Research On Image Median Filtering Forensics And Restoration Based On Deep Learning

Median filtering is a non-linear filtering technique,which is often used to hide the traces of tampering operations and smooth the image.In the field of multimedia security,whether a digital image has undergone median filtering is an important prerequisite for judging the authenticity of an image.Therefore,the forensics of median filtering is of great significance.Although recent research has made some progress,the existing median filtering forensics methods still underperform in the case of low resolution and image compression.In addition,the image is blurred after median filtering due to its characteristics of smooth image.However,the existing median filtering restoration method is difficult to improve the visual quality of the image.To overcome these shortcomings,median filtering forensics method based on deep learning is studied in this thesis,and to solve the degradation of image caused by median filtering,image restoration method based on generative adversarial network is studied in this thesis.The main contributions of this thesis are summarized as follows:1.To realize median filtering forensics,a median filtering forensics network based on discrete cosine transform and attention mechanism is proposed.First,based on the analysis of median filtering,this thesis finds that median filtering mainly affects specific frequencies in the frequency domain.Therefore,this thesis uses discrete cosine transform to transform the image into the frequency domain as the input of the network.Then,in order to enhance the features related to forensics,a preprocessing layer based on the channel attention and self-attention mechanism is designed to weight the information of different frequencies.Finally,a multi-scale feature extraction layer is designed to further extract forensics features in the frequency domain.The comparative experiments show that the forensics method proposed in this thesis performs better than the existing forensics methods under different circumstances,and the accuracy of forensics on small images and compressed images improves significantly.2.To solve degradation problem of median filtered image,a median filtering image restoration network based on generative adversarial model is proposed.First,this thesis uses the median filtering forensics network as the discriminative model.Then,in order to extract the information of multiple scales in the image,a generative model based on the multi-scale encoder-decoder structure is designed.In addition,the feature fusion module and error feedback module are introduced into the decoder structure to enhance the performance of the network.Finally,this thesis improves the loss function of the network.The experimental results prove the superiority of the method and the effectiveness of the network structure proposed in this thesis.