A novel robust watermarking scheme based on neural network

A color image oblivious watermarking scheme based on neural network and discrete wavelet transform (DWT) is proposed in this paper. Three identical watermarks and some different expanded bit streams are adaptively embedded into the low frequency sub-bands generated from three channels for a color image, respectively. Due to the adaptive learning capabilities of neural network, the expanded bit streams could be used to train back propagation (BP) neural networks to represent the relationship among the neighbor wavelet coefficients. Based on the trained neural networks, three watermarking results can be extracted and then are voted to decide the final watermark. Extensive experiments illustrate that the new scheme possesses good robustness against different attacks including noise addition, shearing, scaling, luminance and distortion. And what's more, the scheme has excellent performance in term of imperceptibility and resistance to JPEG compression.