Research On Image Tampering Detection Technology Based On Deep Learning

In recent years,with the popularization of the Internet,digital images have become a significant medium for transmitting information among the world.However,with the fast growth of digital image processing technology,various kinds of user-friendly image editing software have emerged.Users can generate a large number of tampered images through simple operations,and these tampered images will bring serious harm to the justice,military,media and other fields.Therefore,image tampering detection has important research value.At present,the research methods of image tampering detection technology consist of traditional detection methods and detection methods based on deep learning.To be more specific,the traditional image tampering detection method is to manually extract features to detect tampered images and locate the tampered area.Due to the limitations of feature extraction,most of these methods can only detect one tampering operation.The image tampering detection method based on deep learning can detect a variety of tampering operations by extracting richer features through convolutional neural networks,but complex parameter calculation will cause the computing speed to be slow.In addition,post-processing operations will hide part of the tampering features,resulting in being unsatisfactory for the detection of tampered images that have undergone multiple post-processing operations.Therefore,in view of the above problems,this paper has been carried out in the field of image tampering detection technology based on deep learning.The main work includes the following aspects:(1)As in the existing tampered image datasets,there is no public and standard dataset that can be used for the detection of multiple tampering operations.Therefore,this paper has produced a comprehensive dataset suitable for the research.On the basis of the public dataset CASIA,the elimination function of Meitu Xiuxiu is used in this paper to create removal images and the labeling tool named Label Img is used to mark dataset with three kinds of tampering operations and rectangular boxes.(2)In view of the outstanding performance of the Faster R-CNN network in the field of object detection,this paper conduct an image tampering detection method based on the Faster RCNN network.The input tampered image is extracted through VGG16,and the extracted feature maps will be delivered to RPN layer to generate bounding boxes of tampered areas which can be used to subsequent tasks of classification and regression.Experimental results illustrate that the method is more effective for splicing operation,but less effective for copy-paste and removal operations.In addition,the computing speed is very slow.(3)In order to improve the calculation speed,this paper applies the RFCN network to the research of image tampering detection.In the feature extraction part,a full convolutional network is used to reduce parameter calculations,and a position-sensitive feature map is generated through a layer of special convolution.Classification and regression tasks are performed by voting which reduces the amount of calculation.Experimental results show that this method can achieve a detection speed of 12.99 images per second,but there is still room for improvement in detection accuracy.(4)In order to efficiently and accurately detect the three tampering operations of copy-paste,splicing and removal,this paper proposes a research method for multiple tampering detection based on dual-channel full convolutional network.To be more specific,in order to capture more types of tampering features to deal with post-processing operations,the input of the network is composed of RGB images and noise residual images filtered by SRM.The feature extraction part uses the Res Net-50 convolutional layer and GCT module to highlight useful tampering features,promoting the relationship between feature channels to boost detection accuracy.The final location-sensitive feature map is classified and regressed through voting.The experimental results illustrate that the method can detect three tampering methods of copy-move,splicing and removal,and the detection accuracy is increased to 75.55%.