Reproduce “Generalized Lightness Adaptation with Channel Selective Normalization” 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
Mingde Yao, Jie Huang, Xin Jin, Ruikang Xu, Shenglong Zhou, Man Zhou, Zhiwei Xiong
Researchers from the University of Science and Technology of China developed Channel Selective Normalization (CSNorm), a lightweight and plug-and-play module that enables deep learning models to generalize robustly across diverse, unknown lightness conditions in tasks like low-light enhancement, inverse tone mapping, and image retouching, while preserving reconstruction accuracy on known conditions.
Key insights
- Conventional normalization acts as a 'double-edged sword' for lightness adaptation, improving generalization at the cost of information loss and reduced reconstruction accuracy.
- Selectively normalizing only the lightness-relevant feature channels allows for robust generalization without sacrificing critical information needed for high-fidelity image reconstruction.
- Training with frequency-based lightness perturbations and an amplitude loss in the frequency domain effectively guides the network to identify and adapt to true lightness variations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.