Research map / Image generation / Image restoration
CoDe: An Explicit Content Decoupling Framework 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 CoDe, a plug-in framework that decouples image content into DCT frequency components with learnable masks and restores each with modulated sub-networks trained under a content consistency loss, boosting existing IR networks across SR, denoising, and deblurring.
Problem
Image restoration networks treat all image regions with the same unified pipeline, so regions with different complexity/frequency content are poorly reconstructed, causing over-smoothing and wrong textures.
Approach
Decouple LQ/GT into n frequency sub-bands using learnable, adaptive radii on the whole-image type-II DCT spectrum, restore each band with a parameter-shared cascade of sub-nets of increasing size plus per-band reconstruction layers, and supervise each band's DCT-domain output with a weighted L1 content consistency loss; the framework is architecture-agnostic.
Key results
- On classical SR ×2 Set5, DAT improves by 0.34 dB with CoDe (38.58 to 38.72 is 0.14 dB in Table 1; abstract claims 0.34 dB at ×2 Set5 for DAT)
- Applied to GRL, deblurring PSNR rises by 0.12 dB on GoPro and 0.11 dB on HIDE (33.93 to 34.05; 31.65 to 31.76)
- Real image denoising on SIDD: MIRNet-v2 gains 0.11 dB with the framework (39.84 to 39.95)
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CoDe to GRL | GRL · Set5 ×4 | PSNR: 33.10 -> 33.21 | Helped |
| Apply CoDe to GRL | GRL · Set14 ×4 | PSNR: 29.37 -> 29.46 | Helped |
| Apply CoDe to GRL | GRL · BSD100 ×4 | PSNR: 28.01 -> 28.13 | Helped |
| Apply CoDe to GRL | GRL · Urban100 ×4 | PSNR: 28.53 -> 28.61 | Helped |
| Apply CoDe to GRL | GRL · CBSD68 σ=15 | PSNR: 34.45 -> 34.52 | Helped |
| Apply CoDe to GRL | GRL · Kodak24 σ=15 | PSNR: 35.43 -> 35.56 | Helped |
| Apply CoDe to GRL | GRL · McMaster σ=15 | PSNR: 35.73 -> 35.84 | Helped |
| Apply CoDe to GRL | GRL · Urban100 σ=15 | PSNR: 35.54 -> 35.69 | Helped |
| Apply CoDe to GRL | GRL · Set12 σ=15 | PSNR: 33.47 -> 33.64 | Helped |
| Apply CoDe to GRL | GRL · BSD68 σ=15 | PSNR: 32.00 -> 32.16 | Helped |
| Apply CoDe to MIRNet-v2 | MIRNet-v2 · SIDD | PSNR: 39.84 -> 39.95 | Helped |
| Apply CoDe to NBNet | NBNet · DND | PSNR: 39.89 -> 39.98 | Helped |
Surprising findings
- Reducing the width or depth of the original network (overfitting reduction) corrects wrongly restored textures, motivating content-specific sub-networks
Limitations
- Gains are modest (often 0.01–0.15 dB) on several benchmarks
- Frequency decomposition uses a fixed number of components n set per task by hand
- Retrained baselines on 4 V100 GPUs rather than reported numbers
Experimental setup
Datasets: DIV2K DF2K Set5 Set14 BSD100 Urban100 Set12 BSD68 McMaster Kodak24 SIDD DND
Benchmarks: Set5 Set14 BSD100 Urban100 CBSD68 Kodak24 McMaster Set12 BSD68 SIDD DND GoPro
Metrics: PSNR SSIM
Baselines: ELAN Omni-SR HAT SRFormer DAT GRL RNAN SwinIR DRUNet CycleISP AINDNet NBNet
Builds on: EDSR DAT HAT GRL SwinIR
Closest papers by meaning
- IR-CM: The Fast and General-purpose Image Restoration Method Based on Consistency Model (NeurIPS 2024 · similarity 0.87)
- AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.86)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.85)
- IRPO: Boosting Image Restoration via Post-training GRPO (NeurIPS 2026 · similarity 0.84)
- IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models (ICML 2025 · similarity 0.84)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.84)
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.