Reproduce “Equivariant Learning for Unsupervised Image Dehazing” 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
Zhang Wen, Jiangwei Xie, Dongdong Chen
Researchers at Heriot-Watt University developed Equivariant Image Dehazing (EID), a fully unsupervised framework for image dehazing that learns from only hazy images without ground truth or strong priors. The method extends Equivariant Imaging to nonlinear inverse problems, achieving superior perceptual quality and fidelity across scientific and natural imaging benchmarks, for instance, outperforming nine state-of-the-art methods in terms of NIQE, BRISQUE, and FID scores on scientific datasets and PSNR/SSIM on natural datasets.
Key insights
- Equivariant Imaging principles, previously applied to linear inverse problems, can be successfully extended to complex nonlinear and blind inverse problems like image dehazing, even with an unknown forward model.
- Learning a pseudo-haze model through adversarial training allows for effective modeling of unknown haze physics, providing a crucial 'known' forward operator for an unsupervised dehazing framework.
- Combining haze consistency (constraining learning within the haze model's range space) and equivariant consistency (recovering clear patterns from the nullspace via symmetry) is essential for robust unsupervised dehazing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.