Research On Image Tampering Operation Detection Based On Blind Forensics Technology

In today’s society,human beings have to receive and process massive information every day,most of which are digital information.Among them,digital images have become the main carrier for human beings to release and receive information.Digital image is composed of a series of digitized pixels.People can modify the number at will,and naturally,they can process the digital image at will.Although image processing can shine in photography,advertising and other fields,the malicious use of image processing technology will fabricate facts and lead to public opinion,and image tampering belongs to this category.Therefore,the blind forensics technology of tampering image has attracted more and more attention.At present,the tampering image detection algorithm based on deep learning has made great progress,but many of them need to be pre-processed to achieve the purpose of extracting inherent features.At the same time,they only focus on the detection and location of tampered images and ignore the tampering mask.Code extraction.In order to solve this problem,this paper proposes a tampered image detection,localization and mask extraction network that combines FasterRCNN and Full Convolutional Network(FCN).The network uses cascading RPN to modify the alignment of feature maps and anchor points in traditional RPN networks.At the same time,in order to increase the sensitivity of the network to background tampering,bilinear interpolation is used to replace the largest pooling layer in the ROI pooling layer.Finally,perform bounding box regression and classification.Input the regression bounding box and the characteristics of the cascaded RPN network into the FCN network to extract the tampering mask.In the training process of the network model,because there are few tampering data sets,this paper first uses the tampering data set composed of COCO data for pre-training,and then uses three standard data sets CASIA,COVER,and Columbia to fine-tune and test the network.During the training process,we found that the gradient of the network disappeared,so this article separates Faster-RCNN and FCN networks for training.During the experiment,this paper compares the cascading RPN and ROI pooling layer in the tamper detection and location network before and after improvement.It is found that after removing the maximum ROI pooling layer,the network model is improved by 6.24%mAP,and the cascading RPN to Faster-Improvement of RCNN network The F1 score on the three standard data sets has increased by at least 4%,and the AUC has increased by at least 3%.Comparing the falsified mask extracted through the FCN network with other algorithms,it is found that the effect of the algorithm in this paper is significantly better than other algorithms.Finally,the robustness and superiority of the network are evaluated,and the experimental results show that the algorithm in this paper has good robustness and ideal superiority.