Model Order Selection in Reversible Image Watermarking

Digital watermarking is becoming increasingly important in a large number of applications such as copyright protection, content authentication, and document annotation. Thanks to its capability of exactly recovering the original host, reversible image watermarking, a kind of digital watermarking, is favored in fields sensitive to image quality like military and medical imaging. This paper presents theoretical examination and experimental analysis of model order selection in reversible image watermarking. It involves two modeling tools: prediction and context modeling. Classic prediction models are compared and evaluated using specially derived criteria for reversible image watermarking. Among them, the CALIC, a tool combining the Gradient-Adjusted Prediction with a context modeling, stands out as the best by providing the most competitive model-fitness with relatively low complexity. In addition, full context prediction, a model unique to reversible image watermarking, is also discussed. By exploiting redundancy to greater extent, it achieves highly fitted modeling at a very low order. Experimental results demonstrate that it is capable of providing even better performance than the CALIC with only negligible computation.