Research map / Image generation / Image restoration
Quad Bayer Joint Demosaicing and Denoising Based on Dual Encoder Network with Joint Residual Learning
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 DRNet, a dual-encoder network with joint residual blocks and a pixel modulation block for joint demosaicing and denoising of Quad Bayer CFA RAW images.
Problem
Joint demosaicing and denoising of Quad Bayer CFA RAW images, where the enlarged color gaps and noise make reconstruction to RGB difficult for Bayer-oriented methods.
Approach
A dual encoder (Quad Bayer encoder with a pixel modulation block regrouping 2x2 same-color pixels, plus a self-adaptive encoder with joint residual blocks that separately estimate mosaic and noise residuals and a self-adaption block for RAW-to-sRGB domain transfer).
Key results
- DRNet reaches 37.46 dB average PSNR at sigma=5 on Quad Bayer benchmarks, the best among compared methods, per Table 1.
- On SIDD, DRNet achieves 35.72 dB PSNR and 0.8621 SSIM versus 35.25 dB for RSTCA (second best).
- DRNet-M (1.405M params) obtains the second-best average PSNR and SSIM with the lowest inference cost on a mobile processor.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DRNet instead of Restormer | Quad Bayer joint demosaicing and denoising · Urban100, σ=5 | PSNR: 37.67 -> 38.05 | Helped |
| Use DRNet instead of Restormer | Quad Bayer joint demosaicing and denoising · Urban100, σ=25 | PSNR: 32.84 -> 32.95 | Helped |
| Use DRNet instead of RSTCA | Quad Bayer joint demosaicing and denoising · SIDD | PSNR: 35.25 -> 35.72 | Helped |
| Add the noise removal branch to the JRB (ablation) | DRNet · Urban100, σ=25 | PSNR: 31.24 -> 32.95 | Helped |
| Add the noise removal branch to the JRB (ablation) | DRNet · BSD100, σ=25 | PSNR: 32.84 -> 34.86 | Helped |
| Add the mosaic removal branch to the JRB (ablation) | DRNet · Urban100, σ=25 | PSNR: 29.57 -> 32.95 | Helped |
| Add the mosaic removal branch to the JRB (ablation) | DRNet · BSD100, σ=25 | PSNR: 30.29 -> 34.86 | Helped |
| Use PMB instead of a pixelshuffle layer (ablation) | DRNet QBE · Urban100, σ=25 | PSNR: 32.52 -> 32.95 | Helped |
| Use PMB instead of a pixelshuffle layer (ablation) | DRNet QBE · BSD100, σ=25 | PSNR: 33.90 -> 34.86 | Helped |
| Add RRDBs to the QBE (ablation) | DRNet · Urban100, σ=25 | PSNR: 32.68 -> 32.95 | Helped |
| Add RRDBs to the QBE (ablation) | DRNet · BSD100, σ=25 | PSNR: 34.17 -> 34.86 | Helped |
| Add SAB to the SAE (ablation) | DRNet · Urban100, σ=25 | PSNR: 31.24 -> 32.95 | Helped |
Surprising findings
- Removing both JRB and SAB gave better results than removing only one of them, which the authors attribute to SAB being needed for JRB training.
Limitations
- Transformer-based baselines Restormer and RSTCA run out of memory on large inputs, so efficiency comparisons are limited.
- Real-world evaluation is qualitative on smartphone captures and quantitative only on SIDD with simulated Quad Bayer sampling.
Experimental setup
Datasets: DIV2K Flickr2K SIDD Urban100 BSD100 MCM Kodak Set14
Benchmarks: Urban100 BSD100 MCM Kodak Set14 SIDD
Metrics: PSNR SSIM LPIPS inference time
Baselines: PyNet SGNet PIPNet Uformer SAGAN Restormer RSTCA
Builds on: RRDB Restormer PIPNet SGNet
Closest papers by meaning
- Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 2025 · similarity 0.83)
- Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.82)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.82)
- PIDSR: Complementary Polarized Image Demosaicing and Super-Resolution (CVPR 2025 · similarity 0.81)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.81)
- Flexible Residual Binarization for Image Super-Resolution (ICML 2024 · similarity 0.81)
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.