A novel Diamond-Mean predictor for reversible watermarking of images

Reversible watermarking (RW) is the art of embedding secret information in the host image such that after extraction of hidden information, original image is also restored from the watermarked image. Prediction error expansion (PEE) is state of the art technique for RW. Performance of PEE methods depends on the predictor's ability to accurately estimate image pixels. In this paper, a novel Diamond-Mean (D-Mean) prediction mechanism is presented. The D-Mean predictor uses only D-4 neighbors of a pixel, i.e. pixels located at {east, west, north, south}. In the estimation process, apart from edge presence, its orientation and sensitivity is also taken into account. In experimental evaluations, the D-Mean predictor outperforms currently in use MED (median edge detector) and GAP (gradient adjusted predictor) predictors. For, standard test images of Lena, Airplane, Barbara and Baboon, an average improvement of 51.79 for mean squared PE and an average improvement of 0.4 for error-entropy than MED/GAP are observed. Payload vs imperceptibility comparison of the method shows promising results.