Bridge Damage Detection from Low-Light UAV Imagery via Degradation-Aware Mixture-of-Experts Enhancement

Automated bridge inspection using unmanned aerial vehicles (UAVs) is frequently hindered by poor illumination beneath bridges, inside structural components, and within occluded regions. These low-light conditions cause underexposure, signal noise amplification, and contrast reduction, obscuring fine defects like small cracks, spalling, and corrosion stains. This paper investigates whether degradation-aware image restoration can improve bridge damage detection under low illumination and achieve effective transfer from synthetic degradations to real inspection environments.

The authors propose DaL-MoE, a detector-agnostic image restoration front end designed specifically for low-light bridge imagery. Because paired normal-light and low-light UAV field images are difficult to acquire, the method constructs an ISP-aware low-light synthesis pipeline that transforms normal-light images into spatially aligned low-light samples while retaining original object detection and instance segmentation annotations. The DaL-MoE architecture incorporates degradation-aware guidance estimation alongside specialized complementary experts for noise suppression, color adjustment, and structural-detail recovery.

Evaluated on paired synthetic data, DaL-MoE achieves strong quantitative performance with 23.12 dB PSNR and 0.8482 SSIM. When paired with downstream object detectors, the front end substantially improves performance; for instance, it increases the YOLOv11m bounding-box mAP50 from 0.3097 to 0.4923 and mask mAP50 from 0.2281 to 0.3529. In real-world UAV inspection scenarios lacking paired normal-light references, sim-to-real evaluations demonstrate visibly enhanced defect visibility and more complete, reliable damage detections compared to direct inference on raw low-light inputs.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.