Research map / Architectures and efficiency / CNN design
Efficient Look-Up Table from Expanded Convolutional Network for Accelerating Image Super-resolution
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 ECLUT, a LUT-based single-image super-resolution method that uses an expanded convolution layer (output-side overlapping sliding-window summation) to enlarge the receptive field with a single LUT, plus output rescaling to reduce quantization error.
Problem
LUT-based super-resolution is limited by receptive field size, since LUT size grows exponentially with input pixels; existing multi-LUT methods are slow on edge devices.
Approach
Expand the convolution output to r×k×r×k so each query writes overlapping sliding-window outputs that are summed, enlarging the receptive field indirectly while the LUT grows linearly; a scaling factor (0.25) applied to the accumulated result reduces 8-bit quantization error.
Key results
- ECLUT achieves 41ms runtime for 1280x720 ×4 SR on a MEIZU 16s phone, faster than SR-LUT (94ms), SP-LUT (365ms) and MuLUT (242ms) but with lower PSNR than SP-LUT and MuLUT (e.g., Set5 29.91 vs 30.01 and 30.60).
- Rescaling accumulated output by 0.25 (2-bit right shift) recovers the quantization loss, giving Set5 PSNR 29.91 matching the no-quantization result.
- Variant EC-LUT-F (1×3 kernel) reaches Set5 30.24 dB and Urban100 24.18 dB, showing performance grows with more pixels per index.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SR-LUT with EC-LUT | LUT-based ×4 super-resolution · Set5 | PSNR: 29.82 -> 29.91 | Helped |
| replace SR-LUT with EC-LUT | LUT-based ×4 super-resolution · 1280×720 image generation | runtime: 94ms -> 41ms | Helped |
| use EC-LUT instead of Mu-LUT | LUT-based ×4 super-resolution · Set5 | PSNR: 30.60 -> 29.91 | Hurt |
| use EC-LUT instead of Mu-LUT | LUT-based ×4 super-resolution · 1280×720 image generation | runtime: 242ms -> 41ms | Helped |
| apply 2-bit right shift rescaling after quantization (ablation) | EC-LUT · Set5 | PSNR: 29.83 -> 29.91 | Helped |
| apply 1-bit right shift rescaling after quantization (ablation) | EC-LUT · Set5 | PSNR: 29.83 -> 29.91 | Helped |
| apply 3-bit right shift rescaling after quantization (ablation) | EC-LUT · Set5 | PSNR: 29.83 -> 29.82 | No clear effect |
| avoid quantizing the SR network output (ablation) | EC-LUT · Set5 | PSNR: 29.83 -> 29.91 | Helped |
| use a 1×3 kernel instead of a 1×2 kernel (ablation) | EC-LUT · Set5 | PSNR: 29.91 -> 30.24 | Helped |
| use a 2×2 kernel instead of a 1×2 kernel (ablation) | EC-LUT · Set5 | PSNR: 29.91 -> 30.35 | Helped |
Surprising findings
- Expanding the output rather than the input yields a LUT that grows linearly with receptive field, though the single-query design leaves it slightly behind multi-LUT methods in PSNR.
Limitations
- Only two pixels used for single indexing to keep speed, limiting indexing capability versus MuLUT.
- PSNR slightly inferior to SP-LUT and MuLUT on most benchmarks.
- LUT size (9MB) is larger than prior LUT methods such as SR-LUT.
Experimental setup
Datasets: DIV2K Set5 Set14 B100 Urban100 Manga109
Benchmarks: Set5 Set14 BSDS100 Urban100 Manga109
Metrics: PSNR SSIM runtime (ms) model size
Baselines: Nearest Bilinear Bicubic SR-LUT SP-LUT MuLUT FSRCNN CARN-M RRDB
Builds on: SR-LUT SP-LUT sub-pixel convolution rotational ensemble
Closest papers by meaning
- Practical Single-Image Super-Resolution Using Look-Up Table (CVPR 2021 · similarity 0.91)
- Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network (CVPR 2016 · similarity 0.87)
- TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge (NeurIPS 2024 · similarity 0.86)
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (CVPR 2024 · similarity 0.85)
- SAVSR: Arbitrary-Scale Video Super-Resolution via a Learned Scale-Adaptive Network (AAAI 2024 · similarity 0.85)
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning (CVPR 2025 · similarity 0.85)
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.