Robust Watermarking Of RGB Image Against Geometric Attacks And Handwritten Digit Recognition Algorithm Based On Convolutional Neural Network

In aspect of multimedia information security,with the rapid changes in computer technology,people can easily obtain the digital multimedia works through the Internet and various of social software,which brings great convenience to people's daily life and work.But in the meanwhile,the copyright protection of digital multimedia works has become a focus of attention of the community.In this paper,a robust watermarking algorithm based on RGB image is proposed for geometric attacks by analyzed properties of R,G and B channels.Firstly,the pixels in G and B channels of an image are expanded twice and three times,respectively.After that,the image is filtered by a Gaussian low-pass filter so that its mean and histogram can be used for watermark embedding.Furthermore,we embed the same number of bits into R,G and B channels for robustness testing,and it is noting that the robustness in the three channels are different.Experimental results with 100 RGB images have shown that the strategy to expand the pixels in G and B channels are beneficial to improve the robustness of histogram-based watermarking for geometric attacks.On the other hand,this paper explores handwritten digit recognition based on convolutional neural networks(CNN),introduces the basic structure of the CNN and studies the structures and parameters of LeNet-5.The main research work is as follows:(1)This paper exploits the MNIST and USPS of hand-written digits datasets to research and compare the different structures of LeNet-5 network.By changing the number of convolutional layers and pooling layers in LeNet-5,the accuracy rate of handwritten digit recognition in different datasets is observed.(2)We research the influence of the number of the images processed in batch size,gradient descent steps and epochs under the different network structures of LeNet-5 and at the same time,we obtained the good experimental results on the MNIST and USPS datasets.This paper is mainly engaged in the research of multimedia digital watermarking against geometric attacks and obtains some significant results.At the same time,we make exploration on the recognition of handwritten digits based on convolutional neural networks and analyze the influence of different network structures and datasets on the accuracy of handwritten digit recognition.