Research On Robust Watermarking Algorithm For Color Image Based On Autoencoder

As digital media is widely used in daily life,it is more and more urgent to protect digital media from fake and piracy.Digital watermarking technology has become an important research direction in the field of information security.Color image is the most common cover and deep learning can automatically extract image features,so robust color image watermarking based on deep learning is a research hotspot in the field.In view of the problems of existing robust color image watermarking based on deep learning,such as not good enough performance against JPEG lossy compression attacks and lack of visual perception guidance in the process of watermark embedding,this paper carries out the following two aspects of researches:(1)To solve the problem that the existing robust color image watermarking algorithm based on deep learning is not robust enough to resist JPEG lossy compression attack,this paper proposes a new color image JPEG lossy compression simulation network and based on the simulation network,a color image watermarking network based on autoencoder is designed.The standard quantization table and quality factor are used for the first time in the simulation process,which solves the problem that existing algorithms cannot simulate a specific quality factor.The proposed network is designed in accordance with JPEG lossy compression standard flow.Max-pooling is used to simulate 411 sampling,data rearrangement is used to simulate8×8 data blocks,1×1 convolution is used to simulate DCT transform,3D noise mask is used to simulate non-differentiable round operation in quantization,and loss function is used to constrain the difference between the network output and the real compressed image.In order to verify the feasibility of the proposed JPEG lossy compression simulation network in the watermarking field,a color image watermarking network based on autoencoder is designed by referring to the mainstream watermarking network,and the watermark is embedded in the B channel which is insensitive to human eyes.Then,the proposed color image JPEG lossy compression simulation network is pre-trained.In the training stage,it is added to the middle of the watermark embedding network and the extraction network so that the whole framework is an end-to-end system.Experimental results show that the proposed network can accurately simulate JPEG lossy compression with various specific quality factors and verify that the proposed JPEG lossy compression simulation network can effectively improve the performance of watermark network against JPEG lossy compression,and it performs well under various quality factors and is superior to other related methods.(2)Aiming at the problem that existing robust color image watermarking algorithms based on deep learning do not fully consider human visual perception when generating watermarked images,this paper proposes a robust color image watermarking algorithm based on visual perception.Firstly,the chromaticity and texture distribution of the cover images are discussed,and the visual perception loss function is proposed to guide the network to embed watermark into the blue-green region and richly textured region.Secondly,self-attention mechanism is introduced in the stage of watermark embedding and extraction to improve the quality of reconstructed images.Finally,some common signal processing attacks such as the above proposed JPEG lossy compression simulation network,gaussian white noise and mean filtering are added to carry out end-to-end training to improve the robustness of the proposed algorithm.Experimental results show that the proposed algorithm achieves a better balance between invisibility and robustness,and is superior to other related methods.