摘要
针对目前隐写分析算法对图像复杂纹理区域特征表征能力较弱的问题,提出一种基于多尺度特征融合和注意力机制的隐写分析模型。该模型首先使用空域富模型滤波器对输入图像进行预处理,提取噪声成分残差,降低图像本身内容的影响;其次使用多尺度并行网络提取信号,增强对细微特征的学习;然后引入注意力机制对特征进行自适应加权,强调重要通道特征在分类中的作用,同时抑制非重要通道特征对分类的影响;最后提出一种协方差池化对深度神经网络学习后的各特征之间的相关性进行建模,并选取牛顿迭代法求解平方根矩阵,使网络训练更加高效。实验结果表明:在小波权重隐写算法0.5 bit/像素嵌入率的条件下,所提模型准确率达到了88.6%,证明了所提方法的有效性。
Regarding the weak capability of current steganalysis algorithms to characterize features of complex texture regions in images,a steganography analysis model based on multi-scale feature fusion and attention mechanism was proposed.Firstly,the SRM(spatial rich model)filters are used to preprocess the input image to extract the noise component residuals so as to reduce the influence of the content of the image itself.Then,the multi-scale parallel net-work is used to extract the signal to enhance the learning of subtle features.Subsequently,the attention mechanism is introduced to carry out adaptive weighting of the features,emphasizing the role of important channel features in clas-sification,while simultaneously reducing the influence of non-important channel features on the classification.Final-ly,a covariance pooling method is proposed to model the correlation between features after deep neural network learning,and Newton iteration method is selected to solve the square root matrix to make network training more effi-cient.Experimental results demonstrate that the accuracy of the proposed model reaches 88.6%under the condition of a 0.5 bit per pixel embedding rate of the WOW(wavelet obtained weights)algorithm,which proves the effective-ness of the proposed method.
作者
李萌
罗维薇
刘长龙
LI Meng;LUO Weiwei;LIU Changong(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《兰州交通大学学报》
CAS
2024年第3期57-67,共11页
Journal of Lanzhou Jiaotong University
基金
国家自然科学基金(62362047)。
关键词
隐写分析
卷积神经网络
空域富模型
注意力机制
steganalysis
convolutional neural network
spatial rich model(SRM)
attention mechanism
