Image-Moment Based Affine Invariant Watermarking Scheme Utilizing Neural Networks

A new image watermarking scheme is proposed to resist rotation, scaling and translation (RST) attacks. Six combined low order image moments are utilized to represent image information on rotation, scaling and translation. Affine transform parameters are registered by feedforward neural networks. Watermark is adaptively embedded in discrete wavelet transform (DWT) domain while watermark extraction is carried out without original image after attacked watermarked image has been synchronized by making inverse transform through parameters learned by neural networks. Experimental results show that the proposed scheme can effectively register affine transform parameters, embed watermark more robustly and resist geometric attacks as well as JPEG2000 compression.