全局协方差池化与多尺度特征融合的图像隐写检测

Image Steganography Detection Based on Multi-Scale Feature Fusion and Global Covariance Pooling

针对目前图像隐写深度检测模型中池化等操作造成特征图信息丢失,全局平均池化忽视高阶统计量的问题,提出一个基于全局协方差池化与多尺度特征融合的隐写检测模型:首先用多层小尺度卷积核替换多层感知器中的大尺度卷积核,增强特征表达能力的同时提高卷积计算效率;然后利用空洞卷积构建多尺度特征融合模块,减少模型在池化等过程中导致的细节特征信息损失;最后引入全局协方差池化,通过计算二阶统计量协方差作为最后的特征输出,增强检测模型对细节特征的捕捉能力。实验结果表明在不同的隐写算法和不同的嵌入率下,相比于近期典型的Xu-Net、Yedroudj-Net、Zhu-Net模型,所提模型的检测准确率均有显著提升,即使是与最新的Zhu-Net相比,准确率也提升了2.4%~7.3%。

A steganalysis model based on multi-scale feature fusion and global covariance pooling is proposed in response to the feature map information loss caused by pooling and other operations in current image steganography deep learning models,particularly the high order statistic loss by the global average pooling.First,the multi-layer perceptual convolution's large-scale convolution kernel is changed to a small-scale convolution kernel,improving the capacity to express features while lowering the number of parameters.Second,to miti gate the detailed information loss brought on by pooling and other operations,a multi-scale feature fusion module based on the dilated convolution is employed.Finally,to improve the capacity to describe precise features,a global covariance pooling is used,which outputs the second-order statistical covariance.In comparison to established models like Xu-Net,Yedroudj-Net,and Zhu-Net,the suggested model increases detection accuracy.It even surpasses the newly proposed Zhu-Net by 2.4%to 7.30%for various steganographic tech niques and embedding rates.