Research On Image Tampering Detection Based On End-to-End Deep Learning

As a very commonly used information carrier in human society,digital images play an increasingly important role in people's information exchange.However,with the rapid popularity of image acquisition equipment and image editing software,non-professionals can easily acquire images and freely tamper with the content of the images.Nowadays,these tampered images will also use the Internet for faster and more widespread dissemination,which not only has a great negative impact on people's daily life,but also causes serious problems in the media,digital image forensics,politics and even the military.Therefore,more and more researchers have begun to pay attention to the problem of digital image forgery detection.The current digital image forgery detection methods that have been proposed can be mainly divided into image forgery detection methods based on traditional feature extraction and image forgery detection methods based on deep learning.However,these methods generally have some problems: image forgery detection methods based on traditional feature extraction are generally only effective for a certain forgery type,and their robustness and generalization ability are poor;The image forgery detection method based on deep learning requires a large number of forgery image samples to train the network for better detection accuracy,and its time complexity is relatively high.Aiming at the main problems of current forgery detection methods,this thesis conducts the following research on the basis of image forgery detection methods based on traditional feature extraction and image forgery detection methods based on deep learning:1.This thesis fully considers the advantages and disadvantages of image forgery detection methods based on traditional feature extraction and image forgery detection methods based on deep learning,and design and propose a deep learning method that integrates traditional feature extraction image forgery detection methods.This method is an end-to-end dual-encoder U-Net network(D-Unet),which employs an unfixed encoder and a fixed encoder.The unfixed encoder autonomously learns the image fingerprints that differentiate between the tampered and non-tampered regions,whereas the fixed encoder intentionally provides the direction information that assists the learning and detection of the network.In order to improve the global insight of the network,we also designed a spatial pyramid global feature extraction module behind the dual encoder.The experimental results show that this method has good detection accuracy without a large number of forgery images for training,and has good robustness against various attacks.2.Firstly,we got inspiration from the multi-task learning network,and then designed a multi-task learning network based on the idea of multi-task learning to complete the task of image splicing forgery detection.The proposed multi-task learning network contains three decoders to achieve three tasks respectively.The content decoder is used to distinguish the tampered and the non-tampered region,and the edge decoder is used to extract the edge of the tampered region,and the reconstruction decoder is used to complete the task of reconstructing the input image.The experimental results show that this method has both strong robustness and high detection accuracy.