Reproduce “RSHazeDiff: A Unified Fourier-aware Diffusion Model for Remote Sensing 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
Jiamei Xiong, Xuefeng Yan, Yongzhen Wang, Wei Zhao, Xiao-Ping Zhang, Mingqiang Wei
RSHazeDiff introduces a unified Fourier-aware diffusion model specifically designed for remote sensing image dehazing, overcoming limitations of existing methods to produce clearer, more detailed images with accurate colors. This model achieves state-of-the-art quantitative and qualitative performance on various datasets, enhancing critical downstream tasks like vehicle detection.
Key insights
- A phased training strategy, moving from noise-estimation to a combined L1 and MS-SSIM reconstruction loss, effectively refines diffusion model outputs for visually pleasing and accurate dehazing.
- Integrating frequency-domain components, specifically the phase spectrum for structural details and controlled amplitude spectrum for illumination, from the forward diffusion process into reverse sampling significantly improves texture and color fidelity.
- Employing a global learning module that processes the entire image in the Fourier domain can effectively compensate for the local processing of patch-based diffusion models, thereby eliminating artifacts and maintaining global semantic consistency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.