Research map / Image generation / Image restoration
Correlation Matching Transformation Transformers for UHD 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 UHDformer, a dual-space Transformer for ultra-high-definition image restoration that transfers top correlated high-resolution features into a low-resolution Transformer via DualCMT and modulates multi-level features with ACM, using only 0.3393M parameters.
Problem
Restoring ultra-high-definition (UHD) images (low-light enhancement, dehazing, deblurring) is hard because full-resolution processing is expensive and existing UHD methods rely on large models (often 17M-35M parameters).
Approach
Builds a high-resolution learning branch and a low-resolution Transformer branch; the Dual-path Correlation Matching Transformation (DualCMT) selects top C/r channels from max- and mean-pooled high-resolution features by channel similarity to replace low-resolution features in attention queries and forward networks, while the Adaptive Channel Modulator (ACM) modulates multi-level high-resolution f
Key results
- On UHD-LL low-light enhancement (trained on UHD-LL), UHDformer reaches 27.113 PSNR / 0.9271 SSIM with 0.3393M parameters, versus 26.226 PSNR for UHDFour (17.5M).
- On UHD-Haze trained on UHD-Haze, UHDformer achieves 22.586 PSNR, above Uformer's 19.828 and UHD's 18.043.
- On GoPro-trained deblurring, UHDformer reports 27.436 PSNR vs. 24.625 for FFTformer, with about 98% fewer parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DualCMT to CMTA and CMTN (ablation) | UHDformer · UHD-LL | PSNR: 25.431 -> 27.113 | Helped |
| add DualCMT to CMTA (ablation) | UHDformer · UHD-LL | PSNR: 26.740 -> 27.113 | Helped |
| add DualCMT to CMTN (ablation) | UHDformer · UHD-LL | PSNR: 26.616 -> 27.113 | Helped |
| add Max-Pooling to DualCMT (ablation) | UHDformer · UHD-LL | PSNR: 25.642 -> 27.113 | Helped |
| add Mean-Pooling to DualCMT (ablation) | UHDformer · UHD-LL | PSNR: 26.291 -> 27.113 | Helped |
| use DualCMT rather than direct Max- and Mean-Pooling Transformation (ablation) | UHDformer · UHD-LL | PSNR: 26.203 -> 27.113; 0.91dB | Helped |
| set DualCMT squeezing factor to r=2 instead of r=1 (ablation) | UHDformer · UHD-LL | PSNR: 26.634 -> 26.742 | Helped |
| set DualCMT squeezing factor to r=3 instead of r=2 (ablation) | UHDformer · UHD-LL | PSNR: 26.742 -> 26.826 | Helped |
| set DualCMT squeezing factor to r=4 instead of r=3 (ablation) | UHDformer · UHD-LL | PSNR: 26.826 -> 27.113 | Helped |
| set DualCMT squeezing factor to r=5 instead of r=4 (ablation) | UHDformer · UHD-LL | PSNR: 27.113 -> 26.874 | Hurt |
| add ACM (ablation) | UHDformer · UHD-LL | PSNR: 26.485 -> 27.113 | Helped |
| replace ECA with ACM (ablation) | UHDformer · UHD-LL | PSNR: 26.727 -> 27.113 | Helped |
Surprising findings
- Max-pooling and mean-pooling transformation without DualCMT (26.203 PSNR) underperforms the full DualCMT (27.113), indicating the correlation selection matters beyond simple pooling.
Limitations
- Some baselines must resize UHD inputs, so comparisons depend on the resizing strategy rather than full-resolution processing.
- Ablation and hyperparameter study limited to UHD-LL.
Experimental setup
Datasets: UHD-LL UHD-Haze UHD-Blur LOL SOTS-ITS GoPro
Benchmarks: UHD-LL UHD-Haze UHD-Blur LOL SOTS-ITS GoPro
Metrics: PSNR SSIM parameter count
Baselines: SwinIR Restormer Uformer LLFlow LLformer UHDFour GridNet MSBDN UHD D4 DehazeFormer DMPHN
Builds on: Restormer ConvNeXt DehazeFormer UHDFour
Closest papers by meaning
- UHD-processer: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts (CVPR 2025 · similarity 0.89)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.88)
- DreamUHD: Frequency Enhanced Variational Autoencoder for Ultra-High-Definition Image Restoration (AAAI 2025 · similarity 0.86)
- Latent Harmony: Synergistic Unified UHD Image Restoration via Latent Space Regularization and Controllable Refinement (NeurIPS 2025 · similarity 0.85)
- Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · similarity 0.85)
- Decoupling Degradations with Recurrent Network for Video Restoration in Under-Display Camera (AAAI 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.