Research map / Image generation / Image restoration
A Universal Scale-Adaptive Deformable Transformer for Image Restoration across Diverse Artifacts
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 SADT, a transformer for removing structured artifacts (moiré, rain, banding) using scale-enhanced deformable convolution and scale-adaptive deformable attention, trained per task and evaluated on DID, SPAD, and moiré datasets.
Problem
Removing structured artifacts (semi-regular, repetitive patterns such as moiré, rain streaks, and banding) that intertwine with genuine image content and vary in scale, orientation, and distortion.
Approach
Combines cascaded deformable convolutions with parallel large-kernel spatial-channel mixing convolutions (SEDC) and a local self-attention extended with multi-scale dilated grids and predicted deformable sampling offsets (SADA) inside a multi-input multi-output encoder-decoder.
Key results
- Achieves 39.78 dB PSNR on DID debanding, above Restormer (39.50) and MPRNet (39.42).
- Reaches 49.13 dB PSNR on SPAD deraining, 0.23 dB above NeRD-Rain-S with about 42% fewer parameters.
- Obtains 30.77 dB PSNR on TIP2018 moiré removal, outperforming ESDNet (29.81 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SADA to the MIMO variant (ablation) | SADT on DID · DID | PSNR: 38.82 -> 39.61 | Helped |
| Add SEDC to the SADA and MIMO variant (ablation) | SADT on DID · DID | PSNR: 39.61 -> 39.78 | Helped |
| Add SEDC to the MIMO variant (ablation) | SADT on DID · DID | PSNR: 38.82 -> 39.05 | Helped |
| Use SADA dilation Ms {1,2,3,4} instead of {1,1,1,1} (ablation) | SADT on DID · DID | PSNR: 39.58 -> 39.70 | Helped |
| Use SADA dilation Ms {1,3,5,7} instead of {1,2,3,4} (ablation) | SADT on DID · DID | PSNR: 39.70 -> 39.78 | Helped |
Limitations
- Debanding comparisons limited because few open-source deep debanding methods exist, so baselines were adapted from related tasks.
- Training uses only task-specific datasets; generalization to unseen artifact types is not evaluated.
Experimental setup
Datasets: TIP2018 FHDMi LCDMoiré DID SPAD
Benchmarks: TIP2018 FHDMi LCDMoiré DID SPAD
Metrics: PSNR SSIM LPIPS
Baselines: ESDNet RVDNet FHDe2Net MBCNN Restormer MPRNet NeRD-Rain-S DRSformer IDT Uformer-B
Builds on: Restormer deformable convolution deformable attention transformer local self-attention FRFN
Closest papers by meaning
- Bend the Basics: Degradation-Aware Deformable Tokenization for All-in-One Image Restoration (ICML 2026 · similarity 0.85)
- Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · similarity 0.83)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.83)
- Efficient RAW Image Deblurring with Adaptive Frequency Modulation (NeurIPS 2025 · similarity 0.83)
- AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.82)
- DiT4Edit: Diffusion Transformer for Image Editing (AAAI 2025 · similarity 0.82)
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.