Research map / Image generation / Image restoration
Real-World Mobile Image Denoising Dataset with Efficient Baselines
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.
Introduces MIDD, a large real-world mobile image denoising dataset (~400K noisy/clean pairs from 20 phone sensors), a DPreview test set from 294 cameras, and the efficient SplitterNet U-Net-style baseline that runs on mobile GPUs/NPUs.
Problem
Mobile image denoising lacks a sufficiently diverse real-world training dataset, and existing efficient denoising models are either too slow or not validated on mobile hardware across many sensors.
Approach
Capture bursts of 20 RAW Bayer frames per static scene across 20 mobile sensors, use the per-pixel mean of the burst as ground truth, and design a split-tensor U-Net (SplitterNet) optimized for mobile accelerator latency and PSNR.
Key results
- SplitterNet reaches 37.92 dB PSNR and 0.884 SSIM on MIDD at 27.7 ms on Exynos 2200 GPU, beating NOAHTCV (37.81 dB, 30.4 ms) and Megvii (38.05 dB, 56.8 ms in the trade-off).
- Trained on MIDD, SplitterNet scores 21.75 dB on DPreview, 0.62 dB above the SIDD-trained model (21.13 dB).
- SplitterNet processes 8MP images in under one second on all listed mobile GPUs and reaches over 30 FPS at HD on some NPUs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train SplitterNet on MIDD instead of SIDD | SplitterNet · DPreview | PSNR: 21.13 -> 21.75 | Helped |
| Train SplitterNet on MIDD instead of MAI2021 | SplitterNet · DPreview | PSNR: 20.55 -> 21.75 | Helped |
| Train SplitterNet on MIDD instead of SIDD | SplitterNet · MIDD | PSNR: 34.03 -> 37.92 | Helped |
| Train SplitterNet on MIDD instead of MAI2021 | SplitterNet · MIDD | PSNR: 34.69 -> 37.92 | Helped |
| Train SplitterNet on MIDD instead of MAI2021 | SplitterNet · MAI2021 | PSNR: 37.07 -> 37.23 | Helped |
| Train SplitterNet on MIDD and fine-tune on the target dataset instead of training on that dataset alone | SplitterNet · SIDD | PSNR: 36.77 -> 38.79 | Helped |
| Train SplitterNet on MIDD and fine-tune on the target dataset instead of training on that dataset alone | SplitterNet · MAI2021 | PSNR: 37.07 -> 37.62 | Helped |
| Replace NOAHTCV with SplitterNet | Mobile image denoising on MIDD · MIDD | PSNR: 37.81 -> 37.92 | Helped |
| Replace NOAHTCV with SplitterNet | Mobile image denoising on MIDD · MIDD | Runtime: 30.4 -> 27.7 | Helped |
| Replace Megvii with SplitterNet | Mobile image denoising on MIDD · MIDD | PSNR: 38.05 -> 37.92 | Hurt |
| Replace Megvii with SplitterNet | Mobile image denoising on MIDD · MIDD | Runtime: 56.8 -> 27.7 | Helped |
| Train SplitterNet on MIDD instead of SIDD | SplitterNet denoising DPreview images · DPreview | visual denoising quality | Helped |
Surprising findings
- Fine-tuning and testing on MAI2021 yields a MIDD-trained model with lower PSNR than the MAI2021-trained one on MAI2021 test data in some settings, while MIDD-pretrained models perform best after fine-tuning on SIDD (38.79 dB).
Limitations
- Ground truth is the mean of 20 noisy burst frames, so residual blur from motion in scenes is possible.
- DPreview test pairs are aligned via a sliding-window method and use the lowest ISO image as ground truth, which may be imperfect.
- Only RGB-to-RGB denoising on the MIDD test split is reported with a single model architecture; comparisons with other models are limited to the MAI2021 winners and a few baselines.
Experimental setup
Datasets: MIDD SIDD MAI2021 DPreview
Benchmarks: MIDD DPreview SIDD MAI2021
Metrics: PSNR SSIM runtime (ms)
Baselines: NAFNet U-Net ResNet18 NOAHTCV Megvii SIDD-trained SplitterNet MAI2021-trained SplitterNet
Builds on: NAFNet U-Net MAI2021 challenge
Closest papers by meaning
- A High-Quality Denoising Dataset for Smartphone Cameras (CVPR 2018 · similarity 0.92)
- NERDS: A General Framework to Train Camera Denoisers from Raw-RGB Noisy Image Pairs (ICLR 2023 · similarity 0.88)
- Joint HDR Denoising and Fusion: A Real-World Mobile HDR Image Dataset (CVPR 2023 · similarity 0.88)
- Efficient Real-Time Raw-to-Raw Denoising for Extreme Low-Light Ultra HD Video on Mobile Devices (CVPR 2026 · similarity 0.87)
- IDR: Self-Supervised Image Denoising via Iterative Data Refinement (CVPR 2022 · similarity 0.86)
- Toward Convolutional Blind Denoising of Real Photographs (CVPR 2019 · similarity 0.86)
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.