Research map / Architectures and efficiency / CNN design
DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables
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 DnLUT, a lookup-table color image denoiser combining a Pairwise Channel Mixer (RG/GB/BR pair LUTs) and an L-shaped non-overlapping rotation kernel that converts to 3D/4D LUTs for edge-device denoising.
Problem
Deep denoisers are too computationally and memory heavy for edge devices, while existing LUT-based methods process RGB channels independently and so miss inter-channel noise correlations needed for color denoising.
Approach
A Pairwise Channel Mixer that jointly processes RGB channel pairs (RG, GB, BR) with 1x2 depth-2 kernels followed by 1x1 convolutions, plus an L-shaped kernel whose rotations do not overlap pixels, so the network converts to compact 3D/4D LUTs with a larger effective receptive field.
Key results
- On CBSD68 at sigma=25, DnLUT reaches 29.88 dB CPSNR vs 28.56 dB for SPF-LUT, using about 17% of its storage.
- PCM plug-in adds roughly 1 dB CPSNR (e.g., +1.16 dB on CBSD68 for SPF-LUT) to existing LUT methods.
- DnLUT uses about 500 KB storage and 0.1% of DnCNN's energy, with about 20x faster runtime than DnCNN on a mobile CPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PCM to SR-LUT | SR-LUT · CBSD68, σ=25 | CPSNR (dB): 26.71 -> 28.04 (+1.33) | Helped |
| add PCM to MuLUT | MuLUT · CBSD68, σ=25 | CPSNR (dB): 28.11 -> 29.17 (+1.05) | Helped |
| add PCM to RC-LUT | RC-LUT · CBSD68, σ=25 | CPSNR (dB): 28.12 -> 29.22 (+1.10) | Helped |
| add PCM to SPF-LUT | SPF-LUT · CBSD68, σ=25 | CPSNR (dB): 28.56 -> 29.72 (+1.16) | Helped |
| use L-shaped kernel instead of 2 × 2 spatial-wise kernel (ablation) | DnLUT · CBSD68 | CPSNR (dB): 26.71 vs 26.63 | Hurt |
| use L-shaped kernel instead of 2 × 2 spatial-wise kernel (ablation) | DnLUT · Kodak | CPSNR (dB): 27.16 vs 27.05 | Hurt |
| use L-shaped kernel instead of 2 × 2 spatial-wise kernel (ablation) | DnLUT · McMaster | CPSNR (dB): 28.01 vs 27.96 | Hurt |
| combine L-shaped kernel and PCM (ablation) | DnLUT · CBSD68 | CPSNR (dB): 30.15 -> 31.41 | Helped |
| add channel indexing to L-shaped kernel and PCM (ablation) | DnLUT · CBSD68 | CPSNR (dB): 31.41 -> 31.47 | Helped |
| use PCM alone instead of spatial-wise L-shaped convolution (ablation) | DnLUT · CBSD68 | CPSNR (dB): 30.15 -> 30.21 | Helped |
Limitations
- Storage and runtime are evaluated on a single mobile platform (Snapdragon 8 Gen 2) with Java on Android
- Gaussian training uses independent per-channel noise, which may not match real noise statistics
- Real-world evaluation only on SIDD and DnD
Experimental setup
Datasets: DIV2K Flickr2K BSD400 WaterlooED SIDD
Benchmarks: CBSD68 Kodak24 Urban100 McMaster SIDD DnD
Metrics: CPSNR SSIM PSNR runtime energy cost storage
Baselines: SR-LUT MuLUT RC-LUT SPF-LUT CBM3D MC-WNNM DnCNN SwinIR
Builds on: SR-LUT MuLUT SPF-LUT RC-LUT DnCNN
Closest papers by meaning
- Universal Denoising Networks : A Novel CNN Architecture for Image Denoising (CVPR 2018 · similarity 0.88)
- Non-Local Color Image Denoising With Convolutional Neural Networks (CVPR 2017 · similarity 0.88)
- Toward Convolutional Blind Denoising of Real Photographs (CVPR 2019 · similarity 0.86)
- Robust And Interpretable Blind Image Denoising Via Bias-Free Convolutional Neural Networks (ICLR 2020 · similarity 0.86)
- DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · similarity 0.85)
- TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge (NeurIPS 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.