Image Inpainting Forensics Combined Discrete Cosine Transform With Deep Neural Network

With the rapid development of multimedia technology,digital image is one of the multimedia information carriers.Digital image is widely used due to its good characteristics such as scene and intuition.At present,the development and application of various image processing softwares enable users to easily and quickly edit digital images without mastering relevant professional knowledge to meet their own work needs or daily entertainment.But if they are illegally used by the malicious tampers in the fields of military,court evidence,news,etc.,it will bring some disturbance to the society,“seeing is believing”,the traditional idea is also being broken.Therefore,there is an urgent need for a forensic algorithm to identify the authenticity of digital images and to locate tampering regions of falsified digital images.This paper first analyzes the problems that traditional forensic methods require manual extraction about the inpainting features,high computational complexity,and low detection accuracy.After studying the process of image inpainting technology and neural network for feature extraction,this paper proposes a method for image inpainting forensics using deep neural network.First,the forensic network are provided with tags of the same size as the input images,thereby assigning a category tag to each pixel of the digital images.Then,using convolution,pooling and other operations to extract features closely related to the inpainting operation,and finally locating the inpainting area of the digital images.By comparing the forensic results of different forensic schemes,the effectiveness of the proposed forensic scheme is obtained.The experimental results show that compared with the forensic results of other forensic schemes,the proposed algorithm makes the detection accuracy higher than 91%.Then,in order to further analyze the robustness of the proposed scheme,the inpainting images are subjected to certain post-processing operations such as compression and scaling.After experiments,it is found that the detection performances of the proposed forensic method have greatly reduced.Through the analysis of the post-processing,it is known that the compression and scaling are equivalent to the low-pass filter,but not the ideal low-pass filter,which will have a certain influence on the high-frequency information of the images.Therefore,the discrete cosine transform is used to reconstruct the images from the spatial domain to the frequency domain,minimize the using of high-frequency information,and then using the discrete cosine inverse transform to return the images to the spatial domain,and finally sending it to the deep neural network for detection of the inpainting area.So a new image inpainting forensic method combining discrete cosine transform with deep neural network is proposed.The experimental results show that the proposed algorithm achieves high detection accuracy for the inpainting images without post-processing,and has strong robustness to JPEG compression and scaling operations.In summary,the main work of this paper focuses on the research on the location of the tampering area of the inpainting image,which has achieved certain results in both theory and application.These results will positively promote the development of multimedia information security.