Abstract
In digital rights management, watermarks are often embedded in low-resolution images that are subsequently upscaled for distribution. When these upscaled images are corrupted by noise or compression, standard detection methods—which typically involve downscaling—often fail to recover the signal effectively. This paper investigates watermark detection in noisy interpolated images using polyphase decomposition and optimal fusion. By treating interpolation as a source of redundancy rather than a distortion, the proposed method derives an optimal fused correlator that significantly enhances detection robustness.
Introduction
Digital watermarking serves as a critical tool for copyright protection. However, the transmission channel often introduces distortions such as interpolation (resizing) and additive noise. Conventional detection strategies usually attempt to reverse the scaling operation, effectively discarding the additional pixels generated during interpolation. This research argues that these extra pixels contain valuable redundant information about the original watermark signal. By leveraging polyphase decomposition, we can extract multiple estimates of the original image and fuse them optimally to maximize the detection signal-to-noise ratio (SNR).
Proposed Methodology: Polyphase Decomposition and Optimal Fusion
The core of the proposed system is a two-stage process involving embedding and a sophisticated detection architecture.
As shown in Figure 1, the detection process begins by decomposing the high-resolution noisy image $s$ into its polyphase components. For an interpolation factor $M$, the signal is split into $M$ branches. The first branch $s_0$ represents a downsampled version of the received signal. The subsequent branches $s_1, \dots, s_{M-1}$ are filtered using FIR filters $h_m[m]$ to produce additional estimates of the original watermarked image.
To combine these estimates, the authors employ Optimal Fusion. Instead of simple averaging, they use Lagrange optimization to determine weights $a_0, \dots, a_{M-1}$ that minimize the noise variance in the final fused signal $y_f$. This statistical approach ensures that the detector prioritizes the most reliable information across all polyphase components.
Experimental Results
The performance of the fused detector was evaluated using $10^6$ Monte Carlo simulations on the standard Lenna image, comparing it against standard low-resolution and high-resolution detection baselines.
As illustrated in Figure 2, the fused detection scheme (dash-dot line) consistently outperforms both the low-resolution (solid line) and high-resolution (dotted line) approaches. In environments with i.i.d. Gaussian noise, the fused correlator achieves a significantly lower probability of false rejection ($P_{fr}$) for any given probability of false alarm ($P_{fa}$).
Key findings include:
- Noise Robustness: The method showed a massive improvement in Equal Error Rate (EER), outperforming low-res detection by over 15% and high-res detection by over 36% in high-noise scenarios.
- Compression Resilience: Despite being derived for Gaussian noise, the system proved robust against JPEG compression. At a quality factor of 40, the fused detector maintained high reliability where standard high-res methods degraded.
- Versatility: The framework also successfully handled non-integer scaling and multi-stage interpolation (e.g., bicubic followed by bilinear scaling).
Conclusion
This study demonstrates that interpolation in the presence of noise should be viewed as a resource for diversity rather than a mere distortion. By performing detection in an optimally fused domain rather than the original image space, watermark reliability is significantly enhanced. The proposed polyphase architecture provides an elegant and statistically sound methodology for robust watermark recovery in practical, noisy environments.
Summary updated