A geometrical and frequential watermarking scheme using similarities

Watermarking schemes are more and more robust to classical degradations. The NEC system developed by Cox,(1) using both original and marked images, can detect the mark with a JPEG compression ratio of 30. Nevertheless a very simple geometric attack done by the program Stirmark can remove the watermark. Most of the present wartermarking schemes only map a mark on the image without geometric reference and therefore are not robust to geometric transformation. We present a scheme based on the modification of a collage map (issued from a fractal code used in fractal compression). We add a mark introducing similarities in the image. The embedding of the mark is done by selection of points of interest supporting blocks on which similarities are hided. This selection is done by the Stephens-Harris detector. The similarity is embedded locally to be robust to cropping. Contrary to many schemes, the reference mark used for the detection comes from the marked image and thus undergoes geometrical distortions. The detection of the mark is done by searching interest blocks and their similarities. It does not use the original image and the robustness is guaranteed by a key. Our first results show that the similarities-based watermarking is quite robust to geometric transformation such as translations, rotations and cropping.