Reproduce “GT-Mean Loss: A Simple Yet Effective Solution for Brightness Mismatch in Low-Light Image 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
Jingxi Liao, Shijie Hao, Richang Hong, Meng Wang
The GT-mean loss, developed by researchers at Hefei University of Technology, addresses brightness inconsistencies in supervised low-light image enhancement by dynamically weighting an original loss term with a brightness-aligned counterpart. This approach consistently improves image quality across various models and datasets, enhancing metrics such as PSNR, SSIM, and perceptual scores like LPIPS, without requiring architectural changes.
Key insights
- Minor brightness discrepancies in supervised LLIE, often overlooked, can significantly hinder optimal model training by misguiding pixel-level loss functions into a 'futile optimization cycle'.
- Human visual perception prioritizes noise suppression and color accuracy once adequate brightness is achieved, suggesting that loss functions should dynamically de-emphasize brightness errors as they diminish.
- Incorporating a perceptually inspired, dynamic weighting into the loss function can universally enhance existing LLIE architectures without complex redesigns or significant computational overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.