DY-LUT enhances underwater images 300x faster with depth-aware YCbCr lookups
New framework processes 4K underwater images in 7ms, 304x faster than competitors.
Underwater image enhancement has long struggled with wavelength-dependent attenuation that varies with depth. Traditional methods are either too slow for real-time use or fail to adapt spatially. Now, researchers from (affiliations not specified) introduce DY-LUT (Depth-Aware YCbCr Lookup Tables), a framework that combines the efficiency of lookup tables with depth-aware conditioning. The system uses a dual-branch encoder to predict fusion weights and pixel-wise degradation indices from both image and depth features. These outputs condition learnable 4D LUTs, followed by lightweight local refinement. The key innovation is using the YCbCr color space—which separates luminance from chrominance—as the basis for the LUTs, proving more effective than RGB for depth-conditioned lookup.
DY-LUT achieves impressive practical performance. With only 3.56 million parameters, it delivers competitive visual quality on standard benchmarks (UIEB-90 and LSUI) while running 9 to 304 times faster than representative high-capacity baselines. For 4K-resolution images from the UIQAD dataset, adaptive inference maintains real-time speeds of approximately 7ms per frame. The method also benefits downstream tasks like object detection and feature matching. Ablation studies confirm that jointly learned degradation indices further improve adaptive querying. These results provide a physically grounded, efficient route to real-time underwater image enhancement on practical platforms.
- 3.56M-parameter network achieves competitive quality on UIEB-90 and LSUI benchmarks
- Runs 9–304× faster than high-capacity baselines; processes 4K images in ~7ms
- YCbCr color space outperforms RGB for depth-conditioned LUT-based restoration
Why It Matters
Enables real-time underwater vision for autonomous drones, ROVs, and detection systems at a fraction of computational cost.