Deep image watermarking with loss-driven modification

Conventional watermarking algorithms are based on handcrafted features or their fusion. However, how to design and fuse these features efficiently for the performance of watermarking mainly depends on experience. In addition, the embedding is also restricted by a certain predefined function according to prior experiences. In order to improve the performance of watermarking by optimizing these two factors, a novel learning-based blind watermarking algorithm is proposed, motivated by the self-adjustment of deep learning. In the proposed method, the extraction of watermarks is modeled as block classification, which is implemented by a modified convolutional neural network (CNN). And the cross-entropy function is set as the loss function of this CNN to bridge the performance of watermarking, feature extraction, and embedding, which overcomes the problems brought about by the predefined modification criterion. Furthermore, in the modified CNN, a residual module is constructed with fewer batch normalizations, which can efficiently reduce the time of the watermarking process. Experimental results show that the proposed algorithm can achieve robustness against common signal processing attacks.