核心概念
This research paper introduces DFSSM, a novel deep learning model for single image deraining that leverages the strengths of state space models (SSMs) and frequency domain processing to achieve state-of-the-art performance in removing rain streaks and restoring image quality.
統計資料
DFSSM outperforms the state-of-the-art model DRSformer by 0.82dB on Rain200H, 0.58dB on Rain200L, 0.31dB on DID-Data, 1.01dB on SPA-Data, and 0.44dB on LHP-Rain in terms of PSNR.
DFSSM-S, a lightweight version of DFSSM, achieves competitive performance with fewer parameters (7.0M) and lower FLOPs (87.9G) compared to other high-performing methods.
Using SSMs instead of standard Self-Attention for global receptive fields results in similar performance with 6.1% lower computational costs.