The color image steganography in frequency domain based on separation training and image denoising

Color image steganography attracts the attention of scholars because of its secretive and imperceptibility. The color image steganography based on frequency domain has achieved better performance in both traditional steganography and deep learning steganography. However, most current steganographic models based on auto-encoder have limitations in improving the ability of reconstructing secret images. Based on this problem and the existing advantages of steganography in the frequency domain, a color image steganographic method based on separation training and image denoising is proposed. In the face of the performance trade-off between the encoder and the decoder, the proposed method uses the separation training to optimize the model training. In addition, the proposed model adds an image-denoising module which is a Denoising Convolutional Neural Network(DnCNN). Experimental results show that the Peak Signal to Noise Ratio(PSNR) of the stego image and the reconstructed secret image reached 82.31 dB and 39.27 dB respectively, and the Structural Similarity Index Measure(SSIM) reached 0.99. Compared with other models, the proposed model not only has stronger imperceptibility but also has better ability to reconstruct secret images.