Research map / Architectures and efficiency / CNN design
TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge
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 TinyLUT, a LUT-based image restoration framework that uses a separable mapping strategy and dynamic discretization mechanism to cut LUT storage by orders of magnitude while improving super-resolution and denoising accuracy on edge devices.
Problem
LUT-based image restoration suffers exponential LUT storage growth with convolution kernel size, which blocks deployment of CNN-mapped LUTs on resource-constrained edge devices.
Approach
Separable mapping strategy decomposes n×n convolutions and pointwise convolutions into many 1×1 sub-operations mapped to 1D LUTs whose outputs are averaged, turning storage from exponential to linear in kernel size; dynamic discretization splits 8-bit activations into learnable-clipped MSB and LSB branches to further shrink LUTs.</303>
Key results
- TinyLUT-F reaches 31.18 dB PSNR on Set5 (x4) at 171KB, about 0.58 dB above MuLUT-SDY-X2 with roughly 24x less storage.
- TinyLUT-S uses 37KB and gets 27.13 average PSNR/SSIM 0.7710 across five SISR sets, beating SRLUT-S by 0.4 dB on Set5.
- SMS reduces LUT size from 1.274MB (uniform sampling) to 164KB in TinyLUT-S; DDM further reduces it to 37KB with negligible accuracy loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add separable mapping strategy (SMS) (ablation) | TinyLUT-S · Set5, ×4 SISR | PSNR: 29.82 -> 30.24 | Helped |
| Add separable mapping strategy (SMS) (ablation) | TinyLUT-S · Set5, ×4 SISR | LUT size: 1.274MB -> 164KB | Helped |
| Add dynamic discretization mechanism (DDM) (ablation) | TinyLUT-S · Set5, ×4 SISR | PSNR: 30.24 -> 30.22 | No clear effect |
| Add dynamic discretization mechanism (DDM) (ablation) | TinyLUT-S · Set5, ×4 SISR | LUT size: 164KB -> 37KB | Helped |
| Use TinyLUT-F instead of MuLUT-SDY-X2 | ×4 SISR · Set5 | PSNR: 30.60 -> 31.18 | Helped |
| Use TinyLUT-F instead of MuLUT-SDY-X2 | ×4 SISR · Set5 | storage: 4159KB -> 171KB | Helped |
| Use TinyLUT-S instead of SRLUT-S | ×4 SISR · Set5 | PSNR: 29.82 -> 30.22 | Helped |
| Use TinyLUT-F instead of FSRCNN | ×4 SISR on Raspberry 4B · SISR | runtime: 2143ms -> 387ms | Helped |
| Use TinyLUT-F instead of SPF-LUT+DFC | ×4 SISR · Set5 | storage: 2066KB -> 171KB | Helped |
| Use TinyLUT-F instead of DnCNN | image denoising · Set12 and BSD68, noise levels 15/25/50 | average PSNR: 29.61 -> 28.93 | Hurt |
| Use TinyLUT-F instead of ARCNN | image deblocking · Classic5, JPEG quality factor 10 | PSNR-B: 28.76 -> 28.74 | Hurt |
| Use TinyLUT-F instead of ARCNN | image deblocking · LIVE1, JPEG quality factor 10 | PSNR-B: 28.77 -> 28.67 | Hurt |
Surprising findings
- TinyLUT-S achieves 0.4 dB higher Set5 PSNR than SRLUT-S while using 37KB storage versus 1304KB.
Limitations
- Mapping Transformer and Mamba architectures into TinyLUT is difficult.
- Evaluated only on SISR, denoising and deblocking; other restoration tasks left for future work.
Experimental setup
Datasets: DIV2K
Benchmarks: Set5 Set14 Urban100 BSD100 Manga109 Set12 BSD68 Classic5 LIVE1
Metrics: PSNR SSIM PSNR-B latency storage
Baselines: SRLUT MuLUT SPLUT RCLUT SPF-LUT FSRCNN VDSR CARN-M DnCNN SRCNN
Builds on: SRLUT MuLUT SPLUT DnCNN depthwise separable convolutions
Closest papers by meaning
- Look-Up Table Compression for Efficient Image Restoration (CVPR 2024 · similarity 0.87)
- Efficient Look-Up Table from Expanded Convolutional Network for Accelerating Image Super-resolution (AAAI 2024 · similarity 0.86)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.85)
- Omni-Kernel Network for Image Restoration (AAAI 2024 · similarity 0.85)
- Improving the Learning Capability of Small-size Image Restoration Network by Deep Fourier Shifting (NeurIPS 2024 · similarity 0.84)
- DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables (CVPR 2025 · 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.