Research On The Watermarking Of The Animated GIF Image Based Deep Learning

With the highly conformable expression of the animated GIF images,animated interest has been widely concerned on the social platform and aroused the nationwide GIF images creation.With the popularity of the animated GIF images,it has been widely applied on various social media platforms,and gradually started commercial application in advertising commercial fields.The commercial application of the animated GIF image puts forward higher requirements for its copyright protection.Digital watermarking technology has been playing an important role in information security,and robust watermarking is an effective means of copyright protection.Existing watermarking research of the GIF image mainly focuses on fragile watermarking,and there is few robust watermarking method of the animated GIF images.Different from other image formats,the animated GIF image has frame information which need unique design.With the rapid development of deep learning,convolutional neural network has become an effective tool in computer vision.Therefore,this paper propose a robust watermarking method of animated GIF images based deep learning,and the main research results are as follows:(1)A robust watermarking framework based 3D convolutional neural network for the animated GIF images.End-to-end watermark image embedding and extraction are realized.According to the time dimension of GIF image,3D convolutional neural network is used to extract spatial-temporal feature.The Pre-network constructs the time sequence feature of the watermark image,and automatically learns to adapt to the embedded static watermark image features,which can achieve large capacity embedding.By simulating the common image processing operations in the Noise layer,the robustness is obtained.At the same time,the frame deletion and frame replacement noise types are designed to resist the corresponding noise interference.(2)On the basis of the proposed watermarking framework,an adversarial network is added.The adversarial network and the Encoder are trained to promote the training of the Encoder,which can optimize the generated images containing watermarks.Through the phased method,the first stage is mainly to train the Encoder,and the parameters are fixed,and then the Decoder is trained in the second stage.On this basis,because the continuous differentiability requirement can be avoided in the simulation of noise type,a large number of frames containing watermark images can be directly deleted in the second stage.By training the Encoder and Decoder in stages,the influence caused by the training of the Decoder is isolated,and the current goal of the network is focused.It can improve image performance,generate high quality watermarked image,and simulate high intensity frame deletion.