Research On Color Image Splicing Detection And Localization Algorithms

With the development of information technology,digital image has been widely used in all walks of life as a digital media.The rapid development of powerful image editing software make more and more non-professionals edit and tamper digital images without leaving any obvious visible artifacts.The emergence of these tampered images brings great challenges to justice,criminal investigation,reputation protection and even social stability.Therefore,digital image forensics is becoming more and more important.For digital image forensics,this thesis focuses on the following three aspects of research works(1)The existing color image splicing detection algorithms usually process the three color channels of image respectively,thus ignore the holistic property of the three channels and the inherent correlation between them.Therefore,image depth information is introduced into the color image quaternion representation(QR)to improve it.Then,quaternion pseudo-Zernike moments are defined for RGB-D image,and quaternion pseudo-Zernike moments feature are extracted from color forged images.Finally,quaternion back-propagation neural network is combined with the extracted features for color image splicing detection.Experiments on four public datasets(DVMM,CASIA v1.0,v2.0,Wild Web)compare the performance of the proposed splicing detection algorithm in different color spaces,and the performance between the proposed algorithm and several existing algorithms(2)Deep learning-based splicing localization algorithms utilize network training to obtain expected features without extracting explicit features.Among the existing deep learning-based algorithms,the algorithm based on fully convolutional network(FCN)can achieve pixel-level accuracy.However,it still has the following two drawbacks:(a)FCN is not sensitive to image details;(b)condition random field(CRF)is employed as a standalone post-processing step,which is independent of the FCN,thus it is difficult to achieve the simultaneous optimization of FCN and CRF.For the first drawback,region proposal network(RPN)is introduced into FCN to enhance the training of object regions.For the second drawback,the use of CRF is changed to make the whole network an end-to-end learning system.Moreover,the results of three FCNs(FCN8,FCN16,and FCN32)with different upsampling layers are synthetically considered.Experimental results on three publicly available datasets(DVMM,CASIA v1.0,v2.0)verify the accuracy,generalization and robustness of the algorithm(3)The existing deep learning-based image forgery localization algorithms have two drawbacks:(a)with the depth of network growing,their performance sometimes do not improve,and the network may become hard to converge;(b)they consider either original image itself or the image residual as the input of the network.For the previous drawback,the residual block in ResNet is introduced to improve the structure of FCN.For the latter drawback,both of the original image and the image residual are used as the inputs of two different branches of the network,and then are considered in network training.Experiments on three publicly available datasets which are same as(2)show that the proposed algorithm outperform the algorithm presented in(2)in the accuracy,generalization ability and robustness.