Overview
CoRE-UIR (Common and Residual Experts for Universal Image Restoration) is a proposed global-local framework designed for all-in-one remote sensing image restoration. It addresses issues of degradation in images acquired by unmanned aerial vehicles (UAVs) and satellites, which can arise from adverse weather, illumination variations, and imaging artifacts. These degradations may co-occur, leading to global distribution shifts and local structural corruption. The framework aims to provide a unified alternative to task-specific restoration pipelines, which often suffer from implicit degradation cues and redundancy in multi-expert designs.
Research Context
Remote sensing imagery, particularly from UAVs and satellites, is susceptible to various degradations. These can include effects from weather conditions, inconsistent lighting, and artifacts introduced during imaging. The co-occurrence of multiple degradation types can complicate restoration efforts, affecting both the overall image distribution and fine-grained structural details. While all-in-one image restoration methods offer a streamlined approach compared to separate pipelines for each degradation type, existing solutions often exhibit weaknesses. These include a reliance on weak or implicit degradation cues and a high degree of parameter redundancy, particularly in multi-expert designs where experts might have overlapping restoration behaviors.
Approach
The CoRE-UIR framework is structured around the Common-and-Residual Expert Block (CoRE). This block explicitly decomposes restoration capacity. A common dense expert handles degradation-invariant restoration tasks, while low-rank residual experts provide degradation-specific compensation. This design promotes adaptive specialization without duplicating expert functionality.
Key components of the CoRE-UIR approach include:
- Degradation Prior Embedding (DPE): This module adapts frozen CLIP features to create an explicit, restoration-oriented prior.
- Global Feature Modulation (GFM): This component is responsible for aligning global feature statistics prior to the application of local residual compensation.
To support the development and evaluation of CoRE-UIR, a new large-scale dataset, MDVD-108K (Multi-Degradation VisDrone), was constructed. This dataset includes both single and compound degradations relevant to UAV restoration. Additionally, a real-world test set was created for evaluation purposes.
Findings
Extensive experiments were conducted across multiple datasets to evaluate the performance of CoRE-UIR. The key findings include:
- CoRE-UIR improved the overall average Peak Signal-to-Noise Ratio (PSNR) by 1.05 dB.
- The framework demonstrated significant computational efficiency, running 11.83 times faster than the strongest baseline method, BaryIR.
- CoRE-UIR achieved a substantial reduction in peak memory usage, decreasing it by 85.3% relative to BaryIR.
- These results indicate a favorable balance between restoration quality and operational efficiency.
- Evaluations on downstream tasks and images with unseen degradations also validated the generalizability of CoRE-UIR.
Why This Matters
The development of CoRE-UIR offers an efficient and effective solution for restoring degraded remote sensing images. Its ability to handle diverse degradations while maintaining computational efficiency could benefit applications relying on high-quality UAV and satellite imagery, potentially enhancing the reliability of data for various analyses.