Reproduce “HINet: Half Instance Normalization Network for Image Restoration” 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
Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, Chengpeng Chen
Researchers from MEGVII Technology, Fudan University, and Peking University developed HINet, a multi-stage network incorporating novel Half Instance Normalization (HIN) blocks, which achieves state-of-the-art image restoration for denoising, deblurring, and deraining while drastically reducing computational costs. This architecture secured 1st place in the NTIRE 2021 Image Deblurring Challenge.
Key insights
- Normalization layers, particularly a carefully integrated and modified Instance Normalization variant, can significantly enhance performance in low-level image restoration tasks.
- Applying Instance Normalization to only a portion of feature channels provides a crucial balance, re-calibrating statistics for robustness while preserving essential scale and content information vital for dense pixel prediction.
- Multi-stage, U-Net based architectures, when combined with efficient custom building blocks and inter-stage information flow mechanisms, can achieve both superior image quality and substantial computational efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.