Reversible Watermarking Algorithm Based on Embedding Pixel Dependence

Many works have demonstrated that the performance of the prediction error (PE) based reversible watermarking algorithm depends on the precision of the prediction value. However, most algorithms compute the prediction value by exploiting only the correlation between the embedding pixel and its neighboring pixels, thus limiting the watermarking performance of these algorithms. In this work, we obtain twelve prediction candidates from neighboring pixels of the embedding pixel based on the modeling assumption -it is easy to find out that an embedding pixel value is identical or similar to one of its neighboring pixel values in an image. The final prediction value can then be selected from these prediction candidates by using the prediction value and the original pixel value. A more accurate final prediction value can then be obtained. Certainly, the original pixel value is also recovered exactly during the decoding process, in which the final prediction value is obtained according to the twelve prediction candidates, the prediction value and the watermarked pixel value. Experimental results indicate that the variance of the prediction error histogram obtained by the proposed method is about 44.2% less than that proposed by Sachnev et al. Moreover, the mean peak signal-tonoise ratios (PSNRs) are about 1.47 dB and 1.1 dB greater than those of proposed by Sachnev as the watermark capacities are 0-0.04 bits per pixel (bpp) and 0.04-0.5 bpp, respectively. The proposed algorithm performs better than existing reversible watermarking approaches.