Research map / Image generation / Image restoration
AdaRevD: Adaptive Patch Exiting Reversible Decoder Pushes the Limit of Image Deblurring
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 AdaRevD, a reversible multi-sub-decoder image deblurring network on a frozen pretrained encoder (UFPNet/NAFNet) with a degradation-degree classifier for per-patch early exit.
Problem
Image deblurring SOTA methods are limited by lightweight decoders (insufficient decoding capacity) because heavy encoders consume GPU memory, capping performance.
Approach
Inherit a frozen well-trained encoder and build a reversible decoder made of stacked reversible sub-decoders (RevCol-style) to scale decoder capacity with near-constant training memory; a classifier predicts each patch's degradation degree so patches can exit at different sub-decoders.
Key results
- AdaRevD-L reaches 34.60 dB PSNR on GoPro (setting A), vs. UFPNet 34.06 dB.
- RevD-B achieves 37.47 dB average on RealBlur (setting B), and AdaRevD-L reaches 41.19 dB on RealBlur-R vs. UFPNet 40.61 dB.
- Classifier-based patch exiting in AdaRevD-B keeps PSNR at 34.50 dB on GoPro while using 84.3%-85.8% of sub-decoders (RevD-B 34.515 dB at 100%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the reversible multi-sub-decoder to UFPNet | UFPNet · GoPro test | PSNR: 34.06 -> 34.60 | Helped |
| Add the reversible multi-sub-decoder to UFPNet | UFPNet · HIDE test | PSNR: 31.74 -> 32.37 | Helped |
| Add the reversible multi-sub-decoder to UFPNet | UFPNet · RealBlur-R test, setting B | PSNR: 40.61 -> 41.19 | Helped |
| Add the reversible multi-sub-decoder to UFPNet | UFPNet · RealBlur-J test, setting B | PSNR: 33.35 -> 33.96 | Helped |
| Add FourierBlock instead of NAFBlock (ablation) | AdaRevD decoder · GoPro test | PSNR: 34.15 -> 34.51 | Helped |
| Add FuseBlock (ablation) | AdaRevD decoder with FourierBlock · GoPro test | PSNR: 34.24 -> 34.51 | Helped |
| Increase sub-decoder count from 4 to 8 (ablation) | AdaRevD decoder · GoPro test | PSNR: 34.51 -> 34.64 | Helped |
| Use 4 blocks per level instead of 1 (ablation) | AdaRevD decoder · GoPro test | PSNR: 34.51 -> 34.49 | Hurt |
| Continue training UFPNet with encoder trainable (ablation) | UFPNet · GoPro test | PSNR: 34.09 -> 33.71 | Hurt |
| Continue training NAFNet with encoder trainable rather than frozen (ablation) | NAFNet · GoPro test | PSNR: 33.70 -> 33.75 | No clear effect |
| Use classifier-based patch exiting with e4 at τ=0.05 (ablation) | RevD-B · GoPro test | PSNR: 34.515 -> 34.501 | Hurt |
| Use classifier-based patch exiting with e4 rather than e5 at τ=0.05 (ablation) | RevD-B · GoPro test | PSNR: 34.502 -> 34.501 | No clear effect |
Surprising findings
- Continuing training of UFPNet with its encoder unfrozen decreased PSNR from 34.09 dB to 33.71 dB.
- The first sub-decoder cannot decode an accurate blur pattern from the degradation representation, per CKA analysis.
Limitations
- Exit thresholds and degradation-degree classes are dataset-specific and hand-configured.
- Adaptive exit yields only modest compute savings (about 15%) with a small PSNR drop in the default setting.
Experimental setup
Datasets: GoPro HIDE RealBlur-R RealBlur-J
Benchmarks: GoPro HIDE RealBlur-R RealBlur-J
Metrics: PSNR SSIM GPU training memory
Baselines: MPRNet Restormer UFPNet NAFNet FFTformer Stripformer DeepRFT+ MRLPFNet
Builds on: RevCol NAFNet UFPNet DeepRFT APE-SR
Closest papers by meaning
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.85)
- Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (CVPR 2024 · similarity 0.85)
- AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.84)
- Degradation-Aware Feature Perturbation for All-in-One Image Restoration (CVPR 2025 · similarity 0.84)
- Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · similarity 0.84)
- Adaptive Window Pruning for Efficient Local Motion Deblurring (ICLR 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.