Research On Video And Image Reconstruction Algorithm Based On Compressive Sensing

Compressive sensing is a novel signal sampling theory with unique ability of compressing a signal during the process of sampling under the condition that the signal is sparse or compressible, which breaks the constrain of traditional Nyquist sampling theorem that the sampling rate must be no lesser than twice the baseband bandwidth and reduces the amount of data in the process of signal acquisition, storage and transmission. This theory has excellent features, such as low performance requirements of hardware, dynamic allocation of computational complexity and high reconstruction accuracy. So, there is no doubt that this theory can be applied to the field of video and image processing effectively. This thesis puts an emphasis on the research of video and image reconstruction algorithm based on compressive sensing, the main contents are as follow:1) A macro view is given about compressive sensing theory from the aspect of signal sparse representation, measurement matrix and reconstruction algorithm. elaborating on the current several classical compressive sensing reconstruction algorithms, the reconstruction performance is analyzed comparatively through one-dimensional signal and two-dimensional image at the same time.2) Given that the sparsity of signal can’t be a prior knowledge, in order to have a comprehensive consideration of the reconstruction precision and execution efficiency, an improved algorithm is proposed in this thesis. Firstly, by applying Discreet Cosine Transform to the vector of residual correlations to estimate the maximum number of atom needed by the support set; then do appropriate adjustment to large threshold parameter by a factor which is positive correlated with sampling rate and optimize the atoms that chosen by setting a threshold value; finally realize the close approach of signal sparsity and precise reconstruction of the signal step by step under the frame of St OMP. The proposed algorithm shows good adaptive feature without prior information of the sparsity and experimental results demonstrate that this algorithm not only keeps the low reconstruction complexity but also shows better and more stable reconstruction quality than the original St OMP algorithm.3) Surveillance video has strong temporal and spatial correlation and it is essential to maximally exploit the redundancies. Based on the traditional video coding methods and the characteristic of compressive sensing that can fully exploit the signal sparsity, this thesis proposed a video sequence reconstruction strategy for surveillance scenario based on compressive sensing. Key frames of a video sequence are chosen through Euclidean distance algorithm, then the sequence is divided into groups appropriately to formulate intra-frame and inter-frame coding strategy; key frames are measured at a high sampling rate and residual frames at a relatively lower sampling rate to improve compression ratio to the greatest extent; appropriate data blocking and deep sparsification by setting different thresholds to key frame and residual frame modules are applied to improve reconstruction performance. The experimental results show that this scheme has preferable performance for both slow and drastic detail variation video sequence while keeping a certain compression ratio, which has a bright application prospect in the field of video surveillance.