A new encod-decoder steganography scheme based on dense residual connection was proposed to solve the problem of poor quality of encod-decoder images and message images generated by image steganography schemes based on encoder-decoder net-works.Different from the existing end-to-end image steganography networks,the proposed scheme does not need to preprocess the im-age,and adopts dense residual connections to transport the features of the shallow network to each layer of the deep network structure,effectively preserving the details of the feature map,and uses channels and spatial attention modules to filter the features,improving the codec's attention to the complex texture region of the image.Experimental results on LFW,PASCAL-VOC12 and ImageNet datasets show that the proposed method can effectively improve image quality under the premise of ensuring the security of the algorithm,inclu-ding the peak signal-to-noise ratio of dense images and carrier images.The mean values of peak signal-to-noise ratio(PSNR)and struc-tural similarity(SSIM)are 36.2 dB and 0.98 respectively.
关键词
信息隐藏/深度学习/注意力机制/编码-解码结构/密集残差网络
Key words
information hiding/deep learning/attention mechanism/encoding-decoding structure/dense residual network
维普
万方数据