Research map / Image generation / Image restoration
Devignet: High-Resolution Vignetting Removal via a Dual Aggregated Fusion Transformer with Adaptive Channel Expansion
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.
Introduces VigSet, a 983-pair high-resolution real-world vignetting dataset with ground truth, and DeVigNet, a Laplacian-pyramid frequency-aware Transformer (DAFT for low frequency, ACEM for high frequency) for vignetting removal.
Problem
Vignetting removal in high-resolution images relies on physics assumptions and hand-tuned parameters, and there is a lack of real-world paired vignetting datasets with ground truth for objective evaluation.
Approach
Decomposes images with a Laplacian pyramid, removes low-frequency vignetting with a Dual Aggregated Fusion Transformer and enhances high-frequency details with an Adaptive Channel Expansion Module, trained as the first learning-based high-resolution vignetting removal model.
Key results
- On VigSet at 512x512, DeVigNet reaches 22.96 PSNR, 0.79 SSIM, 15.82 MAE, 0.09 LPIPS, beating Uformer (20.95 PSNR, 21.32 MAE).
- On LOL, DeVigNet achieves 21.33 PSNR vs. 18.55 for Uformer; on MIT-Adobe FiveK it reaches 23.10 PSNR vs. 21.92 for Uformer.
- Ablations show removing DAFT drops 512x512 PSNR from 22.96 to 18.25 and removing ACEM drops it to 19.66.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ACEM for high-frequency detail reconstruction (ablation) | DeVigNet · VigSet 512×512 | PSNR: 19.66 -> 22.96 | Helped |
| Add DAFT for low-frequency global feature processing (ablation) | DeVigNet · VigSet 512×512 | PSNR: 18.25 -> 22.96 | Helped |
| Set Laplacian pyramid depth to 3 (ablation) | DeVigNet · VigSet 512×512 | PSNR: 22.96 -> 16.04 | Hurt |
| Set Laplacian pyramid depth to 4 (ablation) | DeVigNet · VigSet 512×512 | PSNR: 22.96 -> 13.94 | Hurt |
| Replace DAFT with DHAN (ablation) | DeVigNet · VigSet 512×512 | PSNR: 22.96 -> 21.84 | Hurt |
| Use DeVigNet instead of Uformer | VigSet 512×512 · VigSet 512×512 | PSNR: 20.95 -> 22.96 | Helped |
| Use DeVigNet instead of Uformer | VigSet 1024×1024 · VigSet 1024×1024 | PSNR: 20.60 -> 22.94 | Helped |
| Use DeVigNet instead of Uformer | VigSet 2048×2048 · VigSet 2048×2048 | PSNR: 20.67 -> 22.94 | Helped |
Surprising findings
- Uformer and some LLIE methods achieve competitive or better LPIPS/SSIM on some datasets, e.g. Uformer has best LPIPS 0.06 on MIT-Adobe FiveK, above DeVigNet's 0.16.
Limitations
- Training only at 512x512 with testing at larger resolutions; performance at high resolution depends on generalization.
- Dataset limited to indoor scenes captured with three devices and ND filters.
- Mixed results on LPIPS compared with some baselines.
Experimental setup
Datasets: VigSet MIT-Adobe FiveK LOL-v1
Benchmarks: VigSet LOL-v1 MIT-Adobe FiveK
Metrics: PSNR SSIM MAE LPIPS
Baselines: SIVC RIVC LIME MSRCR NPE WV-SRIE PM-SRIE JieP RetinexNet KID DSLR ELGAN
Builds on: Uformer DHAN NAFNet
Closest papers by meaning
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.85)
- RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer (NeurIPS 2026 · similarity 0.82)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.81)
- LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 2024 · similarity 0.81)
- AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.81)
- VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook (AAAI 2024 · similarity 0.81)
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.