Research On Video Source Camera Identification

Video source camera identification is a very important research topic in the field of digital multimedia information forensics.Due to the popularity of imaging devices,video images taken by various digital cameras or smart phones have sprung up on social networking platforms.Some of these videos which arc involved in crime,arc the evidence of court litigation,and tracing the source of filming equipment is one of the key links in the judicial forensics.However,large amounts of video datas do not contain any informations about their filming equipments,which forces us to analyze the source device from the video data itself.Video source camera identification of passive forensics is such a technology,in the absence of any embedded metadata about the imaging equipment,digital signature,watermark and other information,analyzing its shooting equipment from the video data itself.In this thesis,the video data taken by digital cameras and smart phones are studied to explore new methods to better obtain the sensor pattern noise and other characteristic information that can fully represent the physical fingerprint of the shooting equipment,so as to further improve the accuracy and effectiveness of video source camera identification.Specific innovations include the following aspects:1.A phase extraction algorithm based on two-dimensional discrete fractional Fourier transform(FFT)on multi-rotation Angle is proposed.To solve the problcm of poor recognition performance of video source cameras due to the limited ability of camera fingerprint representation extracted by existing mode noise extraction methods in processing low-bit rate video signals,this thesis puts forward,firstly,transforms the total noise pattern of each video frame into fractional Fourier transform domains with different rotation angles.Then the module of each element of each transformed matrix is normalized,secondly,the fractional Fourier inverse transformation of each normalized matrix is arried out and its real part is taken,finally,the processing results corresponding to multiple rotation angles of the same frame are averaged,and the mode noise of the video sensor is average od the results after the above processing in all frames.Simulzation results show that compared with the two existing algorithms,the proposed algorithm unpruves the performancc of video source camera identification.2.A feature extraction method based on the null space projection transformation of the odd order statistical moment eigenvector of video frame noise is proposed.There is room for further improvement in the recognition rate of low-bit rate video source digital cameras with time-varying statistical characteristics.In this thesis,firstly,the total noise pattern of each video frame is converted into a one-dimensional vector to calculate their odd order statistical moment matrix.In order to reduce the computation,principal component analysis is used to reduce the dimension of these matrices,and only the main components of each matrix are obtained.Then these components of all frames constitute the initial feature vector of each video.In order to minimize the intra-class distance and maximize the inter-class distance of the initial feature vector,linear discriminant analysis is carried out.Since many eigenvalues of the distance matrix within the class are close to zero,we solve the null space solution vector of this matrix,and then transform the eigenvector of each video into the projection matrix formed by the solution vector to form the final feature vector.We find that these new feature vectors are linearly separable in pairs.Therefore,a set of perceptrons is used to identify the video source camera.Computer simulation results show that the proposed method is superior to the traditional methods based on correlation and support vector machine.3.An empirical mode decomposition based sensor pattern noise preprocessing algorithm is proposed.At present,most of the researches on source camera recognition are carried out by extracting sensor pattern noise representing camera fingerprint information from images.However,when these methods are applied to video taken by smart phones,the recognition performance may be improved.Considering the characterization of camera fingerprint sensor model belongs to a special type of high frequency noise component,and the raw pattern noise extracted from the frame not only contain other noise from similar functions in the processing units also have video content details of high frequency components,in this thesis,aiming at a few selected frames in video,the raw pattern noise of every frame respectively according to the row and column is converted into a one dimensional vector,and then the vector is processed with the empirical mode decomposition,so,the intrinsic mode function IMFs with different instantaneous frequency components are get.By means of numerical simulation,the selection of the high frequency component resulting from the empirical mode decomposition is completed.This method is applied to video source recognition of three smart phones.The experimental results show that the proposed algorithm has better recognition performance than the existing five algorithms.