A Robust Blind Watermarking Scheme Based on Attention Mechanism and Neural Joint Source-Channel Coding

Abstract

Due to the importance of protecting digital images in the modern world, digital image watermarking draws the attention of academia because of its applications in proving the authenticity and integrity of digital images. The performance of current watermarking algorithms needs to be improved to cope with the growing demand for better imperceptibility, greater robustness, and larger capacity. In response to the above requirements, a new robust blind watermarking algorithm is proposed in this paper, which mainly utilizes convolutional neural networks to embed watermarks in DCT domain. An attention module is introduced to calculate the confidence that each image pixel block contains watermark information, thus making our algorithm robust enough to deal with Crop attack. In addition, the introduction of a learning based solution for joint source and channel coding enables our algorithm to take grayscale images with simple backgrounds as input watermarks while maintaining great imperceptibility and robustness. Extensive experiments demonstrate the superiority of our algorithm.