Research map / Image generation / Image restoration
Bridging Knowledge Gap Between Image Inpainting and Large-Area 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.
Adapts a pre-trained LaMa inpainting model with dual-branch residual-background features fused via gated fusion modules, and trains with coarse masks so it is robust to mask quality for large-area visible watermark removal.
Problem
Visible watermark removal struggles with large-area watermarks and depends heavily on the quality of predicted watermark masks.
Approach
Fuse residual background features from a watermark component cleaning branch and a background content embedding branch (both with transposed attention) into a LaMa inpainting backbone through gated fusion modules, and train with randomly eroded/dilated coarse masks so that inference only needs a coarse mask.
Key results
- On ILAW with coarse masks, the proposed method reaches 26.81 PSNR / 0.924 SSIM / 0.094 LPIPS, versus 26.13 PSNR for the best baseline (SLBR).
- In a user study of 46 participants on 27 real-world images, the method received 882 votes versus 206 for SLBR and 154 for SplitNet.
- Blind removal with none/white masks gives 25.77 PSNR and 0.100 LPIPS, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add feature adaptation branches to LaMa (ablation) | LaMa · ILAW | PSNR: 23.62 -> 26.81 | Helped |
| Use dual WCC+BCE branches instead of BCE-only adaptation (ablation) | LaMa · ILAW | PSNR: 25.67 -> 26.81 | Helped |
| Use pretrained LaMa parameters (ablation) | Ours · ILAW, fixed mask | PSNR: 25.99 -> 26.81 | Helped |
| Train with augmented coarse masks (ablation) | Ours · ILAW, fixed mask | PSNR: 26.38 -> 26.81 | Helped |
| Use 3 TA blocks in WCC+BCE instead of 2 (ablation) | LaMa feature adaptation · ILAW | PSNR: 23.89 -> 26.81 | Helped |
| Use 3 TA blocks in WCC+BCE instead of 6 (ablation) | LaMa feature adaptation · ILAW | PSNR: 22.82 -> 26.81 | Helped |
| Use transposed attention instead of convolutional feature extraction (ablation) | LaMa feature adaptation with GFM · ILAW | PSNR: 24.04 -> 26.81 | Helped |
| Use transposed attention instead of conventional attention (ablation) | LaMa feature adaptation with GFM · ILAW | PSNR: 23.50 -> 26.81 | Helped |
| Use GFM instead of convolution for feature fusion (ablation) | LaMa with transposed attention · ILAW | PSNR: 26.02 -> 26.81 | Helped |
| Use the proposed method rather than SLBR | Watermark removal · ILAW, fixed and coarser mask | PSNR: 26.09 -> 26.66 | Helped |
| Use the proposed method rather than SLBR | Watermark removal · Real-world dataset | votes: 206 -> 882 | Helped |
Surprising findings
- Removing the pretrained LaMa initialization leaves residual watermarks, but the paper notes that LaMa alone without feature adaptation scores only 17.97 PSNR under fixed masks.
Limitations
- Real-world evaluation uses only 27 images and a user study with 46 participants.
- Relies on coarse masks supplied manually or by a segmentation model at test time.
Experimental setup
Datasets: ILAW (synthesized, Places365 backgrounds) Places365 Challenge
Benchmarks: ILAW real-world dataset (27 high-resolution watermarked images)
Metrics: PSNR SSIM RMSE RMSE_w LPIPS user study votes
Baselines: WDNet SplitNet SLBR DENet LaMa MAT CoordFill SCATCL
Builds on: LaMa Restormer (transposed attention) fast Fourier convolution perceptual loss
Closest papers by meaning
- Removing Interference and Recovering Content Imaginatively for Visible Watermark Removal (AAAI 2024 · similarity 0.88)
- Image Watermarks are Removable using Controllable Regeneration from Clean Noise (ICLR 2025 · similarity 0.85)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.85)
- END^2: Robust Dual-Decoder Watermarking Framework Against Non-Differentiable Distortions (AAAI 2025 · similarity 0.85)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.85)
- WMarkGPT: Watermarked Image Understanding via Multimodal Large Language Models (ICML 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.