Research map / Image generation / Image restoration
Efficient RAW Image Deblurring with Adaptive Frequency Modulation
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 FrENet, a U-Net with frequency-domain FrE-Blocks and an Adaptive Frequency Positional Modulation module for efficient RAW-to-RAW deblurring, which also extends to sRGB deblurring.
Problem
RAW image deblurring is underexplored, and existing methods struggle to handle frequency-dependent blur while staying computationally efficient.
Approach
Operates directly in the FFT domain inside a compact U-Net; AFPM generates position-dependent kernels and biases from each patch's normalized distance to the spectrum center to modulate frequency components, combined with frequency-domain skip connections and a NAFNet-style channel attention branch.
Key results
- FrENet+ reaches 45.63 dB PSNR on Deblur-RAW, versus 44.04 dB for LoFormer-L, which uses 8.98G MACs against FrENet's 2.22G MACs for FrENet at 44.73 dB.
- FrENet+ achieves 33.87 dB PSNR on RealBlur-J, 0.97 dB above LoFormer-L's 32.90 dB.
- Removing the frequency skip connection drops Deblur-RAW PSNR from 44.73 to 44.39 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| FrENet versus LoFormer-L | RAW image deblurring · Deblur-RAW | PSNR: +0.69dB | Helped |
| Extend FrENet+ to sRGB without architectural modifications | FrENet+ · RealBlur-J | PSNR: +0.97dB PSNR | Helped |
| Add frequency skip connection while retaining spatial skip connection (ablation) | FrENet · Deblur-RAW | PSNR: 44.39 -> 44.73 | Helped |
| Replace adaptive modulation with average pooling (ablation) | FrENet · Deblur-RAW | PSNR: 44.35 -> 44.73 | Hurt |
| Use 2 × 2 rather than 8 × 8 AFPM division (ablation) | FrENet · Deblur-RAW | PSNR: 44.44 -> 44.73 | Hurt |
| Use 4 × 4 rather than 8 × 8 AFPM division (ablation) | FrENet · Deblur-RAW | PSNR: 44.52 -> 44.73 | Hurt |
| Add global branch to the local AFPM branch (ablation) | FrENet FrE-Block · Deblur-RAW | PSNR: 44.67 -> 44.73 | Helped |
| Add local AFPM branch to the global branch (ablation) | FrENet FrE-Block · Deblur-RAW | PSNR: 44.48 -> 44.73 | Helped |
Limitations
- Heavy reliance on FFT/IFFT costs substantial computation for high-resolution images.
- AFPM's simplified positional encoding may not capture complex spatial dependencies.
Experimental setup
Datasets: Deblur-RAW GoPro RealBlur-R RealBlur-J HIDE
Benchmarks: Deblur-RAW GoPro HIDE RealBlur-R RealBlur-J
Metrics: PSNR SSIM MACs Params
Baselines: NAFNet DeepRFT Restormer FFTFormer LoFormer-S LoFormer-L DeepRFT+ LoFormer-B
Builds on: NAFNet Restormer LoFormer SimpleGate
Closest papers by meaning
- AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.86)
- Complementary Advantages: Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising (CVPR 2025 · similarity 0.85)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.85)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.85)
- Efficient Visual State Space Model for Image Deblurring (CVPR 2025 · similarity 0.85)
- Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring (CVPR 2024 · similarity 0.85)
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.