融合Gabor滤波与Transformer的图像水印方法  

Image Watermarking Approach Fusing Gabor Filtering with Transformer

作　　者：张天骐[1] 谭霜 沈夕文 唐娟 ZHANG Tianqi;TAN Shuang;SHEN Xiwen;TANG Juan(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《信号处理》2025年第4期694-705,共12页Journal of Signal Processing

基　　金：重庆市自然科学基金项目(cstc2021jcyj-msxmX0836)。

摘　　要：图像水印在数字版权保护和身份验证领域中具有关键意义,是保护图像信息安全和确保数据可信性的重要技术手段。目前,大多数已发表的基于深度学习的图像水印方法都是基于卷积神经网络设计的,此类方法存在无法充分捕捉图像的全局信息和细节信息,以及忽略图像高频信息具备稳定和不可感知特点等问题,为了克服上述问题,该论文提出一种融合Gabor滤波与Transformer的图像水印模型。该模型由嵌入网络、提取网络和判别网络组成:在嵌入网络设计了水印信息处理模块对水印信息引入冗余和扩展操作,以增加水印信息在传输过程中的鲁棒性;在嵌入网络引入Gabor滤波的思想在特征提取模块通过卷积分支来捕捉局部特征,通过Transformer分支捕捉全局信息,来充分挖掘图像的稳定特征;在提取网络中融合标准卷积和差分卷积,来准确感知图像的细微信息,进而提高水印的提取精度;引入判别网络与嵌入网络形成对抗训练关系,评估生成水印图像的真实性和质量,从而提升嵌入网络生成水印图像的视觉质量。分别在COCO、ImageNet和VOC2012数据集下进行综合对比实验,结果表明,该文方法针对不可感知性和鲁棒性,相比于相关水印模型取得了更优的指标,具有较为突出的增强性能与泛化能力。此外,还进行了相关的消融实验,结果进一步验证了该模型的可靠性和有效性。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 watermar

关 键 词：图像水印 不可感知 鲁棒性 卷积神经网络 TRANSFORMER 

分 类 号：TN911.7[电子电信—通信与信息系统]