Robust Multi-Bit Watermarking for Free-View Television Using Light Field Rendering

Huawei Tian, Yao Zhao, Zheng Wang, Rongrong Ni, Lunming QinIEICE Trans. Information and Systems 2013 · DOI

Introduction

Free-View Television (FTV) is an emerging immersive technology that allows viewers to navigate a scene by generating arbitrary "virtual views" from a set of fixed camera positions. This is typically achieved through Light Field Rendering (LFR), which uses interpolation to synthesize new viewpoints. From a copyright perspective, FTV presents a unique challenge: a watermark must not only survive in the original camera feeds but also persist through the mathematical blending used to create these unsampled viewpoints. This paper addresses the need for robust multi-bit watermarking—storing metadata like signatures—that remains detectable despite the geometric and color transformations inherent in the rendering process.

Methodology

Watermark Construction and Embedding

The proposed scheme treats watermarking as a communication task using Direct-Sequence Code Division Multiple Access (DS-CDMA). A multi-bit message is encoded by modulating each bit with a pseudo-random sequence. These sequences are summed to create a 1-D watermark signature, which is then converted into a 2-D format via a zigzag scan.

The watermark embedding process using DS-CDMA and DCT domain modulation across the light field image array.

As shown in the embedding workflow, the watermark is embedded into the mid-range Discrete Cosine Transform (DCT) coefficients of every original view frame. This specific targeting balances the need for imperceptibility with the robustness required to survive signal processing.

Rendering Persistence

The core innovation of this research is the mathematical proof that the watermark survives LFR. Since LFR relies on linear interpolation (such as Bilinear or Nearest Neighbor) to calculate pixel values for virtual views, the authors demonstrate that the correlation properties of the watermark are preserved in the synthesized image.

Mathematical representation of Bilinear and Nearest Neighbor interpolation used in Light Field Rendering to synthesize virtual views.

By embedding the same watermark consistently across corresponding pixels in all original views, the system ensures that any interpolated pixel in a virtual view still contains a coherent, detectable watermark signal. To retrieve the message, a correlation detector compares the rendered view against regenerated pseudo-random sequences.

Experimental Results

The researchers evaluated the method using the Buddha and Dragon benchmarks from the Stanford Light Field Archive. Key findings include:

  • Imperceptibility: Watermarked images maintained a Peak Signal-to-Noise Ratio (PSNR) of approximately 35dB, ensuring changes remain invisible to the human eye.
  • Robustness to Rendering: The Bit Error Rate (BER) remained near zero across various virtual camera trajectories and rotations.
  • Attack Resistance: The scheme proved resilient against common image processing attacks, including Gaussian/Median filtering, Salt & Pepper noise, and JPEG compression (down to a 30% quality ratio). It also survived significant cropping, remaining detectable even when 50% of the image was removed.

Conclusion

This paper successfully demonstrates a multi-bit watermarking scheme for FTV that is robust to the interpolation processes of Light Field Rendering. By leveraging the linearity of LFR and the spread-spectrum properties of DS-CDMA, the method provides a reliable way to protect copyright in immersive media environments. Future work will focus on "blind" extraction where virtual camera parameters are unknown and optimizing the system for real-time high-resolution broadcasting.

Summary updated

BibTeX
@article{tian2013robust,
  author = {Huawei Tian and Yao Zhao and Zheng Wang and Rongrong Ni and Lunming Qin},
  title = {{Robust Multi-Bit Watermarking for Free-View Television Using Light Field Rendering}},
  journal = {IEICE Trans. Information and Systems},
  year = {2013},
  doi = {10.1587/transinf.e96.d.2820},
}
Figure