Abstract: In the prediction-residual reconstruction framework, multi-hypothesis prediction based on temporal correlation is the key step of compressed video sensing reconstruction. This paper studies the accuracy prediction method by utilizing rich features based on deep learning, and a novel feature-domain multi-hypothesis reconstruction network for compressed video sensing (FMH_CVSNet) is proposed. In FMH_CVSNet, the feature domain multi-hypothesis prediction module (FMH_Module) is firstly proposed, which improves the prediction ability by reasonably constructing the motion estimation module and the hypothesis weight calculation module based on the characteristics of video signal. Secondly, the two-stage multi-reference motion compensation mode is proposed, which makes the constructed hypothesis sets much better for sequences with different motion and the further improves the prediction accuracy. Simulation results show that FMH_CVSNet achieves better reconstruction performance under various experimental conditions, improves by 4.76dB compared with the traditional multi-hypothesis algorithm 2sMHR and improves by 3.87dB compared with CNN based compressed video sensing reconstruction algorithm VCSNet-2.

Key words: compressed video sensing, deep learning, multi-hypothesis prediction, adaptive hypothesis weight, multiple reference frame, video motion feature