Research On Imagr Content Identification Method Based On Deep Learning

With the popularization of digital technology,digital image processing technology has developed rapidly,and it has become very easy to change image content.Even non-professionals can easily edit image content using image editing tools.In the last decades,with the continuous emergence of image processing tools,image splicing technology has become more and more complex.Various forged images maybe used in news media,scientific research,forensic medicine,medical diagnosis,cultural media,military applications and other fields.The existence and spread of fake images reduce the credibility of image information,and have caused seriously negative impact in many fields.It has changed the traditional cognition of "seeing is believing".In this thesis,we study the method of image splicing detection and deepfake detection.The main works are as follows:We proposed an image splicing detection method based on Deep Convolutional Neural Network that can effectively distinguish fake images from real images.The proposed DCNN model consists of three parts:data processing,feature extraction and detection.In the data processing part,we expanded Columbia and CASIA databases through scale transformation,mirroring,rotation and other operations to obtain a large capacity tampered image database,which can enable the model to extract the discrimination features between natural images and forged images.In the feature extraction part,we use four convolution layers and four pooling layers to extract the features of the input image.The weight and offset of each convolution layer are automatically updated by error back propagation algorithm.We proposed a new activation function named tReLU,which solves the failure problem of the ReLU activation function.Experimental results show that the proposed activation function can improve the training efficiency and detection accuracy of the DCNN model.In the detection part,we used the softmax function to classify all the features.Experimental results show that DCNN can effectively distinguish fake images from real images with high accuracy,and the proposed algorithm is superior to other algorithms.The DCNN model is very robust to the common image processing such as JPEG compression,salt and pepper noise,Gaussian noise and so on.We proposed a deepfake images detection method based on Residual Neural Network that can effectively distinguish deepfaked images and composite images from real images.In the proposed method,we presented a new ResNet network model named ResNet26 by improving Building_Block of ResNet18.Compared with the ResNet18 model,the ResNet26 model has stronger nonlinearity and higher fitting ability,which can detect deepfaked images and composite images,concurrently.Considering that there is subtle difference in texture between fake and natural images,we converted the image from the RGB color space to the YCbCr color space so that ResNet26 model can be able to learn fine feature well.Finally,we use Scharr operator to extract the edge information of Cb and Cr channels.The gray level co-occurrence matrices of the two edge image information matrices are used as the input of the ResNet26 model.We have verified the superior performance of the proposed ResNet26 model on multiple databases,and the experimental results show that our method has good ability to detect both deepfaked images and spliced images,and the model has a strong generalization ability.