Research map / Image generation / Image restoration
Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up 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 Pan-LUT, a network-free learnable look-up table framework (PGLUT, SDLUT, AOLUT) for pan-sharpening that processes very large remote sensing images with under 700K parameters and sub-millisecond GPU inference.
Problem
Deep learning pan-sharpening methods are computationally expensive at inference, scale poorly with image size, and depend on GPUs, limiting real-time processing of large remote sensing images.
Approach
Replaces deep network operations with three learnable LUTs: a PAN-guided 5D spectral LUT with pentalinear interpolation, a 4D spatial-detail LUT with rotation-based indexing and quadrilinear interpolation, and a 5D adaptive output LUT for channel aggregation, trained end-to-end with smoothness and monotonicity regularizers.
Key results
- On WV2 reduced-resolution, Pan-LUT reaches 40.86 dB PSNR with 0.66M parameters and 0.38 ms at 2Kx2K, versus about 41-42 dB for the strongest DNN baselines that are far slower.
- On WV2 full-resolution real-world scenes, Pan-LUT attains D_lambda 0.0571, D_S 0.0640, and QNR 0.8829, outperforming all compared DNN methods on these metrics.
- Pan-LUT processes 9Kx9K images on an 11GB GPU and 15Kx15K on a 24GB GPU, while most DNN baselines run out of memory at 4Kx4K or below.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Pan-LUT instead of Pan-Mamba | Pan-sharpening · WorldView-II reduced-resolution | PSNR: 42.2354 -> 40.8555 | Hurt |
| Use Pan-LUT instead of TA-DiffHQP | Pan-sharpening · GaoFen2 reduced-resolution | PSNR: 47.7716 -> 43.7466 | Hurt |
| Use Pan-LUT instead of TA-DiffHQP | Pan-sharpening · WorldView-III reduced-resolution | PSNR: 31.3369 -> 29.7376 | Hurt |
| Use Pan-LUT instead of TA-DiffHQP | Pan-sharpening · WorldView-II full-resolution | QNR: 0.8025 -> 0.8829 | Helped |
| Add SDLUT to the method (ablation) | Pan-LUT · WorldView-III reduced-resolution | performance | Helped |
| Add PGLUT to the method (ablation) | Pan-LUT · WorldView-III reduced-resolution | performance | Helped |
| Add AOLUT to the method (ablation) | Pan-LUT · WorldView-III reduced-resolution | performance | Helped |
| Increase PGLUT size from 9 to 17 (ablation) | Pan-LUT · WorldView-III reduced-resolution | PSNR: 0.01 dB | Helped |
| Increase SDLUT size from 3 to 9 (ablation) | Pan-LUT · WorldView-III reduced-resolution | performance | Helped |
| Increase SDLUT size above 9 (ablation) | Pan-LUT · WorldView-III reduced-resolution | performance | Hurt |
| Increase AOLUT size beyond 9 (ablation) | Pan-LUT · WorldView-III reduced-resolution | performance | Helped |
| Set monotonicity weight λ_m to 10 instead of 0 (ablation) | Pan-LUT · WorldView-III reduced-resolution | PSNR: 29.5321 -> 29.7376 | Helped |
Surprising findings
- Pan-LUT falls short of most DNN methods on reduced-resolution PSNR yet outperforms them on full-resolution D_lambda, D_S and QNR, suggesting better real-world generalization.
- Enlarging LUT size beyond 9 bins yields only minor PSNR gains while parameters grow from 236K to 5M, indicating capacity redundancy.
Limitations
- LUT storage grows exponentially with input dimension (O(N D^n)), limiting high-dimensional inputs.
- Reduced-resolution PSNR is lower than most DNN-based methods.
- No neural network component, so modeling capacity is limited relative to hybrid designs.
Experimental setup
Datasets: WorldView-II GaoFen2 WorldView-III
Benchmarks: WorldView-II reduced-resolution GaoFen2 reduced-resolution WorldView-III reduced-resolution WorldView-II full-resolution real-world scenes
Metrics: PSNR SSIM SAM ERGAS D_lambda D_S QNR inference time (ms) parameter count
Baselines: Brovey IHS SFIM GS PNN PanNet MSDCNN Pan-GAN GPPNN SFDI UCGAN PanFlow
Builds on: SRLUT MuLUT image-adaptive 3D LUT
Closest papers by meaning
- Unsupervised Pan-Sharpening via Mutually Guided Detail Restoration (AAAI 2024 · similarity 0.87)
- SSUN-Net: Spatial-Spectral Prior-Aware Unfolding Network for Pan-Sharpening (AAAI 2025 · similarity 0.86)
- Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · similarity 0.86)
- Frequency-Adaptive Pan-Sharpening with Mixture of Experts (AAAI 2024 · similarity 0.86)
- SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening (NeurIPS 2024 · similarity 0.85)
- A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (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.