Research map / Image generation / Image restoration
Omni-Kernel Network for Image Restoration
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 OKNet, an efficient convolutional encoder-decoder for image restoration built around an omni-kernel module (large 63x63 and strip depth-wise convolutions, dual-domain global attention, and a 1x1 local branch) placed in the bottleneck.
Problem
Transformer-based image restoration models have quadratic complexity in input size, making them impractical, while prior CNN large-kernel methods have limited receptive fields and no global field.
Approach
An omni-kernel module with three parallel branches (global via frequency-domain channel and spatial attention, large via 63x63 and 1x63/63x1 depth-wise convolutions, and local via 1x1 depth-wise convolution) inserted only at the bottleneck to get multi-scale receptive fields efficiently.
Key results
- OKNet achieves 40.79 dB PSNR on SOTS-Indoor, outperforming DehazeFormer-L by 0.74 dB with about 14% of its FLOPs.
- On DPDD combined category, OKNet reaches 26.18 dB PSNR, 0.2 dB above Restormer with 46% fewer parameters (per the text).
- On CSD desnowing, OKNet reaches 37.99 dB PSNR, 0.7 dB above IRNeXt (37.29 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OKNet instead of DehazeFormer-L | Image dehazing · SOTS-Indoor | PSNR: 40.05 -> 40.79 | Helped |
| Use OKNet instead of IRNeXt | Image desnowing · SRRS | PSNR: 31.91 -> 31.70 | Hurt |
| Add a square large-kernel convolution to the baseline (ablation) | OKNet-S · SOTS-Indoor | PSNR: 31.32 -> 35.07 | Helped |
| Add strip-based convolutions to the square large-kernel convolution (ablation) | OKNet-S · SOTS-Indoor | PSNR: 35.07 -> 35.29 | Helped |
| Add the local branch (ablation) | OKNet-S · SOTS-Indoor | PSNR: 35.29 -> 35.48 | Helped |
| Add DCAM/FCA (ablation) | OKNet-S · SOTS-Indoor | PSNR: +1.52 dB | Helped |
| Add DCAM/SCA (ablation) | OKNet-S · SOTS-Indoor | PSNR: +1.03 dB | Helped |
| Add FSAM (ablation) | OKNet-S · SOTS-Indoor | PSNR: 2 dB | Helped |
| Combine FCA and SCA (ablation) | OKNet-S · SOTS-Indoor | PSNR: 31.32 -> 35.82 | Helped |
| Add FSAM to the global branch with FCA and SCA (ablation) | OKNet-S · SOTS-Indoor | PSNR: 31.32 -> 36.12 | Helped |
Surprising findings
- A 63x63 depth-wise convolution in the bottleneck gives a 3.75 dB gain over the baseline without the global branch, indicating very large kernels help restoration in CNNs.
Limitations
- Separate models must be trained per dataset rather than a single unified model.
- Global branch relies on frequency-domain operations and the 63x63 kernel cannot cover the full field at inference on larger images, requiring a dual-domain workaround.
Experimental setup
Datasets: RESIDE (SOTS-Indoor/Outdoor) Dense-Haze NH-HAZE O-HAZE I-Haze NHR DPDD CSD SRRS Snow100K
Benchmarks: SOTS-Indoor SOTS-Outdoor Dense-Haze NH-HAZE O-HAZE I-Haze NHR DPDD CSD SRRS Snow100K
Metrics: PSNR SSIM MAE LPIPS FLOPs Params
Baselines: GridDehazeNet MSBDN FFA-Net PMNet MAXIM-2S DeHamer SDCE DehazeFormer-L Fourmer DPDNet KPAC DRBNet
Builds on: RepLKNet SLaK LaKDNet MAN Restormer IRNeXt
Closest papers by meaning
- xUnit: Learning a Spatial Activation Function for Efficient Image Restoration (CVPR 2018 · similarity 0.89)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.86)
- Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · similarity 0.86)
- Boosting Image Restoration via Priors from Pre-trained Models (CVPR 2024 · similarity 0.86)
- Efficient Visual State Space Model for Image Deblurring (CVPR 2025 · similarity 0.86)
- Learning from History: Task-agnostic Model Contrastive Learning for Image Restoration (AAAI 2024 · similarity 0.86)
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.