A fast and efficient image watermarking scheme based on Deep Neural Network

Abstract

In this work, a robust image watermarking method is proposed based on the LWT (lifting wavelet transform) and DNN (Deep Neural Network). Watermark embedding uses wavelet transforms that help in maintaining a high value of imperceptibility and robustness. Different frequency bands are tested to find the optimum balance between robustness and imperceptibility. To check the robustness of proposed work, various attacks like compression attacks, noise attacks and filtering attacks are used. Deep Neural Network is trained to identify the changes made by these attacks on the different frequency bands. The use of high-frequency sub-band (LH (low-high)/LH1/HL (high-low) 2) for watermark insertion enhances the invisibility of scheme. On the contrary, various sub-bands perform differently (in terms o/f robustness) against the different types of attacks. The proposed scheme is tested on 600 images and the average PSNR (peak signal to noise ratio) is found to be 44.1148 dB with the variance of 0.5146 and standard deviation of 0.7173. Also, time analysis shows it to suitable for real-time applications. Comparative analysis suggests that the proposed method depicts improved performance over state-of-the-art techniques in almost all the parameters for most of the cases. (C) 2021 Elsevier B.V. All rights reserved.