Research map / Image generation / Image restoration
Removing Interference and Recovering Content Imaginatively for Visible Watermark 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 RIRCI, a two-stage visible watermark removal framework that first separates the watermark component and then restores background with a dual-path network built on a Global and Local Context Interaction module.
Problem
Visible watermark removal often leaves residual watermarks because watermark removal and background restoration share one branch, and existing methods fail when watermarks heavily obscure the background.
Approach
Decouples watermark component exclusion (stage 1, U-Net with dual decoders for mask and component) from background restoration (stage 2), where a dual-path network combines intrinsic background recovery and inpainting-style content imagination, fused by non-local blocks; the GLCI module combines local/global MLPs with spectral transform and scSE for bidirectional feature propagation.
Key results
- On HWVOC, RIRCI reaches 39.52 PSNR and 0.9855 SSIM, versus 38.72 PSNR for the next best (SplitNet).
- On PW, RIRCI reaches 41.89 PSNR and 0.9866 SSIM, versus 40.35 PSNR for SplitNet.
- Ablation: replacing GLCI with 3x3 convolutions drops HWVOC PSNR from 39.52 to 39.82-free variant #1 at 38.42 ... (ablation table shows lower PSNR for all variants than the full model).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Separate stage-1 prediction of the background component from watermark-free image prediction (ablation) | RIRCI · HWVOC | PSNR: 38.74 -> 39.52 | Helped |
| Use GLCI instead of FFC (ablation) | RIRCI · HWVOC | PSNR: 38.34 -> 39.52 | Helped |
| Add the content imagination sub-network alongside content restoration (ablation) | RIRCI · HWVOC | PSNR: 39.15 -> 39.52 | Helped |
| Add the content restoration sub-network alongside content imagination (ablation) | RIRCI · HWVOC | PSNR: 38.78 -> 39.52 | Helped |
| Use GLCI instead of Conv 3×3 (ablation) | RIRCI · HWVOC | PSNR: 38.42 -> 39.52 | Helped |
| Use scSE and spectral transform instead of Conv 3×3 for both GLCI feature transformations (ablation) | RIRCI · HWVOC | PSNR: 38.79 -> 39.52 | Helped |
| Use scSE instead of Conv 3×3 for global-to-local feature propagation (ablation) | RIRCI · HWVOC | PSNR: 38.66 -> 39.52 | Helped |
| Use spectral transform instead of Conv 3×3 for local-to-global feature propagation (ablation) | RIRCI · HWVOC | PSNR: 39.03 -> 39.52 | Helped |
| Predict the background component instead of the watermark-free image in stage 1 (ablation) | RIRCI · HWVOC | F1: 0.8576 -> 0.8802 | Helped |
| Use RIRCI rather than the strongest reported baseline | visible watermark removal models · PW | RMSE_w: 12.9250 -> 13.5077 | Hurt |
Surprising findings
- Predicting the watermark-free image directly in stage 1 instead of the background component reduces watermark segmentation F1 (0.8576 vs 0.8802), suggesting conflict between recovery and segmentation.
Limitations
- Relies on the watermark being composited with a known opacity model; generalization to real-world watermarks is only partially tested via cross-dataset validation.
- Cross-dataset results are lower than in-domain results on HWVOC when trained on PW.
Experimental setup
Datasets: PASCAL VOC2012 HWVOC PW
Benchmarks: HWVOC PW
Metrics: PSNR SSIM RMSE RMSE_w F1 IoU
Baselines: WDNet SplitNet SLBR LaMa
Builds on: SLBR LaMa FFC scSE
Closest papers by meaning
- Bridging Knowledge Gap Between Image Inpainting and Large-Area Visible Watermark Removal (AAAI 2025 · similarity 0.88)
- Image Watermarks are Removable using Controllable Regeneration from Clean Noise (ICLR 2025 · similarity 0.85)
- Invisible Image Watermarks Are Provably Removable Using Generative AI (NeurIPS 2024 · similarity 0.84)
- WMarkGPT: Watermarked Image Understanding via Multimodal Large Language Models (ICML 2025 · similarity 0.83)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.83)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · 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.