OWDiff: Overlapped Wavelet Diffusion Fixes Low-Light Image Blurring
New framework boosts PSNR by 0.58 dB while eliminating blocking artifacts.
Low-light image enhancement remains a challenge for both traditional methods and recent diffusion-based approaches. The existing DiffLL model suffers from two key issues: blocking artifacts caused by the Haar Wavelet Transform (WT) and blurred edges or over-smoothed textures due to the limitations of its High-Frequency Restoration Module (HFRM). To address these, the authors introduce OWDiff, which incorporates an Overlapped Wavelet Transform (OWT) that considers correlations across neighboring regions, structurally preventing blocking artifacts. Additionally, they design a low-frequency-guided High-Frequency Enhance Block (HFEBlock) that strengthens detail recovery, resulting in sharper edges and more reliable textures.
Experiments on the LOLv1 and LOLv2-real datasets show OWDiff consistently outperforms existing LLIE methods both qualitatively and quantitatively. Compared to DiffLL, OWDiff achieves an average PSNR gain of 0.58 dB, a 1.64% relative improvement in SSIM, and a 5.9% relative reduction in LPIPS. The method maintains computational efficiency while delivering superior visual quality. Published in IEICE Transactions on Information and Systems, the code is publicly available, enabling easy replication and adoption for real-world applications like surveillance, photography, and medical imaging where low-light conditions degrade performance.
- OWDiff uses Overlapped Wavelet Transform (OWT) to eliminate blocking artifacts from Haar wavelet-based methods.
- Achieves 0.58 dB PSNR gain, 1.64% SSIM improvement, and 5.9% LPIPS reduction over DiffLL on LOLv1/v2-real datasets.
- Low-frequency-guided HFEBlock restores sharper edges and textures without extra computational cost.
Why It Matters
Clearer low-light images impact photography, surveillance, and medical diagnostics without requiring expensive hardware.