WFormer: A Transformer-Based Soft Fusion Model for Robust Image Watermarking

Most deep neural network (DNN) based image watermarking models often employ the encoder-noise-decoder structure, in which watermark is simply duplicated for expansion and then directly fused with image features to produce the encoded image. However, simple duplication will generate watermark over-redundancies, and the communication between the cover image and watermark in different domains is lacking in image feature extraction and direction fusion, which degrades the watermarking performance. To solve those drawbacks, this paper proposes a Transformer-based soft fusion model for robust image watermarking, namely WFormer. Specifically, to expand watermark effectively, a watermark preprocess module (WPM) is designed with Transformers to extract valid and expanded watermark features by computing its self-attention. Then, to replace direct fusion, a soft fusion module (SFM) is deployed to integrate Transformers into image fusion with watermark by mining their long-range correlations. Precisely, self-attention is computed to extract their own latent features, and meanwhile, cross-attention is learned for bridging their gap to embed watermark effectively. In addition, a feature enhancement module (FEM) builds communication between the cover image and watermark by capturing their cross-feature dependencies, which tunes image features in accordance with watermark features for better fusion. Experimental results show that the proposed WFormer outperforms the existing state-of-the-art watermarking models in terms of invisibility, robustness, and embedding capacity. Furthermore, ablation results prove the effectiveness of the WPM, the FEM, and the SFM.