Research map / Image generation / Super-resolution
KeDuSR: Real-World Dual-Lens Super-Resolution via Kernel-Free Matching
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 KeDuSR, a dual-lens super-resolution network that aligns the telephoto reference via global+local warping and uses kernel-free matching (LR-corner queries vs LR-center keys) to warp reference details to the corner region, plus a new well-aligned real dataset DuSR-Real.
Problem
Dual-lens (wide-angle LR, telephoto Ref) super-resolution suffers from the resolution gap between LR and Ref when matching densely, which misses best-matched patches and breaks structure consistency; existing datasets also lack aligned real LR-Ref-HR triples.
Approach
Align the Ref to the LR center region with global homography plus flow-guided DCN local warping, then match LR corner patches (queries) against LR center patches (keys) in the LR domain, so the matching avoids the LR-Ref resolution gap and the matched index is used to warp the aligned Ref to the corner.
Key results
- On DuSR-Real, KeDuSR-ℓ reaches 27.66 dB PSNR / 0.8890 SSIM full-image, vs. 26.77 dB for the second-best DCSR-ℓ.
- On CameraFusion-Real, KeDuSR-ℓ reaches 27.53 dB PSNR vs. 26.16 dB for ZeDuSR-ℓ and 26.02 dB for DCSR-ℓ.
- In cross-dataset generalization trained on DuSR-Real, KeDuSR-ℓ achieves 26.55 dB on RealMCVSR-Real vs. 25.46 dB for DCSR-ℓ.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KeDuSR-ℓ instead of DCSR-ℓ | Dual-lens SR · DuSR-Real, full image | PSNR: 26.77 -> 27.66 | Helped |
| use KeDuSR-ℓ instead of DCSR-ℓ | Dual-lens SR · RealMCVSR-Real, full image | PSNR: 26.28 -> 27.05 | Helped |
| use KeDuSR-ℓ instead of ZeDuSR-ℓ | Dual-lens SR · CameraFusion-Real, full image | PSNR: 26.16 -> 27.53 | Helped |
| replace traditional dense feature matching with center and corner warping (ablation) | KeDuSR · DuSR-Real | PSNR: 27.03 -> 27.66 | Helped |
| add corner warping (ablation) | KeDuSR with center warping · DuSR-Real | PSNR: 27.16 -> 27.66 | Helped |
| add center warping (ablation) | KeDuSR with corner warping · DuSR-Real | PSNR: 27.50 -> 27.66 | Helped |
| use adaptive SIFT-based alignment instead of a fixed homography (ablation) | KeDuSR global warping · DuSR-Real | PSNR: 27.51 -> 27.66 | Helped |
Surprising findings
- Finetuning DCSR with self-supervised real-image adaptation (DCSR-SRA) performs far worse than DCSR trained directly on the real pairs.
- The perceptual/adversarial hybrid-loss variant KeDuSR (27.18 dB) scores lower in PSNR than the reconstruction-only variant KeDuSR-ℓ (27.66 dB), though it has better LPIPS.
Limitations
- Corner-region matching is affected by dataset misalignment; some CameraFusion pairs still have small misalignments due to lens switching.
- Large LR inputs must be split into 128x128 patches for memory, which can cause blocking artifacts in other methods (claimed avoided by KeDuSR).
- Dataset is small (420 training triples, 305 scenes) and is built with manual removal of misaligned triples.
Experimental setup
Datasets: DuSR-Real RealMCVSR-Real CameraFusion-Real
Benchmarks: DuSR-Real RealMCVSR-Real CameraFusion-Real
Metrics: PSNR SSIM LPIPS latency
Baselines: RCAN SwinIR ESRGAN BSRGAN TTSR MASA DATSR DCSR SelfDZSR ZeDuSR
Builds on: DCSR SelfDZSR ZeDuSR TTSR MASA SpyNet
Closest papers by meaning
- Zero-Shot Dual-Lens Super-Resolution (CVPR 2023 · similarity 0.89)
- Camera Lens Super-Resolution (CVPR 2019 · similarity 0.87)
- Robust Reference-Based Super-Resolution via C2-Matching (CVPR 2021 · similarity 0.86)
- Robust Reference-Based Super-Resolution With Similarity-Aware Deformable Convolution (CVPR 2020 · similarity 0.85)
- Continuous Optical Zooming: A Benchmark for Arbitrary-Scale Image Super-Resolution in Real World (CVPR 2024 · similarity 0.84)
- Sequence Matters: Harnessing Video Models in 3D Super-Resolution (AAAI 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.