Research On Reconstruction Algorithm Based On Sparse Representation And Multi-hypothesis In Compressed Video Sensing

The traditional video coding framework based on Nyquist sampling theory adopts the method of first sampling and then compressing.The video signal is efficiently encoded and compressed,then transmitted to the client for downloading and decoding,such as broadcast and video–on-demand systems.But this solution has problems such as high computational complexity and poor real-time performance in the encoder.Wireless video surveillance and mobile video telephony have developed rapidly in recent years.This type of technology prioritizes the wireless and portable factors of the client,so it requires simple video compression algorithm with low complexity.Due to the limited power consumption of the terminal hardware devices,the traditional video coding framework is no longer suitable for this field.Compressed Sensing(CS)theory proposes: sparse or compressible signals can be sampled with less than Nyquist sampling rate,and accurately recovered by some suitable reconstruction algorithms.It solves the problem of resource limitation on the encoding side of some application areas.The process of sampling and compressing a video signal based on CS theory is called Compressed Video Sensing(CVS).Based on the different sparsity characteristics and motion features of video signals,this paper completed the following two parts of research work:1.Based on the sparse characteristics of video signals in different representation domains,this paper proposes a Dual-Sparsity Reconstruction algorithm based on Multi-dimension Reference Frames(MRF-DSR).Firstly,a dual-sparse reconstruction model is proposed,which uses the group sparse and laplace sparse to describe the reconstructed videos.Secondly,the concept of multi-dimensional reference frame is proposed,a half-pixel reference frame and a scaling reference frame are introduced to obtain much more similar blocks;Finally,a diamond shape fast searching algorithm is proposed to achieve wide searching range with lower complexity.Simulation results show that compared with the existing optimal compressed video sensing reconstruction algorithms,MRF-DSR algorithm has better reconstruction performance on both subjective and objective criteria.2.Multi-dimensional Reference Frames have the advantage of excellent matching performance,and Multi-Hypothesis prediction-residual reconstruction algorithm has the advantage of low complexity.Taking advantage of these two algorithms,this paper proposes a Multi-Hypothesis reconstruction algorithm based on Multi-dimension Reference Frames(MRF-MH).In which,the optimal similar block number setting scheme and the double matching criterion are proposed to preserve much more similar blocks so as to improve matching accuracy and Multi-Hypothesis prediction accuracy.MRF-DSR algorithm can reconstructfor the videos' contours and details clearly,while MRF-MH algorithm is effective for suppressing high-frequency noise.taking advantage of DSR and MH,this paper proposes Video Motion Features based Multi-Hypothesis-Dual-Sparsity Reconstruction algorithm(VFMH-DSR),which can obtain the optimal reconstruction results for videos with different motion features.