Research On Double JPEG Compression Detection And Location Algorithm Based On Difference DCT Coefficients And CNN

With the popularity of the Internet and the rapid development of multimedia technology,more and more digital images are being transmitted on the Internet.The constant emergence of various image acquisition devices and editing software has provided great convenience for people to shoot and modify photos.If some of the falsified photos are used in photography competitions,news reports or court forensics,it will have a bad impact on society.Therefore,in the past decade,people's attention and demand for the authenticity and originality of digital images has increased significantly,leading to the rise of digital image forensics research.More than 90% of digital images are saved in JPEG format,and most imaging devices and post-processing software output images in JPEG format.If you tamper with JPEG images and save new fake images in JPEG format,it will definitely introduce the trace of double JPEG compression.Therefore,forensic research on double JPEG images has high practical value.This thesis starts with the frequency domain characteristics of JPEG images,and combines the convolutional neural network to detect and locate double JPEG.The main research results of this thesis are as follows:Firstly,aiming at the problem that the existing algorithm has low detection rate when the small size image and the first compression factor are larger than the second compression factor,a double JPEG compression detection algorithm based on DCT domain CNN is proposed,which divides the image into texture and smooth region.The differential DCT matrix of the texture region is extracted,and the histogram is used to represent the feature.Then the CNN model is built,the parameters affecting the model training are discussed,and the best CNN model is established to classify the extracted features to detect JPEG compressed images.The experimental results show that the algorithm designed in this chapter can classify 64×64 primary and secondary JPEG compressed images,and the average classification accuracy is above 89%,even in the difficult case that the first compression factor is greater than the second compression factor.The algorithm in the thesis can also maintain the detection accuracy of about 80%,and its detection rate is the highest among the three algorithms.Secondly,based on the CNN model proposed in the previous paper,a tampering region localization algorithm for JPEG tampering composite images is designed.The JPEG images are superimposed into the CNN to obtain the probability that each block is compressed once,and the probability is integrated and the threshold is set.A tampering area is determined when each block probability is greater than the current threshold.The experimental results show that the proposed algorithm can achieve the locating of JPEG tampering composite image tampering region.The average positioning accuracy is above 0.85,which is higher than the comparison literature by about 10%,and it has strong robustness.Finally,the thesis develops a simulation demonstration system that is easy for users to understand and use,and can intuitively display the detection effect of the algorithm.