Research map / Image generation / Image restoration
Progressive High-Frequency Reconstruction for Pan-Sharpening with Implicit Neural Representation
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 PIF-Net, a progressive high-frequency pan-sharpening network combining a band-limited multi-scale high-frequency generator, a spatial-spectral implicit image function, and progressive high-frequency injection, trained from low to high frequency.
Problem
Pan-sharpening requires restoring high-frequency edge details in multispectral images, but DNNs trained with spectral bias fit low frequencies first and misalign multi-scale features when fused in discrete feature spaces.
Approach
Treats pan-sharpening as coarse-to-fine high-frequency restoration: band-limited sinusoidal signals from the PAN gradient are injected progressively across training stages, and spatial and spectral features are fused with an implicit neural representation (SIIF) that decodes pixels at continuous coordinates to avoid misalignment.
Key results
- PSNR of 42.0635 dB on WorldView-II, improving on the second-best by 0.13 dB; 30.9309 dB on WorldView-III (+0.32 dB); 49.1388 dB on GaoFen2 (+1.66 dB).
- Ablation on WV2: removing PHIM costs about 0.11 dB PSNR, and replacing SIIF with a CNN reduces PSNR from 42.06 to 41.77 dB.
- On full-resolution real GaoFen2 scenes, QNR of 0.8476 is the best among compared methods (second best SFIIN 0.8466).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PIF-Net instead of the strongest listed baseline, MDCUN | Pan-sharpening · WorldView-II | PSNR: 41.9269 -> 42.0635 | Helped |
| Use PIF-Net instead of the strongest listed baseline, Band Aware | Pan-sharpening · WorldView-III | PSNR: 30.6050 -> 30.9309 | Helped |
| Use PIF-Net instead of the strongest listed baseline, SFIIN | Pan-sharpening · GaoFen2 | PSNR: 47.4712 -> 49.1388 | Helped |
| Replace BMHG with a dilated network (ablation) | PIF-Net · WorldView-II ablation | PSNR: 41.3760 -> 42.0635 | Hurt |
| Replace SIIF with CNN (ablation) | PIF-Net · WorldView-II ablation | PSNR: 41.7689 -> 42.0635 | Hurt |
| Remove PHIM and fuse only full-bandwidth high-frequency signals (ablation) | PIF-Net · WorldView-II ablation | PSNR: 41.9474 -> 42.0635 | Hurt |
Limitations
- Evaluated only on three satellite datasets with reduced-resolution protocol and 200 real full-resolution scenes without ground truth.
- Ablation reports modest gains for individual modules (about 0.1 dB for PHIM).
Experimental setup
Datasets: WorldView-II WorldView-III GaoFen2
Benchmarks: WorldView-II WorldView-III GaoFen2 full-resolution real GaoFen2 scenes
Metrics: PSNR SSIM SAM ERGAS Dλ DS QNR
Baselines: SFIM Brovey GS IHS GFPCA BAM SFIIN GCPNet Band Aware MDCUN
Builds on: SIREN (sinusoidal implicit representations) LIIF-style local implicit image function Wald protocol Lindell et al. 2022 band-limited INR
Closest papers by meaning
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · similarity 0.90)
- Unsupervised Pan-Sharpening via Mutually Guided Detail Restoration (AAAI 2024 · similarity 0.89)
- Frequency-Adaptive Pan-Sharpening with Mixture of Experts (AAAI 2024 · similarity 0.88)
- Deep Unfolded Network with Intrinsic Supervision for Pan-Sharpening (AAAI 2024 · similarity 0.87)
- Binarized Neural Network for Multi-spectral Image Fusion (CVPR 2025 · similarity 0.86)
- SSUN-Net: Spatial-Spectral Prior-Aware Unfolding Network for Pan-Sharpening (AAAI 2025 · 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.