Research On Color Image Splicing Localization Algorithms Based On Deep Convolutional Neural Network

With the rapid development of digital image technology and the popularity of image editing software tools,users can easily edit and tamper images without leaving obvious visual traces.The emergence of forged images negatively affects many aspects of our lives,such as news,academics,politics,the Internet,and so on.Therefore,it is more and more important to study effective digital image forensics.To solve the drawbacks of the existing deep learning-based color image splicing localization algorithms,this thesis studies the following two aspects:(1)The existing color image splicing localization algorithms based on fully convolutional network(FCN)have achieved good performance,however they have the following two drawbacks:(a)they ignore the inherent correlation between color channels and the integrity of three channels;(b)they use conditional random field(CRF)to refine the results of FCNs,however the CRF doesn't model higher level consistency,such as region-level consistency.To solve the above drawbacks,an improved color image splicing localization algorithm is proposed.To solve the first drawback,quaternion FCN(QFCN)is proposed to generalize FCN to quaternion domain by using color image quaternion representation method and quaternion convolution.To solve the second drawback,QFCN and superpixel-enhanced CRF are formed an end-to-end learning system.Finally,the predictions of QFCN32,QFCN16 and QFCN8 are fused to get final result.The experimental results on three public forged datasets(CASIA v1.0,CASIA v2.0 and DVMM)demonstrate that the proposed algorithm achieves better performance than the existing algorithms in the accuracy,generalization and robustness abilities.(2)The existing color image splicing localization algorithm based on two-stream Faster R-CNN has the following two drawbacks:(a)it ignores the inherent correlation between color channels and the integrity of three channels and only considers single-scale features by Res Net;(b)it only locates spliced regions at block-level by using a bounding box.In order to solve the above drawbacks,this thesis proposes an improved two-stream R-CNN model for pixel-level color image splicing localization.To solve the first drawback,quaternion Res Net(QRes Net)is proposed to generalize Res Net to quaternion domain,and then feature pyramid network(FPN)based on QRes Net is used to extract multi-scale features.Regarding the second drawback,QFCN and superpixel-enhanced CRF are used to locate spliced regions at pixel-level.Moreover,attention region proposal network(ARPN)combining attention mechanism and region proposal network(RPN)is designed to pay more attention to important regions.The experimental results on the three publicly forged datasets show that the proposed algorithm outperforms the algorithm presented in(1).