Research map / Label-efficient and robust learning / Domain adaptation and generalization
OSMamba: Omnidirectional Spectral Mamba with Dual-Domain Prior Generator for Exposure Correction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OSMamba, a frequency-domain Mamba exposure correction network with omnidirectional spectral scanning and a dual-domain diffusion prior generator that recovers details without ground truth at inference.
Problem
Exposure correction for under- and over-exposed images, where frequency-domain CNN methods have local receptive fields and regression-only training struggles to restore details in severely degraded regions.
Approach
Adapts Mamba to the Fourier spectrum with four-direction (row, column, positive and negative diagonal) zig-zag scanning of amplitude and phase, and distills a ground-truth-based prior from a teacher extractor into a diffusion-based generator that uses only the input image at inference.
Key results
- On LCDP, OSMamba reaches 24.53 dB PSNR and 0.8773 SSIM, versus 23.89 dB for CoTF.
- On SICE, OSMamba improves PSNR by 1.01 dB over the second-best method (average 22.82 dB).
- Ablation: replacing OS-SSM with VSSM drops PSNR by 0.26 dB; removing the diffusion generator drops PSNR by 0.52 dB on LCDP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OSMamba instead of the strongest reported baseline | Exposure correction · MSEC average | PSNR: 23.57 -> 23.78 | Helped |
| Use OSMamba instead of the strongest reported baseline | Exposure correction · SICE average | PSNR: 21.51 -> 22.82 | Helped |
| Use OSMamba instead of the strongest reported baseline | Exposure correction · LCDP average | PSNR: 23.89 -> 24.53 | Helped |
| Replace OS-SSM with VSSM (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.27 | Hurt |
| Replace OS-SSM with Attention (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.31 | Hurt |
| Use Cross-Scan No.1-2 instead of full OS-Scan (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.30 | Hurt |
| Use Cross-Scan No.3-4 instead of full OS-Scan (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.32 | Hurt |
| Use OS-Scan No.1-2 only (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.34 | Hurt |
| Use OS-Scan No.3-4 only (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.37 | Hurt |
| Use Cross-Scan No.1-4 instead of full OS-Scan (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.35 | Hurt |
| Remove prior guidance (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 23.72 | Hurt |
| Use a spatial-domain extractor without the generator (ablation) | OSMamba · LCDP | PSNR: 24.53 -> 24.01 | Hurt |
Limitations
- Diffusion prior is a compact 1x1xM vector with only T=4 steps rather than full image-space diffusion.
- Evaluated only on exposure correction benchmarks with PSNR/SSIM; HDR reconstruction left as future work.
Experimental setup
Datasets: MSEC SICE LCDP
Benchmarks: MSEC SICE LCDP
Metrics: PSNR SSIM
Baselines: FECNet CoTF LACT MMHT MSEC ERL+FECNet LCDPNet ECLNet ENC+DRBN Zero-DCE RUAS SCI
Builds on: Mamba VMamba FECNet DDPM Restormer
Closest papers by meaning
- Low-Light Image Enhancement via Generative Perceptual Priors (AAAI 2025 · similarity 0.82)
- ECMamba: Consolidating Selective State Space Model with Retinex Guidance for Efficient Multiple Exposure Correction (NeurIPS 2024 · similarity 0.82)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.79)
- URWKV: Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration (CVPR 2025 · similarity 0.79)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.79)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.