Compositional-Degradation UAV Image Restoration: Conditional Decoupled MoE Network and A Benchmark

A research team has introduced DAME-Net (Degradation-Aware Mixture-of-Experts Network), a novel AI architecture designed to restore images captured by drones (UAVs). Unlike current methods that treat multiple image degradations—like rain, haze, and noise—as a single, entangled problem, DAME-Net explicitly decouples the perception of each degradation factor from the restoration process. This is achieved through a Factor-wise Degradation Perception Module (FDPM) that provides clear, interpretable cues for each type of damage, preventing the corrections for one factor from interfering with another.

The model's core innovation is its Conditioned Decoupled MoE Module (CDMM), which uses these degradation cues to route information through specialized 'expert' networks. This allows for selective, factor-specific correction while suppressing irrelevant interference. To train and evaluate this approach, the team created the Multi-Degradation UAV Restoration (MDUR) benchmark, the first large-scale dataset for this problem, featuring 43 different degradation configurations from single issues to complex four-factor composites. Experiments show DAME-Net consistently outperforms existing unified restoration methods, with particularly significant gains on unseen and higher-order composite degradations that mimic real-world conditions.

Crucially, the improvements aren't just cosmetic. Downstream validation experiments confirm that images restored by DAME-Net lead to better performance in practical applications like UAV object detection, which is critical for infrastructure inspection, emergency response, and large-area mapping. By moving from implicit, entangled representations to explicit, decoupled processing, this research provides a more robust and generalizable framework for handling the messy reality of aerial imagery, paving the way for more reliable vision systems in autonomous drones.