Research map / Image generation / Image restoration
DL2G: Degradation-guided Local-to-Global Restoration for Eyeglass Reflection Removal
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 DL2G, a three-stage eyeglass reflection removal framework combining a multiplicative degradation model, a local structure-aware diffusion model, and a global consistency refinement module, plus a 2,000-pair extension (ReyeR+) of the ReyeR dataset.
Problem
Eyeglass reflection removal must remove localized reflections, restore lost eye texture in strongly degraded regions, and keep color and illumination consistent with the input; existing additive/linear reflection models fit this poorly.
Approach
Model eyeglass reflection as a multiplicative degradation map (I = T ∘ D) to get a preliminary result, then apply a diffusion model trained with sliding-window local patch sampling to restore eye details, and finally fuse non-eye features from the input via a transformer-based refinement module for illumination consistency.
Key results
- On ReyeR+, DL2G reaches 34.801 PSNR, 0.947 SSIM and 0.040 LPIPS, the best of all compared methods (next best DSRNet at 33.555 PSNR, ER2Net at 0.050 LPIPS).
- On the SHIQ specular highlight dataset, DL2G achieves 28.145 PSNR and 0.946 SSIM, above TSHRNet (25.575 / 0.933) and JSHDR (34.131 PSNR but 0.860 SSIM).
- Ablation shows the full model (34.801 PSNR) improves over DEM+LSDM (34.754) and DEM+GCR (32.490), and the LSDM alone reaches 33.441 PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DEM, LSDM, and GCR rather than DEM alone (ablation) | DL2G · ReyeR+ | PSNR: 31.633 -> 34.801 | Helped |
| Use DEM, LSDM, and GCR rather than LSDM alone (ablation) | DL2G · ReyeR+ | PSNR: 33.441 -> 34.801 | Helped |
| Add GCR after DEM + LSDM (ablation) | DL2G · ReyeR+ | SSIM: 0.934 -> 0.947 | Helped |
| Add LSDM to DEM + GCR (ablation) | DL2G · ReyeR+ | PSNR: 32.490 -> 34.801 | Helped |
| Use DL2G rather than ER2Net | eyeglass reflection removal · ReyeR+ | PSNR: 34.301 -> 34.801 | Helped |
| Use DL2G rather than ER2Net | eyeglass reflection removal · ReyeR+ | SSIM: 0.935 -> 0.947 | Helped |
| Use DL2G rather than DSRNet | eyeglass reflection removal · ReyeR+ | LPIPS: 0.046 -> 0.040 | Helped |
| Use DL2G rather than JSHDR | specular highlight removal · SHIQ | PSNR: 34.131 -> 28.145 | Hurt |
| Use DL2G rather than TSHRNet | specular highlight removal · SHIQ | SSIM: 0.933 -> 0.946 | Helped |
Surprising findings
- TSHRNet scores only 20.257 PSNR on ReyeR+ despite a competitive SSIM of 0.879, suggesting a mismatch between its outputs and the eyeglass task.
- The additive reflection degradation models used in prior work were found to fit eyeglass reflections poorly, motivating the multiplicative formulation.
Limitations
- Fails on eyeglasses not worn on eyes, since training data focuses on eyes behind the glasses.
- Fails on eyeglasses with special materials where the front scene is fully projected, because the degradation model cannot separate the front scene from the background.
Experimental setup
Datasets: ReyeR ReyeR+ SHIQ
Benchmarks: ReyeR+ SHIQ
Metrics: PSNR SSIM LPIPS
Baselines: Watanabe et al. ER2Net IBCLN Robust SIRR DSRNet TSHRNet JSHDR SpecularityNet
Builds on: Retinex theory conditional diffusion model (DDPM) ReyeR dataset transformer image-to-image backbone (Restormer)
Closest papers by meaning
- Anti-Glare: Tightly Constrained Optimization for Eyeglass Reflection Removal (CVPR 2017 · similarity 0.81)
- Reflection Removal Using Low-Rank Matrix Completion (CVPR 2017 · similarity 0.81)
- Learning Where and What to Restore for Composite Image Restoration (NeurIPS 2026 · similarity 0.81)
- Learning Latent Transmission and Glare Maps for Lens Veiling Glare Removal (CVPR 2026 · similarity 0.80)
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.79)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.79)
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.