real-world facade orthomosaic (GSD ≈ 0.3 mm) generated from UAV imagery, followed by a quantitative comparison of crack length and width measurement accuracy before and after SR application. The results showed that Real-ESRGAN delivered the best performance. Most notably, the application of SR dramatically reduced the average relative error in crack width measurement from a prohibitive 149.11% to a practical 10.03%, while also more than halving the error in length measurement from 4.80% to 1.93%. This study demonstrates that SR is not merely a visual enhancement technique but a practical solution that enables the acquisition of high-precision, quantitative data comparable to that of a detailed safety inspection, all from safe and efficient UAV operations.