Research map / Image generation / Image restoration
DCTMamba: Advancing JPEG Image Restoration Through Long-Sequence Modeling and Adaptive Frequency Strategy
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 DCTMamba, a U-Net-style Mamba network that scans DCT coefficients from low to high frequency and normalizes them by image size, for JPEG artifact removal.
Problem
Mamba's sequential scanning is ill-suited to non-causal image data, and JPEG restoration must recover high-frequency components lost after DCT quantization, with DCT coefficient distributions varying with image size.
Approach
Apply DCT inside Mamba so the scan runs from low to high frequencies (coarse-to-fine causal ordering), normalize DCT coefficients by image size (Scale-Adaptive Normalization), and run a parallel convolutional branch for local detail.
Key results
- On Classic5 Q10, DCTMamba reports 30.25 dB PSNR vs 30.12 dB for FBCNN and Zhao et al.
- On Classic5 Q40, 34.46 dB PSNR / 0.908 SSIM, the best among compared methods in Table 1.
- Ablation on Classic5 Q40: full model 34.46 dB vs 34.21 dB baseline without Mamba/conv modules; removing SASSB drops it to 34.37 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCTMamba instead of the strongest reported baseline | JPEG restoration · Classic5 Q40 | PSNR: 34.41 -> 34.46 | Helped |
| Use DCTMamba instead of the strongest reported baseline | JPEG restoration · LIVE1 Q40 | PSNR: 34.55 -> 34.63 | Helped |
| Use DCTMamba instead of the strongest reported baseline | JPEG restoration · BSDS500 Q40 | PSNR: 34.34 -> 34.43 | Helped |
| Add the Mamba module to the Global-to-Local Block (ablation) | DCTMamba · Classic5 Q40 | PSNR: 34.21 -> 34.46 | Helped |
| Add the Mamba module to the Global-to-Local Block (ablation) | DCTMamba · Classic5 Q40 | SSIM: 0.905 -> 0.908 | Helped |
| Add the convolution module to the Global-to-Local Block (ablation) | DCTMamba · Classic5 Q40 | PSNR: 34.33 -> 34.46 | Helped |
| Add the convolution module to the Global-to-Local Block (ablation) | DCTMamba · Classic5 Q40 | SSIM: 0.906 -> 0.908 | Helped |
| Scan in the DCT frequency domain instead of directly in the time domain (ablation) | DCTMamba · Classic5 Q40 | PSNR: 34.29 -> 34.46 | Helped |
| Scan in the DCT frequency domain instead of directly in the time domain (ablation) | DCTMamba · Classic5 Q40 | SSIM: 0.906 -> 0.908 | Helped |
| Add Scale-Adaptive State Space Block (ablation) | DCTMamba · Classic5 Q40 | PSNR: 34.37 -> 34.46 | Helped |
| Add Scale-Adaptive State Space Block (ablation) | DCTMamba · Classic5 Q40 | SSIM: 0.907 -> 0.908 | Helped |
| Use the full loss combination instead of setting λssim to 0 (ablation) | DCTMamba · Classic5 Q40 | PSNR: 34.42 -> 34.46 | Helped |
Limitations
- Gains over prior methods are small (e.g., about 0.1 dB on Classic5 Q10).
- Evaluation limited to grayscale Y-channel results on a few standard datasets.
Experimental setup
Datasets: DIV2K Flickr2K BSD500 WED Classic5 LIVE1 BSDS500
Benchmarks: Classic5 LIVE1 BSDS500
Metrics: PSNR SSIM PSNR-B
Baselines: ARCNN DnCNN DCSC MWCNN RDN QGAC FBCNN Zhao et al. RNAN
Builds on: Mamba S4 MambaIR VMamba
Closest papers by meaning
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.88)
- JDEC: JPEG Decoding via Enhanced Continuous Cosine Coefficients (CVPR 2024 · similarity 0.85)
- MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.85)
- DiMSUM: Diffusion Mamba - A Scalable and Unified Spatial-Frequency Method for Image Generation (NeurIPS 2024 · similarity 0.83)
- Making Old Film Great Again: Degradation-aware State Space Model for Old Film Restoration (CVPR 2025 · similarity 0.83)
- FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining (ICML 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.