Research On Image Inpainting Forensics With Multi-Link Features

With the widely used of computer network and image acquisition equipment,the images has been widely accepted as information dissemination carriers,which widely used in various fields of society.However,with the increasing maturity of image editing software function,more and more non-professionals can easily edit the image content,posing a great challenge to the authenticity and integrity of images.Therefore,in order to detect the authenticity and integrity of images,image passive forensics has become an important research content in the field of information security.Although image inpainting is an effective image editing technique,there are few researches for inpainting forensics.Therefore,aiming at patch-based inpainting,this paper proposed an image inpainting forensics scheme.The main research content and innovations of this paper are as follows:(1)Patch-based image inpainting will cause the relatively high similarities between an inpainted patch and its reference patches.Therefore,to assess the similarity of image patches,this paper constructed a comprehensive feature descriptor combining the root mean squared error,the weighted zero connectivity length and the position distance between a patch pair.At the same time,multi-link feature vector is obtained,which is a vector of feature values computed between a patch and a set of most similar patches.This feature can characterize the inpainting information more comprehensively and help to improve the performance of image inpainting forensics.(2)In order to classify the image patches efficiently,a distribution model has been established for the elements of the multi-link feature based on two deformed Laplacian distributions.By analyzing the distribution differences of multi-link feature elements of different classes,the parameters of the distribution model are deduced,and then the discriminant criterion of the inpainting classifier is obtained.(3)On the basis of the initial location of inpainting region,this paper designed a false alarm patches removal strategy.The strategy explored the size,self-relation degree of a suspicious region,and the relation between two image regions.Then,false alarmed uniform region,reference regions and small suspicious regions can be removed layer by layer.By applying this strategy,the false alarm performance issignificantly improved while keeping the detection accuracy unchanged.Compared with the state-of-the-art inpainting forensics methods,the proposed inpainting forensics scheme has better visual quality in the detection of the inpainted area,objective performance indicators like the true positive rate and the false positive rate have been improved dramatically.Moreover,it is very robustness against different inpainting operations.