Visual attention based image zero watermark scheme with ensemble similarity

Most existing digital image watermarking schemes tend to have the inherent conflicts between imperceptibility and robustness because watermarks are embedded by parameter modification. Image watermark techniques resolve this dilemma by extracting invariant features from the image as “embedded” watermark. In this paper we propose an image zero watermark scheme based on visual attention regions of images. In the proposed scheme visual attention model carefully selects top-N salient areas, where a set of selected Scale Invariant Feature transform (SIFT) descriptors are extracted as a watermark. The distance of each pair of SIFT descriptors from the reference and test images are calculated by Kullback-Leibler (KL) divergence after mapping into the high dimensional space respectively. The final distance of two sets of SIFT descriptor are determined by ensemble similarity. The experimental results indicate that the proposed scheme outperforms image zero watermark scheme based on color and edge histograms (CEH) and is robust to attacks of geometric distortion, contrast/luminance distortion and JPEG compression.