Research map / Image generation / Image restoration
Efficient Visual State Space Model for Image Deblurring
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 EVSSM, an efficient visual state space model for image deblurring that uses a single-direction selective scan preceded by flip/transpose geometric transformations, plus a cheaper frequency-domain FFN (EDFFN).
Problem
Transformers and CNNs for image restoration either have quadratic cost or limited non-local receptive fields, and existing visual SSMs (e.g., multi-direction VMamba-style scanning) multiply computational cost for high-resolution deblurring.
Approach
Scan the feature map in one direction only, but apply a fixed geometric transformation (flip or transpose alternating across EVSS modules) before each scan so that the SSM adaptively sees different directions at almost no extra cost; add depth-wise convolutions on B, C, Δ; and move the frequency-domain quantization in the FFN to the end of the block to cut the cost of DFFN.
Key results
- EVSSM reaches 34.51 dB PSNR on GoPro, above FFTformer's 34.21 dB and the next best listed method.
- On RealBlur-J EVSSM gets 34.34 dB PSNR, at least 0.58 dB above the next best method (MLWNet 33.84 dB).
- EVS block with geometric transformation matches one-direction scan cost (126G FLOPs, 88.8ms) while improving PSNR from 33.89 to 34.15 dB on GoPro ablation; four-direction scan costs 148G FLOPs and 182.6ms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EVSSM instead of FFTformer | image deblurring · GoPro test | PSNR (dB): 34.21 -> 34.51 | Helped |
| use EVSSM instead of GRL | image deblurring · HIDE test | PSNR (dB): 31.65 -> 31.99 | Helped |
| use EVSSM instead of Uformer-B | image deblurring · HIDE test | SSIM: 0.9530 -> 0.9503 | Hurt |
| use EVSSM instead of MLWNet | image deblurring · RealBlur-R test | PSNR (dB): +0.58dB | Helped |
| use EVSSM instead of MLWNet | image deblurring · RealBlur-J test | PSNR (dB): +0.5dB | Helped |
| add geometric transformations before one-direction scans (ablation) | EVSSM · GoPro | PSNR (dB): 33.89 -> 34.15 | Helped |
| replace EVS with two-direction scanning (ablation) | EVSSM · GoPro | PSNR (dB): 34.15 -> 33.08 | Hurt |
| replace EVS with four-direction scanning (ablation) | EVSSM · GoPro | PSNR (dB): 34.15 -> 33.05 | Hurt |
| add flip transformation (ablation) | EVSSM · GoPro | PSNR (dB): 33.89 -> 34.05 | Helped |
| add transpose transformation (ablation) | EVSSM · GoPro | PSNR (dB): 33.89 -> 34.03 | Helped |
| add both flip and transpose transformations (ablation) | EVSSM · GoPro | PSNR (dB): 33.89 -> 34.15 | Helped |
| replace EDFFN with DFFN (ablation) | EVSSM encoder/decoder · GoPro | PSNR (dB): 33.68 -> 33.67 | No clear effect |
Surprising findings
- Multi-direction scanning (two- and four-direction) performed slightly worse than the simple geometric-transformation approach in ablation, attributed to naive summation fusion.
Limitations
- Only simple geometric transformations (flip and transpose) were explored.
- Evaluated mainly on deblurring benchmarks; other restoration tasks only briefly tested.
- Paper claims a quarter of the computational cost versus existing SSM methods without a detailed breakdown.
Experimental setup
Datasets: GoPro RealBlur HIDE
Benchmarks: GoPro HIDE RealBlur-J RealBlur-R RESIDE-6K real-world deraining dataset (Yang et al.)
Metrics: PSNR SSIM FLOPs runtime parameters
Baselines: Restormer FFTformer NAFNet Stripformer MLWNet GRL CU-mamba MIMO-UNet+ MPRNet Uformer
Builds on: Mamba VMamba FFTformer Restormer
Closest papers by meaning
- Scan Clusters, Not Pixels: A Cluster-Centric Paradigm for Efficient Ultra-high-definition Image Restoration (CVPR 2026 · similarity 0.88)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Omni-Kernel Network for Image Restoration (AAAI 2024 · similarity 0.86)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.85)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.85)
- Learning from History: Task-agnostic Model Contrastive Learning for Image Restoration (AAAI 2024 · similarity 0.85)
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.