Research On Image Forgery Detection Based On Deep Learning

With the continuous development of digital media and the devices of digital image processing,digital images are easy to be modified by criminals.In recent years,the altered images are utilized to commit crimes from time to time,so there is an urgent demand for studying the image forensics technology,aiming to authenticate the originality and authenticity of images through blind forensics technology,and then discriminate and localize the image tampering according to the traces left by image tampering.Image inpainting,image splicing,and image copy-move forgery are three most common and easiest ways to achieve image tampering,which can be implemented by some softwares,such as Photoshop,GIMP,and Meitu.Existing image forgery detection techniques mainly include traditional methods and deep learning-based methods.Traditional methods need to extract features manually,which are of poor applicability and low efficiency.Deep learning-based methods have made great progress in image tampering detection during recent years,which extracted features automatically by the convolution neural network.The networks can extract more discriminant and generalized features than traditional methods.Howevever,there are some shortcomings with the existing deep learning-based methods,mainly including:1)the detection result of forgery regions with small size is far from satisfactory;2)lack of feature recalibration,resulting in insufficient discrimination of extracted features.Specifically,focusing on the existing shortcomings,some image content forgery detection issues,including image inapinting detection,image splicing detection,and image copy-move forgery detection are studied.The main research contents of this thesis include the following aspects:(1)A feature pyramid network for diffusion-based image inapinting detection is proposed.Existing diffusion-based image inpainting detection methods only extract the change of Laplacian on the direction perpendicular to image gradients as the discriminative feature,which is not ideal for the detection of inpaint regions with small size,and is not robust to some image post-processing operations.In this thesis,the feature pyramid network is utilized to extract the features of different scales,and then the multi-scale features are fused to localize the inpainting regions of differnet inpainting size.In addition,a stage-wise binary cross entropy loss function is proposed,which adopts different loss functions in different training stages to accelerate convergence during training and improve the performance on localization of small inpainting regions.Experimental results have shown that the proposed network can not only improve the detection of small inapinting regions greatly,but also has strong robustness against several image post-processing operations.(2)A multi-task attention network for image splicing localization is proposed.The existing image splicing detection methods do not take the edges of the images into account and lack of study of low level features,which leads to inaccurate localization of the splicing edge and the splicing area.Aiming at image splicing detection,this thesis designs a multi-task network to obtain the edges of images,the edges of splicing regions,and the splicing maasks,which will lead to improvment on the localization of the edges of inapinting regions.In addition,the expression ability of discriminant features is improved by introducing the fusion of low level features in shallow network and high level features in deep networks.Finally,squeeze-excitation attention mechanism is proposed to recalibrate the fused features.The network would pay more attention to the features of spliced regions.Experimental results have evaluated the effectiveness of each module of the proposed network.The proposed method can improve the localization results of spliced region,especially the edges of the splice region.(3)A CNN-Transformer network for image copy-move source/target distinguishment is proposed.Most of the existing methods can only locate the copy-move region,while cannot distinguish the source and target regiona.The networks that can distinguish the source and target regions needs training in stages,and cannot localize the source/target regions end-to-end.Besides,these methods lack of the use of global information in images.In this thesis,a generative adversarial network based on CNNTransformer is proposed for copy-move source/target distinguishment.Both CNN and Transformer are introduced in the generated network.CNN is used to extract local features,while self-attention mechanism in Transformer is used to extract global features in images,which makes the network focus on the regions scattered in different locations of images but with related structures.Feature coupling layer is utilized to interact the features between CNN and Transformer.Finally,Pearson correlation coefficient layer is introduced to measure feature similarity so as to improve location accuracy of similar areas.Experimental results have evaluated the advantages of joint of CNN and Transformer.The proposed method can improve the performance on localization of copy-move regions,especially the performance on localization of source regions.