Research On Spectral Feature Modeling Methods For Multimedia Resampling Detection

With the rapid development of Internet and 5G communication technology,multimedia,as the mainstream information carrier,which includes speech,image,video and many other forms,plays an important role in people's daily life.At the same time,after continuous development and improvement,multimedia editing software has become more and more simple and popular.In the process of multimedia tampering,in order to make it verisimilitude from the senses,the tampered areas are often resampled,such as the scaling of speech(fast forward and slow playback),the geometric transformation of image(scaling,rotation,distortion,etc.).Therefore,the detection of multimedia resampling can assist multimedia tampering identification,which has high theoretical research and practical value.The existing multimedia resampling detection is still difficult and challenging.In the resampling detection brought by speech scaling,most of the existing algorithms extract features from one-dimensional signals,which can not well describe the traces of resampling,and the detection accuracy is not high enough.In the blind estimation of image rotation angle,most of the existing use the Fourier peak value to estimate,but the post-processing of JPEG will bring additional peak interference to the spectrum,which has a great impact on the estimation of rotation angle.In view of the above problems,this paper mainly analyzes the resampling characteristics of speech and image media,and separately gives a more effective detection method.The main research work of this paper is as follows:(1)In view of the resampling characteristics brought by speech scaling,this paper proposes a speech resampling detection method based on local features of spectrogram.Unlike most of the existing methods,which extract features directly from one-dimensional speech signals,this paper first transforms speech into spectrogram,and then exposes the traces of resampling on spectrogram.When the speech is resampled,the bandwidth of the spectrogram changes with the change of the sampling factor.Firstly,the relationship between the period of resampled speech and the harmonic frequency on the spectrogram is analyzed theoretically.Secondly,according to the change rule of the texture feature on the spectrogram,the local binary pattern operator is applied to model and extract the statistical feature.Finally,the support vector machine is used to detection the resampled speech.The experimental results show that compared with the existing algorithms,the proposed algorithm has significantly improved detection performance under different sampling factors,and has better robustness for MP3 compression with different bit rates.(2)According to the resampling characteristic of image rotation,a blind estimation method of image rotation angle based on cyclic spectrum analysis is proposed in this paper.After the image is transformed in the frequency domain,the image covariance shows the cyclostationary in the spectrum,and the positions of characteristic peaks is different with different rotation angle,so the rotation angle of the image can be estimated according to the coordinates of the characteristic peaks displayed in the spectrum.In order to reduce the interference of JPEG compression to the image,this paper first calculates the theoretical peak trajectory of the rotation angle,and searches the peak value in the four neighborhood of the theoretical trajectory;then uses the wavelet transform based on the semi soft threshold to denoise the searched peak value,so as to reduce the noise interference brought by JPEG compression;finally applies the super sorting algorithm to process the processed peak value in descending order,the closest point to the theoretical trajectory is selected within limits to estimate the rotation angle.The experimental results show that the proposed algorithm has better detection accuracy and smaller detection error in the case of JPEG compression.