Reproduce “Bridge Damage Detection from Low-Light UAV Imagery via Degradation-Aware Mixture-of-Experts Enhancement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hu Wang, Hongxu Pu, Zhiqi Hu, Fangzhou Lin, Wang Wang
A new framework, DaL-MoE, enhances low-light UAV imagery of bridges using a degradation-aware mixture-of-experts network and an ISP-aware data synthesis pipeline to improve multi-class damage detection. This approach boosted YOLOv11m's box [email protected] from 0.3097 to 0.4923 on synthetic data and from 0.123 to 0.252 on real-world low-light bridge images.
Key insights
- Realistic paired low-light training data can be synthesized by simulating the physical image formation process through an ISP-aware degradation pipeline, enabling supervised learning.
- Explicitly estimating and integrating degradation-aware guidance (illumination and noise prior) into the restoration process allows for adaptive and targeted enhancement of low-light regions.
- An adaptive mixture-of-experts design, with specialized branches for noise, color, and detail, can effectively balance diverse restoration goals and dynamically combine their outputs for task-oriented visual recovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.