Research On Multi-hypothesis Prediction-Based Reconstruction Algorithms For Compressed Video Sensing

The traditional video acquisition is based on the Nyquist sampling theory. A video signal is first sampled at a high-speed, then a large amount of redundant data is discarded via complex compression algorithms in traditional video coding to achieve efficient storage and transmission. This approach causes huge waste of sampling resources, as well as high coding complexity, making the traditional video coding schemes unsuitable for the new application scenarios with a resource-deprived sampling side, e.g., the wireless video surveillance and the wireless multimedia sensor networks, etc.Compressed sensing(CS) conducts sampling and compression simultaneously, saving enormous sampling resources while reducing the sampling complexity significantly, thus is suitable for the application scenarios with a resource-deprived sampling side. Compressed video sensing applies the compressed sensing technology to video acquisition and reconstruction, which implements independent acquisition and joint reconstruction.Exploiting inter-frame correlation to achieve high video reconstruction quality is main technology in compressed video sensing, as well as the research focus in this thesis. Since multi-hypothesis motion compensation enables to effectively excavate the correlation among video frames, this thesis focuses on the research of compressed video sensing reconstruction algorithms based on multi-hypothesis prediction. The major work and research results of this thesis are as follows:1. The existing Tikhonov-regularized multi-hypothesis prediction algorithms use all possible search blocks as hypotheses, which causes a high computation load and impairs the prediction accuracy. Furthermore, the Tikhonov regularization weighted by Euclidean-distance cannot precisely reflect the similarity between current block and a hypothesis, causing regularization distortion problem. To address these issues, this thesis proposes an optimal multi-hypothesis selection scheme and an improved Tikhonov regularization metric combining Euclidean distance and correlation coefficient. Furthermore,by combining with the multi-reference frame technique, this thesis proposes a multi-reference frame-based optimal multi-hypothesis prediction algorithm(MRMH). Additionally, we also apply frame-based DPCM quantization to the CS measurements to improve compressionefficiency. Simulation results show that the proposed optimal multi-hypothesis selection scheme greatly reduces the computational complexity of multi-hypothesis prediction while improving the prediction accuracy effectively and the proposed MRMH algorithm can obtain high video reconstruction quality.2. Most of the existing compressed video sensing multi-hypothesis prediction algorithms are proposed in measurement domain. This prediction approach will cause block artifacts and decrease reconstruction accuracy due to the restriction of inflexible block partitioning. To address this issue, this thesis proposes a two-stage multi-hypothesis reconstruction scheme(2sMHR) and designs two implementation schemes: a frame-wise scheme and a GOP-wise scheme. Simulation results show that the proposed 2sMHR can effectively reduce block artifacts and obtain high video reconstruction quality. Meanwhile, we also analyze the rate-distortion performance of the proposed 2sMHR under frame-based DPCM quantization.