Reproduce “You Only Look Yourself: Unsupervised and Untrained Single Image Dehazing Neural Network” on OpenResearch

About this paper

Boyun Li, Yuanbiao Gou, Shuhang Gu, Jerry Zitao Liu, Joey Tianyi Zhou, Xi Peng

You Only Look Yourself (YOLY) introduces an unsupervised and untrained deep neural network for single image dehazing, eliminating the need for large-scale hazy-clean image datasets and mitigating domain shift. This method disentangles scene radiance, transmission map, and atmospheric light from a single input image, achieving competitive dehazing quality with supervised approaches and enabling realistic haze synthesis.

Key insights

• High-quality single image dehazing is achievable by optimizing a deep neural network on a per-image basis without explicit ground-truth supervision or prior training.
• The atmospheric scattering physical model can serve as a robust self-supervision signal to guide the disentanglement of image components within a deep learning framework.
• Incorporating domain-specific priors (e.g., Color Attenuation Prior) and variational inference for global parameters (e.g., atmospheric light) within an untrained network architecture enables effective unsupervised restoration.

Summary by alphaXiv.