Research map / Image generation / Image restoration
Reversible Decoupling Network for Single Image 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 RDNet, a reversible multi-column encoder with a transmission-rate-aware prompt generator for single image reflection removal, reporting new PSNR records on five benchmarks.
Problem
Single image reflection removal loses high-level semantic information through layer-by-layer propagation and dual-stream interactions follow fixed patterns, while real-world reflection strength varies widely and training data is limited.
Approach
Replace lossy dual-stream interactions with a multi-column reversible encoder (RevCol-style, with bidirectional cross-level connections) that decouples transmission and reflection features losslessly, and modulate features with a prompt derived from an estimated transmission rate (channel-wise alpha/beta).
Key results
- Averaged over four real-world SIR2/Real20 subsets without extra Nature data, RDNet reaches 25.95 dB PSNR / 0.908 SSIM, above DSRNet's 25.40 / 0.905.
- With Nature data, RDNet reaches 26.65 dB PSNR / 0.917 SSIM on average, versus 25.75 dB for DSRNet and 25.34 dB for Zhu et al.
- Ablation: replacing reversible connections with a U-Net connection drops PSNR by 2.6 dB; a DSRNet-style dual-stream variant is 0.28 dB lower.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RDNet instead of DSRNet | single-image reflection removal · Real20 | PSNR: 24.23 -> 24.43 | Helped |
| Use RDNet instead of DSRNet | single-image reflection removal · Objects | PSNR: 26.28 -> 25.76 | Hurt |
| Use RDNet instead of DSRNet | single-image reflection removal · Postcard | PSNR: 24.56 -> 25.95 | Helped |
| Use RDNet instead of DSRNet | single-image reflection removal · Wild | PSNR: 25.68 -> 27.20 | Helped |
| Use RDNet instead of the strongest listed baseline | single-image reflection removal with Nat. training data · Real20, Objects, Postcard, Wild | PSNR: 23.91 -> 25.58; 26.74 -> 26.78; 24.83 -> 26.33; 26.49 -> 27.70 | Helped |
| Add transmission-rate-aware techniques (ablation) | RDNet · Real20, Objects, Postcard, Wild average | PSNR: 25.52 -> 26.65 | Helped |
| Adjust the input using the estimated transmission rate without a prompt (ablation) | RDNet · Real20, Objects, Postcard, Wild average | PSNR: 25.52 -> 25.99 | Helped |
| Add a transmission-rate-aware prompt while adjusting the input (ablation) | RDNet · Real20, Objects, Postcard, Wild average | PSNR: 25.99 -> 26.03 | No clear effect |
| Use a transmission-rate-aware prompt without input adjustment (ablation) | RDNet · Real20, Objects, Postcard, Wild average | PSNR: 26.03 -> 26.65 | Helped |
| Use a single-stream decoupling design instead of dual-stream (ablation) | RDNet · Real20, Objects, Postcard, Wild average | PSNR: 26.37 -> 26.65 | Helped |
| Add reflection loss (ablation) | RDNet · Real20, Objects, Postcard, Wild average | PSNR: 25.99 -> 26.65 | Helped |
| Use reversible connections instead of U-Net connections (ablation) | RDNet · Real20, Objects, Postcard, Wild average | PSNR: 24.05 -> 26.65 | Helped |
Surprising findings
- Adjusting the input image with estimated transmission rates performed worse than modulating features with the prompt, since input-level corrections were hard to recover from when estimates were wrong.
Limitations
- Performance peaks at 4 columns; 2 and 6 columns give lower PSNR, so the design needs tuning.
- Two-stage training depends on a separately trained transmission-rate estimator.
- SSIM falls short of Zhu et al. on the Objects subset and the Nature comparison SSIM is marginally lower.
Experimental setup
Datasets: Real20 PASCAL VOC Nature SIR2 (Objects, Postcard, Wild)
Benchmarks: Real20 Objects Postcard Wild Nature
Metrics: PSNR SSIM
Baselines: ERRNet IBCLN RAGNet Dong et al. YTMT DSRNet Zhu et al.
Builds on: RevCol GLOM DSRNet YTMT FocalNet
Closest papers by meaning
- Single Image Reflection Removal Exploiting Misaligned Training Data and Network Enhancements (CVPR 2019 · similarity 0.88)
- The Reversible Residual Network: Backpropagation Without Storing Activations (NeurIPS 2017 · similarity 0.86)
- Single Image Reflection Separation via Dual-Stream Interactive Transformers (NeurIPS 2024 · similarity 0.84)
- Dr2Net: Dynamic Reversible Dual-Residual Networks for Memory-Efficient Finetuning (CVPR 2024 · similarity 0.83)
- FIRM: Flexible Interactive Reflection ReMoval (AAAI 2025 · similarity 0.83)
- CoDe: An Explicit Content Decoupling Framework for Image Restoration (CVPR 2024 · similarity 0.83)
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.