Spread spectrum watermarking for real images: Is everything so hopeless?

In this paper we perform the capacity analysis of known-host-statitistics watermarking methods based on spread spectrum (SS) under Additive Wight Gaussian noise (AWGN) attack. The reason of our research is based on the paradox that being non-effective in theory SS-based practical watermarking systems outperform known-host-state methods when a commonly accepted benchmarking strategy (Stirmark benchmark) is used. We show that the gap in capacity of SS-based techniques with respect to quantization-based techniques at high WNR regime could be significantly reduced, if the embedding scenario is designed using a proper stochastic model of the host image at the encoder. We show how the practical solution of watermark power allocation problem relates to the optimal one. In particular, we model the host image as an Autoregressive Process of the first order (AR(1)) and show the results of power allocation based on the water-pouring principle. It is pointed out that water-mark spectrum in a real system that is properly shaped using Contrast Sensitivity Function (CSF) well-approximates the water-pouring solution. Finally, we perform several tests to analyze the modified SS capacity facing AWGN attack. Experimental results show that host interference in this case is significantly weakened, which leads to a noticeable capacity improvement.