| Median filtering is a common used method in image tampering,which will change the correlations among adjacent pixels to some extent.To implement the blind detection of median filtering,researchers have proposed many approaches.The main problems of the these method exist in the difficulties on median filtering feature extraction,especially in the case of low resolution and low JPEG compression quality factor.In recent years,deep learning has been successfully applied to various fields related to computers.This paper studies the application of deep learning in image median filtering forensics.In this paper,we address this problem by proposing a novel approach based on DCT-domain convolutional neural network(CNN).Two different median filtering forensics CNN are designed,called ATMnet and AWMnet.Firstly,the traces left by the median filtering operation in the image are analyzed,and the variation of the image frequency distribution caused by the filtering operation is studied.Images are converted to the DCT domain by DCT transform before further processing.Secondly,considering that the traces of the median filter may only exist in a specific frequency band of the DCT coefficient matrix,it is necessary to select the most relevant DCT coefficients.Therefore,we proposed a frequency threshold thresholding method and a weighting method to process the DCT coefficients to highlight the importance of different DCT coefficients.Thirdly,based on these two method,we have developed two new kinds of CNN layer,called adaptive thresholding layer and adaptive weighting layer,that can suppress the useless DCT coefficients and automatically extract median filtering traces.Furthermore,these two layers can help the CNN to focus on the median filtering artifacts rather than the image contents.Based on the proposed adaptive layer,two median forensic CNN architecture are designed.We use multi-scale competitive convolution to enable the network to adaptively select and fuse features of different scales.Finally,the one of the proposed median filter forensic CNNs is applied to the recovery problem of median filtered images,and a restoration model De MFGAN based on generated confrontation network(GAN)is designed.The discriminator of De MFGAN is AWMnet proposed in this paper.The generator is a special network with skip-connection and no pooling operation to ensure that information will not be lost during the transmission process.In order to focus on the detail portion of an image,the generator is designed to learn only the residuals of the filtered image and the original image.To further improve the performance of our De MFGAN,we proposed a new loss function which consists of an adversarial loss,a spatial domain loss and a DCT-domain loss.Experimental results demonstrate that our proposed methods are able to accurately detect the median filtering manipulation and outperform some state-of-theart methods,especially in the scenarios of low resolution and JPEG compression quality factor.Moreover,the proposed method can also distinguish between some different image tampering operations.By using the proposed two CNN architectures,median filtering forensics no longer needs hand-designed features,and is able to adaptively learn the most useful manipulation information.In the recovery problem of median filtered images,the proposed De MFGAN also achieved good performance.the experimental results shows that the De MFGAN is able to specifically recover the details of the median filtered image and has better performance than other deblurring methods. |
