.png){kind=icon}
.png){kind=icon}
Return
A robust image watermarking framework based on U2-net encoder and loss function weight assignment
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-024-01640-1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the field of watermarking, the process of guiding the watermark embedding is crucial, which can determine the imperceptibility and robustness of the watermarking algorithms. Traditional algorithms for guided watermark embedding are susceptible to feature extraction capabilities and their robustness cannot meet the demands of practical applications. Recently, some watermarking frameworks based on deep learning have been proposed, which employ an encoder to automatically extract features and guide the watermark embedding, and show better robustness in practical applications. However, some encoders also suffer from insufficient feature extraction capability due to their network structure, resulting in unsatisfactory visual quality. To this end, an encoder based on improved U-2-Net structure is proposed, which performs multi-scale and multi-dimensional feature extraction on images to generate high-resolution feature maps with edge information. These feature maps help guide the watermark to be embedded in the optimal pixel space, thus improving the visual quality. In addition, a weight assignment strategy that combines square growth and linear growth is proposed to balance visual quality and robustness in training. The experimental results demonstrate that our method outperforms existing end-to-end watermarking schemes in terms of imperceptibility while ensuring high robustness.
Keywords:
Watermarking
U-2-Net
Weight assignment
Loss function