Median Filtering Forensics Research Based On Automatic Feature Learning

Recent research has demonstrated that median filtering can be adopted as a postprocessing procedure to cover the tampering traces in many image forgeries.Hence,median filtering detection has become an important branch in digital image forensics.Most existing median filtering forensic techniques detect median filtering traces using handcraft features,but their performance degrades with smaller image size and stronger compression.To get more detailed features,automatic feature learning becomes a viable approach.To get more detailed description of image local texture,a discriminative multi-scale sparse coding based approach is proposed in this paper.At first,a multi-scale patch domain modeling is done to capture median filtering patterns at different levels.And then by associating label information with each dictionary item,the learned overcomplete dictionary can be more discriminative and the samples with different labels tend to select different dictionary atoms.Hence,different samples have different texture features.Based on the discriminative sparse codes,maxpooling is then employed over all corresponding patches to obtain representative features for each specific patch size.The final features are derived by concatenating all the features of various patch sizes.The experiment results show that the proposed approach can achieve median filter detection in JPEG images and can distinguish median filtering operation with other image manipulations in most cases.Median filtering is an image post-processing operation without changing image content,therefore the images with median filtering operation and without median filtering operation have a lot shared information,which can degrade detection performance to some degree.Therefore,a low rank shared dictionary based sparse coding median filtering forensics algorithm is proposed in this paper.Through separating the shared dictionary part with the overcomplete dictionary,a more accurate image representation can be achieved.More specifically,by limiting the shared dictionary low-rank can avoids the shared dictionary containing the characteristic dictionary elements related to classes and by limiting the sparse coefficients similar can make the shared sparse coding coefficient contains little information about classes.Through setting the low rank and coding similarity limitations to the shared dictionary,the feature expression capability of the characteristic dictionary can be significantly improved.Extensive experiments show that our proposed algorithm can obtain a better performance by only using a 512 dimensional characteristic overcomplete dictionary.Additionally,our proposed algorithm can localize the positions of median filtering operation in compressed JPEG images,which is of importance in practical applications.