Reversible watermarking via extreme learning machine prediction

In this paper, we attempt to construct a novel framework of reversible watermarking. This work is based on the difference-image histogram shift. De-correlation is the core of high capacity data-hiding in histogram-shift techniques. For the sake of higher payload, we choose the down-sample pattern as reference set. For each layer, prediction points are obtained in terms of points from the reference set. The full-resolution image quality reconstructed determines to reversible watermarking performance. When existing the prior knowledge, an effective regression method named extreme learning machine is utilized to estimate missing pixels. It can yield high-quality recovery image. Compared to other better algorithms on state of the art, the proposed method achieves higher capacity gain of watermarked images with the similar distortion.