Research map / Image generation / Image restoration
RoPaSS: Robust Watermarking for Partial Screen-Shooting Scenarios
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 RoPaSS, a two-stage encoder-decoder watermarking scheme that uses flipping-induced symmetry plus homography estimators to resynchronize and extract watermarks from partially screen-shot images.
Problem
Screen-shooting resilient watermarks fail when only a portion of the watermarked image is captured, because existing synchronization relies on image contours or edges that are missing in partial captures.
Approach
Flipping operations build symmetric watermarks that act as an internal synchronization reference; a serial chain of watermark, symmetry, and homography estimators then predicts corner offsets to restore partially captured images before decoding.
Key results
- Under partial screen-shooting with 50% of image captured, RoPaSS reaches 93.58% extraction accuracy, versus about 52% for Hidden Code.
- Full-capture frontal extraction accuracy is 98.33% at 20 cm, with PSNR 35.902 dB, the highest among compared schemes.
- Ablation: full two-estimator synchronization reaches 97.64% ACC versus 57.13% without Ests and Estw.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RoPaSS instead of Hidden Code | Partial screen-shooting watermark extraction · Partial screen-shooting 50% | ACC: 52.42 -> 93.58% | Helped |
| use RoPaSS instead of Hidden Code | Partial screen-shooting watermark extraction · Partial screen-shooting 30% | ACC: 51.40% -> 85.58% | Helped |
| remove symmetry estimator Ests and watermark estimator Estw (ablation) | RoPaSS synchronization framework · Partial screen-shooting | ACC: 97.64% -> 57.13% | Hurt |
| remove symmetry estimator Ests (ablation) | RoPaSS synchronization framework · Partial screen-shooting | ACC: 97.64% -> 51.75% | Hurt |
| remove watermark estimator Estw (ablation) | RoPaSS synchronization framework · Partial screen-shooting | ACC: 97.64% -> 72.84% | Hurt |
| train the encoder and decoder end-to-end without synchronization (ablation) | Proposed Enc & Dec · Partial screen-shooting | ACC: 59.94% | Hurt |
| use RoPaSS instead of SSRW | Watermarked images · Visual quality | PSNR (dB): 35.674 -> 35.902 | Helped |
| use RoPaSS instead of SSRW | Watermarked images · Visual quality | SSIM: 0.9816 -> 0.9744 | Hurt |
Surprising findings
- End-to-end training without synchronization yields below 60% accuracy even with the partial screen-shooting noise layer.
Limitations
- Performance degrades at long shooting distances (60 cm) due to pixel fusion.
- Accuracy drops below 90% at 30% screen-shooting percentage and for larger shooting angles under partial capture.
- Watermark is visibly dominated by symmetric patterns, and content adaptivity is limited.
Experimental setup
Datasets: COCO USC-SIPI
Benchmarks: USC-SIPI
Metrics: extraction accuracy PSNR SSIM LPIPS
Baselines: SSRW StegaStamp PIMoG Hidden Code
Builds on: PIMoG StegaStamp HomoNet HRNet U-Net
Closest papers by meaning
- ScreenMark: Watermarking Arbitrary Visual Content on Screen (AAAI 2025 · similarity 0.86)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.85)
- Watermarking One for All: A Robust Watermarking Scheme Against Partial Image Theft (CVPR 2025 · similarity 0.85)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.84)
- Hidden in the Noise: Two-Stage Robust Watermarking for Images (ICLR 2025 · similarity 0.83)
- ROBIN: Robust and Invisible Watermarks for Diffusion Models with Adversarial Optimization (NeurIPS 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.