Research map / Image generation / Image restoration
JDEC: JPEG Decoding via Enhanced Continuous Cosine Coefficients
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 JDEC, a local implicit neural representation that decodes quantized JPEG DCT spectra directly into images across quality factors, using a continuous cosine formulation to estimate dequantized spectra.
Problem
JPEG artifact removal and decoding: existing restoration networks take the already-decoded image as input and are often tied to specific quality factors, discarding the information present in the quantized DCT spectra.
Approach
A continuous cosine formulation module estimates dominant frequencies, amplitudes and a quantization-map embedding from latent features, so a local implicit decoder can take quantized spectra (luma and chroma) as input and jointly perform dequantization and upsampling for all quality factors with one model.
Key results
- On ICB at q=10, JDEC reaches 32.55 dB PSNR versus 32.18 dB for FBCNN and 32.46 dB for Swin2SR (trained on extra data).
- On LIVE-1 at q=100, JDEC achieves 45.14 dB versus 42.23 dB for FBCNN and 43.07 dB for JPEG.
- JDEC has 38.9M parameters and runs in 224.79 ms on a 560x560 image versus 71.95 ms for FBCNN (70.1M params), so it is slower than FBCNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use JDEC instead of FBCNN | JPEG decoding · LIVE-1 q=100 | PSNR (dB): 42.23 -> 45.14 | Helped |
| use JDEC instead of FBCNN | JPEG decoding · ICB q=100 | PSNR (dB): 41.61 -> 43.61 | Helped |
| remove the CCF frequency estimator (ablation) | JDEC · LIVE-1 q=10 | PSNR (dB): 27.76 vs 27.95 | Hurt |
| remove sub-block conversion (ablation) | JDEC · LIVE-1 q=10 | PSNR (dB): 27.69 vs 27.95 | Hurt |
| remove group embedding and CCF (ablation) | JDEC · LIVE-1 q=10 | PSNR (dB): 26.90 vs 27.95 | Hurt |
| replace CCF with the Fourier feature formulation in Eq. (13) (ablation) | JDEC · LIVE-1 q=10 | PSNR (dB): 27.88 vs 27.95 | Hurt |
Surprising findings
- Existing methods such as QGAC and FBCNN show performance drops at very high quality factors (q 90-100), while JDEC keeps improving.
- Removing the group spectra embedding and CCF together causes a severe drop, with PSNR of about 26.9 dB at q=10 versus 27.95 dB for the full model.
Limitations
- Slower and more FLOP-heavy than the CNN-based FBCNN in the reported timing comparison.
- Extreme low quality (q=0) needed a separate JDEC+ variant trained over all quality factors, so the main model does not cover that case.
- Training data is limited to DIV2K and Flickr2K with synthetic OpenCV JPEG compression.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: LIVE-1 BSDS500 ICB
Metrics: PSNR PSNR-B SSIM bits-per-pixel rate-distortion curves bit-error-rate
Baselines: DnCNN QGAC FBCNN DMCNN IDCN Swin2SR JPEG
Builds on: SwinIR SwinV2 LIIF-style local implicit image functions Park et al. spectrum embedding for transformers QGAC
Closest papers by meaning
- DCTMamba: Advancing JPEG Image Restoration Through Long-Sequence Modeling and Adaptive Frequency Strategy (AAAI 2025 · similarity 0.85)
- Practical Learned Lossless JPEG Recompression With Multi-Level Cross-Channel Entropy Model in the DCT Domain (CVPR 2022 · similarity 0.83)
- Faster Neural Networks Straight from JPEG (NeurIPS 2018 · similarity 0.83)
- CosAE: Learnable Fourier Series for Image Restoration (NeurIPS 2024 · similarity 0.82)
- JPEG Inspired Deep Learning (ICLR 2025 · similarity 0.80)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.80)
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.