A Novel Zero-Watermarking Algorithm Based on Separated Learning of Residual-DenseNet
Hanwen Dong, Yusuke Sao, Toshiyuki Uto · 2025
This paper presents a robust zero-watermarking technique with multiple learning of residual-densely connected convolutional networks (DenseNets) for images. For a usage of feature vector as the labels, we separate a training process of residual DenseNet (RDNet) into pre-learning, labeling, and post-learning. After pre-training RDNet from the original image, the proposed watermarking approach derives a feature vector as the labels by the pre-trained RDNet. From the resulting feature vector and the corresponding attacked sample images, the RDNet is trained for image zero-watermarking. Experimental results show that our designed RDNet-based zero-watermarking has good robust performance against different attacks.