Research map / Image generation / Image restoration
Enhancing RAW-to-sRGB with Decoupled Style Structure in Fourier Domain
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 FourierISP, a neural ISP that uses the Fourier amplitude/phase decomposition to separately learn structure (phase) and color style (amplitude) for mobile-RAW-to-DSLR-sRGB mapping.
Problem
Mapping smartphone RAW images to DSLR-quality sRGB is hard because the mapping involves both color differences and spatial-structure differences (resolution, misalignment), and shared deep representations entangle these, limiting quality.
Approach
Decouples style (Fourier amplitude) and structure (Fourier phase) in the frequency domain, using a Phase Enhance Subnet for structure, an Amplitude Refine Subnet for color, and a Color Adaptation Subnet that modulates structural features with amplitude via SFT-style blocks, supervised with phase and amplitude losses.
Key results
- On ZRR, FourierISP reaches 21.65 dB PSNR, 0.7546 SSIM and 0.182 LPIPS, improving PSNR by 0.08 dB over the best prior method (LiteISP/LWISP at 21.55/21.57).
- On ZRR aligned with RAW, PSNR reaches 23.93 dB, a 0.17 dB gain over the best baseline (LiteISP at 23.76).
- On MAI after one-epoch fine-tuning from ZRR, PSNR reaches 24.99 dB versus 24.34 dB for the best baseline (MWISP-GAN), with LPIPS 0.0850.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove the phase branch and phase loss (ablation) | FourierISP · ZRR | SSIM: 0.7413 -> 0.7546 | Helped |
| remove the amplitude branch and amplitude loss (ablation) | FourierISP · ZRR | PSNR: 21.37 -> 21.65 | Helped |
| replace AmplitudeLoss with ColorHistLoss (ablation) | FourierISP · ZRR | PSNR: 21.65 -> 21.13 | Hurt |
| reduce channel count to 16 (ablation) | FourierISP · ZRR | PSNR: 21.65 -> 21.64 | Hurt |
| increase channel count to 48 (ablation) | FourierISP · ZRR | PSNR: 21.65 -> 21.67 | Helped |
| use FourierISP instead of LiteISP | RAW-to-sRGB mapping · ZRR (Align GT with RAW) | PSNR: +0.17dB | Helped |
| use FourierISP instead of the SOTA method | RAW-to-sRGB mapping · ZRR | PSNR: +0.08dB | Helped |
Surprising findings
- Replacing the amplitude loss with a color-histogram loss degrades results, indicating coarse histogram matching is poorly suited to RAW-to-RGB.
Limitations
- Relies on optical flow alignment for ZRR evaluation and baseline LiteISP training, limiting transferability of that baseline
- Fixed loss weights (0.1) set by experience without sensitivity analysis
- Reducing channels to 16 lowers visual quality despite similar PSNR
Experimental setup
Datasets: ZRR MAI
Benchmarks: ZRR MAI
Metrics: PSNR SSIM MS-SSIM LPIPS
Baselines: PyNet AWNet MWISP MWISP-GAN LiteISP LWISP
Builds on: LiteISP HINBlock FFC SFT
Closest papers by meaning
- ISPDiffuser: Learning RAW-to-sRGB Mappings with Texture-Aware Diffusion Models and Histogram-Guided Color Consistency (AAAI 2025 · similarity 0.86)
- DiffRAW: Leveraging Diffusion Model to Generate DSLR-Comparable Perceptual Quality sRGB from Smartphone RAW Images (AAAI 2024 · similarity 0.85)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.84)
- Efficient RAW Image Deblurring with Adaptive Frequency Modulation (NeurIPS 2025 · similarity 0.83)
- ChromaFusionNet (CFNet): Natural Fusion of Fine-Grained Color Editing (AAAI 2024 · similarity 0.83)
- Intrinsic Phase-Preserving Networks for Depth Super Resolution (AAAI 2024 · 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.