Research On GIF Dynamic Image Robust Watermarking Based On Deep Learning

In the era of new media,Graphics Interchange Format(GIF)images have been widely applied and spread in various social networks in a special image format,such as dynamic emoticons,dynamic advertisements,etc.,and their continuous dynamic sense has strong interactivity and attraction.At the same time,it can intuitively create fantastic visual effects and atmosphere.In order to achieve the balance between GIF image transmission information and copyright protection,more attention should be paid to the copyright protection of this multimedia form.Digital watermarking technology is one of the research hotspots in the field of information security because of its important application value in the link of copyright confirmation and copyright monitoring.Traditional research on GIF dynamic image watermarking usually focuses on fragile watermarking,while existing robust watermarking methods ignore the balance between imperceptibility and robustness.Therefore,this paper aims to improve the robustness of GIF dynamic image watermarking,combines the good performance of deep learning network in static image watermarking scheme,and develops the design according to the characteristics of GIF dynamic image carrier.The main contributions of the paper are as follows:(1)Aiming at the problem that it is difficult to balance the invisibility and robustness of GIF watermarking,a watermarking method based on attention mechanism and frequency domain enhancement is proposed.In the first stage,for the specific time dimension of GIF image,3DCNN convolution is used to capture the interframe motion information of the time dimension,and an improved attention module of channel dimension is introduced into the encoder to ensure that the watermark is embedded in the insensitive position of each pixel between frames.In the second stage,the frequency domain enhancement module is introduced.For GIF dynamic images obtained in the first stage,the frequency domain enhanced carrier with watermark residual signal is added to the carrier frame by frame.In the third stage,the decoder is retrained to extract high-quality watermark information.Experiments show that compared with existing methods,the proposed method can improve the robustness of watermarking network on the premise of ensuring the invisibility of watermarking.(2)Aiming at the problem that the current GIF watermarking is difficult to resist unknown distortion,a new GIF watermarking method based on noise self-learning network is proposed.After embedding the watermark,a noise self-learning network is introduced to construct different noise forms through the automatic feature learning ability of the deep neural network,which can improve the robustness of the model without using simulated noise or real noise,and avoid the high cost of sacrifice.It also enables the network to learn unknown distortion and have resistance to it,while maintaining a high level of invisibility due to training with the encoder and decoder at the same time.The experimental results show that compared with the existing methods,the proposed method can resist unknown distortion such as color change and noise combination,and also has a good performance in simulating noise.