Image Watermarking Approach Fusing Gabor Filtering with Transformer

‍ ‍Image watermarking is of key significance in the field of digital copyright protection and identity authentication. Moreover， it serves as a crucial technical means for safeguarding image information security and ensuring data trustworthiness. In recent years， deep learning has been widely adopted in the field of watermarking， owing to its powerful and superior learning capabilities. Currently， most of the published deep learning-based image watermarking methods are designed based on convolutional neural networks. However， among other shortcomings， these methods often fail to adequately capture global and detailed image information and ignore high-frequency information images that possess stable and imperceptible characteristics. To address these issues， this thesis proposes a fusion of Gabor filtering and a Transformer’s image watermarking model. The model consists of an embedding network， extraction network， and discriminative network. In the embedding network， the watermarking information processing module is designed to incorporate redundancy and expansion operations into the watermarking information to enhance the robustness of the watermarking information in the process of transmission. Additionally， in the embedding network， the feature extraction module captures local features and global information using convolutional and Transformer branching， respectively， to fully explore the stable features of the image. In the extraction network， ordinary convolutional branching is fused with the Transformer to effectively and exhaustively explore the stable and imperceptible characteristics of image features. The fusion of ordinary and differential convolutions in the extraction network aims to accurately perceive the subtle information of the image， thereby improving the extraction accuracy of the watermark. A discriminative network is introduced to establish an adversarial training relationship with the embedding network to evaluate the authenticity and quality of the generated watermark image， enhancing the visual quality of the watermark image produced by the embedding network. Comprehensive comparison experiments are conducted using the Microsoft Common Objects in Context， ImageNet and pattern analysis， statistical modeling， and computational learning visual object classes 2012 datasets. The results demonstrate that the method proposed in this paper achieves superior indexes for imperceptibility and robustness， compared with related watermarking models， and provides significant improvements in enhancement performance and generalizability. Relevant ablation experiments further validate the reliability and effectiveness of the model.