An End-to-End Watermarking Framework Integrates Matrix Decomposition and Deep Networks

To protect image content security and digital copyright, an end-to-end watermarking framework integrates matrix decomposition and deep networks. The framework includes an encoder, an adversarial network, a noise layer, and a decoder. The feature of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R_{1,1}$</tex> in the matrix <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R$</tex> of the image after QR decomposition (QRD) is extracted by the encoder through SE block. 256 watermark bits are embedded into <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R_{1,1}$</tex> using convolutional network, and the output watermarked image is transmitted to the adversarial network. The combination of the two is trained to enhance the encoding ability of the encoder. The decoder is used to extract watermark information. The experimental data shows that when the embedding strength is 1.5, the PSNR value is larger than <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$50 d B$</tex>, the SSIM is larger than 0.95, and the BER value is 0.47 %. Under the same PSNR, the proposed scheme shows better robustness than other schemes.