A robust document image watermarking scheme using deep neural network

Watermarking is an important copyright protection technology which generally embeds the identity information into carrier imperceptibly. Then can be extracted to prove from watermarked even after suffering various attacks. Most of existing watermarking technologies take nature images as carriers. Different natural images, document are not so rich in color and texture, thus have less redundant carry watermarks. This paper proposes end-to-end image scheme using deep neural network. Specifically, encoder a decoder designed embed extract watermark. A noise layer added simulate attacks that could encountered reality, such Cropout, Dropout, Gaussian blur, noise, Resize, JPEG Compression. text-sensitive loss function limit embedding modification on characters. An strength adjustment strategy proposed improve quality with little extraction accuracy. Experimental results show outperforms three state-of-the-art methods terms robustness quality.