Application Of Motion Analysis To Distributed Video Coding

Owing to the existence of non-stationary motion,some video segment may show a violent or gentle characteristic along the temporal axis,or exhibit different global or local spatial attribute.In DVC(Distributed Video Coding),it is important to analyze the characteristic of the motion non-stationarity from the encoder and decoder respectively,and apply the analysis results to the key modules of the encoder and decoder.Application of motion analysis can match variable video content,allocate the computation and communication resources of the codec dynamically and thus improve the performance of the entire system.The main researches are included in the following several aspects.Firstly,an interframe correlation based AKFS(Adaptive Key Frame Selection)is proposed for solving the motion non-stationarity problem along the temporal axis.Two low-level image features are presented to compute the interframe correlation.These features measure the motion situation of the video sequence globally and detailedly respectively,which is justified by the MPEG-7 motion activity descriptor.Then,an adaptive selection of GOP(Group of Pictures)size based on interframe correlation with low computational complexity is executed,matching variable video content and allocating the resources of the codec dynamically.Furthermore,a revised algorithm with an extra restriction on GOP size is proposed after observing simulation results from the interpolation process at the decoder,in order to reach the best performance when adaptive GOP size is applied.Compared to the PKFS(Periodical Key Frame Selection),it is proved that the AKFS can gain 0.6?2dB RD(Rate Distortion)performance improvement,and its revised algorithm can further achieve a 0.2-1 dB advanced RD performance over the unrevised algorithm.Secondly,a MSEC(Motion State based spatio-temporal Encoding mode Classification)is proposed aiming at the temporal and spatial motion non-stationarity.A frequency statistic on the interframe difference image is presented to classify the motion states for the guidance of the encoding process.Then,different frame-level and block-level encoding mode is adopted corresponding to the temporal and spatial dimensions respectively based on the different motion states.Meanwhile,a reference direction is indicated according to the different motion states between the current frame and the reference frames from the two different directions as the forward and backward directions respectively.Different encoding mode enables a dynamic allocation of the resources at the encoder and decoder,an effective elimination of the noise,and thus improves the objective and subjective performance of the entire DVC system.The experiments show that the MSEC can obtain up to 1dB RD performance improvement with low computational complexity and exhibit better subjective visual reconstruction,compared with the DVC algorithm without encoding mode classification.Finally,a STSM based SIG(Spatio-Temporal Saliency Map based Side Information Generation)is proposed for the application of motion non-stationary information to the decoder.Since the nonlinear motion is made by the motion subject with local motion,a STSM extraction algorithm is executed to differentiate the ROI(Region of Interest)and non-ROI.Then,a motion vector refinement algorithm and a motion vector smoothness algorithm based on the STSM are performed only at the decoder,in order to promote the precision and smoothness of the motion vectors with low computational complexity.The experiments indicate that the STSM based SIG exceeds the SIG in the traditional DVC with 0.5dB RD performance.