Mapping based Residual Convolution Neural Network for Non-embedding and Blind Image Watermarking

Abstract

Traditional image watermarking algorithms directly modify the host image by watermark embedding, which is hard to balance the contradiction between the robustness and imperceptibility. Inspired by the human brain’s associative memory, this paper proposes a non-embedding and blind image watermarking algorithm via mapping based Residual Convolution Neural Network (Mapping-based RCNN). For preprocessing, median filter is applied on the host image to enhance the robustness of the algorithm to against various attacks. After that, Discrete Cosine Transform and Singular Value Decomposition are adopted to extract the corresponding image information matrix. To obtain the mapping relationship between host image and watermark image, the information matrix is input into the designed Mapping-based RCNN structure for network training. The Mapping-based RCNN is a non-embedding watermarking algorithm, which not only overcomes the imperceptibility shortcoming but also wins good robustness compared with traditional watermarking algorithms. Experimental results show that the proposed algorithm can successfully extract the watermark images under various attacks, and is more robust than existing watermarking algorithms.