Research map / Image generation / Image restoration
Physical Marker: Revealing Invisible Hyperlinks Hidden in Printed Trademarks
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 a deep hiding scheme that embeds hyperlinks only inside the opaque area of vector brand logos (via a transparency mask), with background-variation distortion and a feature prompt subspace modulator, for print-camera resilient link extraction.
Problem
Hiding and robustly extracting hyperlinks from camera-captured printed brand logos that are vector graphics with transparent backgrounds, where prior PCR watermarking methods fail due to missing backgrounds, diverse real backgrounds, and homogeneous logo regions.
Approach
Restrict encoding to the logo region using a mask derived from the logo's transparency, train with a background-variation distortion layer simulating diverse backgrounds, and add a feature prompt subspace modulator (RIM + FSM) that projects features onto learned low-rank subspaces to improve invisibility and help the decoder localize the embedded region.
Key results
- Proposed method reaches 99.83% extraction accuracy on the digital test set versus 92.88% for StegaStamp, with PSNR 31.59 dB and SSIM 0.907.
- In real print-camera tests on two phones, the method keeps above 90% accuracy across 15–40 cm distances and ±15°–±45° angles (e.g., 97.60% at 25 cm on vivo X100 Pro).
- Ablation shows FPSM raises PSNR from 27.18 to 31.59 dB when added to the mask and noise-layer model and improves accuracy to 98.95%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Physical Marker instead of the strongest listed baseline, StegaStamp | Printed-logo link extraction on METU Trademark Dataset test · METU Trademark Dataset test | ACC (%): 92.88 -> 99.83 | Helped |
| Use Physical Marker instead of LIM | Encoded-logo visual quality on METU Trademark Dataset test · METU Trademark Dataset test | PSNR (dB): 35.36 -> 31.59 | Hurt |
| Use Physical Marker instead of LIM | Encoded-logo visual quality on METU Trademark Dataset test · METU Trademark Dataset test | SSIM: 0.953 -> 0.907 | Hurt |
| Add a mask to the encoder (ablation) | Physical Marker · Ablation evaluation | ACC (%): 84.60 -> 94.82 | Helped |
| Add a background noise layer (ablation) | Physical Marker with mask · Ablation evaluation | ACC (%): 94.82 -> 98.07 | Helped |
| Add a background noise layer (ablation) | Physical Marker with mask · Ablation evaluation | PSNR: 30.49 -> 27.18 | Hurt |
| Add FPSM (ablation) | Physical Marker with mask and background noise layer · Ablation evaluation | ACC (%): 98.07 -> 98.95 | Helped |
| Add FPSM (ablation) | Physical Marker with mask and background noise layer · Ablation evaluation | PSNR: 27.18 -> 31.59 | Helped |
| Add FPSM (ablation) | Physical Marker with mask and background noise layer · Ablation evaluation | SSIM: 0.809 -> 0.907 | Helped |
| Use Physical Marker instead of Fang | Real print-camera link extraction at 15cm · vivo X100 Pro | ACC (%): 65.08 -> 90.19 | Helped |
| Use Physical Marker instead of Fang | Real print-camera link extraction at ±45° · vivo X100 Pro | ACC (%): 55.14 -> 90.39 | Helped |
| Use Physical Marker instead of Fang | Real print-camera link extraction on textured background · vivo X100 Pro real background | ACC (%): 51.43 -> 90.24 | Helped |
Surprising findings
- Adding the background noise layer alone reduced PSNR from 30.54 to 30.49 yet raised accuracy to 94.82%, while adding FPSM without the noise layer lowered PSNR to 27.18 dB; the authors also note the noise layer costs visual quality.
Limitations
- Relies on manually located and cropped regions in real-world tests.
- Lower visual quality (PSNR/SSIM) than Fang and LIM in the reported comparison.
- Real-world evaluation uses only 20 images and two phone models.
Experimental setup
Datasets: Large Logo Dataset METU Trademark Dataset
Benchmarks: METU Trademark Dataset
Metrics: PSNR SSIM ACC (bit extraction accuracy)
Baselines: Fang StegaStamp WRAP LIM
Builds on: StegaStamp HiDDeN LIM WRAP U-Net
Closest papers by meaning
- Embedding Robust Watermarking into Pattern to Protect the Copyright of Ceramic Artifacts (AAAI 2025 · similarity 0.82)
- Silent Branding Attack: Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.81)
- IWRN:A Robust Blind Watermarking Method for Artwork Image Copyright Protection Against Noise Attack (AAAI 2025 · similarity 0.79)
- Robust Message Embedding via Attention Flow-Based Steganography (CVPR 2025 · similarity 0.79)
- EditGuard: Versatile Image Watermarking for Tamper Localization and Copyright Protection (CVPR 2024 · similarity 0.79)
- Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models (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.