Dictionary Learning-based Multi-view Compressive Sensing Video Coding

With features of three-dimensional and interactivity,3D stereoscopic video represents thedevelopment direction of future video applications，and to obtain depth information and othergeometric information in the decoder，multi-view videos should be compressed in encoder whichare obtained by multiple cameras shooting the same scene. In multi-view video coding, the biggestchallenge is the hardware pressure brought by massive video data.In this thesis, compressive sensing and distributed video coding are combined together to givefull play to their respective advantages and a new distributed compressive video sensing codingscheme is proposed to process multi-view video coding, the main research methods and results areas follows:(1) Contourlet transform based compressive sensing：to solve the traditional orthogonaltransform’s problem of limited sparse approximation capacity in two-dimensional signal, Contourlettransform is adapted to realize the sparse approximation of video frame, and with the feature ofmulti-directional Contourlet can compensate for the directional shortcomings of discrete wavelettransform;(2) Signal sparse representation by learning dictionary：the MOD and K-SVD are used to traindictionary in block-based compressive sensing, and improved OMP algorithm is adapted for sparsecoding in MOD and K-SVD which can accelerate the convergence of training dictionary andimprove the real-time performance of reconstruction;(3) Distributed compressive video sensing based on dynamic sampling rate allocation： Thedistributed video coding scheme is adapted in multi-view video coding and video frames aredivided into two kinds of frame the K frame and the CS frame which is compressed andreconstructed with different sparse basis and sampling rate. What’s more, motion estimation andinterpolation are adapted to compute time side information and spatial side information which areboth used to train over-complete dictionary, thus the reconstruction quality is future enhanced withthe introduced of spatial correlation.Experimental results show that multi-view video coding base on distributed compressive videosensing can give full play to the advantages of compressive sensing and distributed video coding.Benefit from the strategy of Encoded separately, joint decoding large amount of calculation istransferred from the encoder to the decoder and the reconstruction quality improves to a certain extent. At last, with the help of compressive sensing, the scheme can reduce the coding rateeffectively while enhancing the anti-error performance.