Research On Image Splicing Blind Forensics Based On Two Stream Convolutional Neural Network

With the popularization of Internet technology and the continuous development of smartphones,it is easier for people to access and use digital images.At the same time,the increasing power of image tampering technology makes people pay more and more attention to the authenticity of images.Image splicing blind forensics as image authenticity and integrity of technology has received more attention and to maintain social fairness and protect the truth is of great significance.With the increasing maturity of convolutional neural network technology,new research ideas are pointed out for splicing blind forensics,which can effectively improve the positioning accuracy of splicing blind forensics.At present,there are still some problems with splicing blind forensics.Due to the development of image post-processing technology,splicing tampering traces are hidden so that more and more difficult to be found,the location of tampered areas is becoming more and more challenging.At present,the mainstream splicing blind forensics method cannot fully extract tampering features in the feature extraction stage,and the pixel-level localization is not accurate.Based on this,this paper proposes an image splicing blind forensics algorithm（Two Stream Region-CNN,TW R-CNN）.The specific research contents are as follows:（1）Aiming at the problem of insufficient feature extraction of splicing tamper,this paper puts forward the two stream feature extraction network,that is,an illuminant stream and an image stream.Firstly,GGE（Generalized Grey-World,GGE）and IIC（Inverse-Intensity Chromaticity,IIC）Illuminant Map（IM）are introduced into the illuminant stream to extract the local features,that is,the illuminant color inconsistent between the tampered region and real region.Secondly,the image stream extracts the global features of the input image,such as unnatural tampering with boundary,the difference of the strong contrast.Finally,the two stream feature is fused at multiple scales based on the feature pyramid.The fused feature maps contain rich context information,which provides sufficient tamper features for the subsequent tampered region location.（2）Aiming at the problem of inaccurate pixel-level positioning,this paper proposes two improvements:1)the Spatial Attention Region Proposal Network（Attention-RPN）is proposed,which extracts the tamper candidate region containing pixel-level position information,and the feature representation of tamper candidate region is enhanced,2)K-means is used to cluster analysis of the anchor size of Attention-RPN,the anchor size of splicing blind forensics algorithm is calculated to improve the accuracy of bounding box regression.The algorithm in this paper is trained on the synthesized dataset,and tested on the Columbia,NIST16 and CASIA standard dataset.When using the GGE IM,the F₁ scores on the three datasets reach 73.1%,79.4%and 43.9%respectively,and when using the IIC IM,the F₁ scores on the three datasets reach 73.2%,81.0%and 45.2%respectively.Comparing the algorithm in this paper with other mainstream splicing blind forensics methods,when using GGE IM,F₁ scores on Columbia and NIST16 increased by 3.4%and 7.2%respectively;when using IIC IM,F₁ scores on Columbia and NIST16 increased by 3.5%and 8.8%respectively.Experimental results show that TW R-CNN proposed in this paper achieves better results on splicing blind forensics.