Reversible Watermarking

Background

The straightforward approach to reversible watermarking is by using lossless compression (Fridrich, Goljan, & Du, 2002, etc.). In order to make room for data embedding, a part of the host is compressed and substituted by the compressed data and the watermark. The imperceptibility of the watermarking imposes the substitution somewhere into the least significant bits of the host. Obviously, the least significant bits are noisy and, consequently, the compression ratio is rather low. The compression based approaches are either of rather low capacity, or, in order to gain in capacity, of rather high complexity. The best performances reported so far for compression based reversible watermarking are about 1 bpp (Celik, Sharma, & Tekalp, 2006).

Another approach is histogram shifting reversible watermarking. In the original approach of Ni, Zhicheng et al. (2006), image graylevels are shifted in order to create, into the image histogram, a gap of one graylevel adjacent to the most populated graylevel, i.e., to the histogram peak. Data embedding follows immediately by scanning the image and by flipping the pixels of the most populated graylevel to the one of the gap if the bit to be embedded is “1” and by not flipping for “0.” In one watermarking level, the original method provides an embedding capacity of the order of the maximum histogram bin. Sharper histograms (Laplacian distributed), as the ones of the difference between adjacent pixel or of the prediction error, are used to increase the embedding capacity (Lin et al., 2008). In a single level of embedding, the histogram shifting reversible watermarking provides up to about 0.1-0.3 bpp at a very low distortion. Bit-rates greater than 1 bpp can be obtained by multiple embedding. Obviously, the distortion increases with each level of embedding.