关键词: 视频压缩感知, 深度学习, 帧间组稀疏表示, 混合递归网络, 运动估计, 增强重构

Abstract: The traditional iterative optimized based video compression sensing algorithms are limited by long running time and low adaptability of parameters,resulting in low practicability and generalization. Taking advantage of the powerful computing power, fast speed and learnable parameters of neural networks, this paper first proposes a group sparse representation network (VGSR-Net), which maps the image block group to a higher-dimensional sparse domain through convolution, and uses a learnable threshold to denoise and extract inter-frame correlation. On this basis, a two-stage recursive enhance reconstruction network (2sRER-VGSR-Net) is proposed. First, we perform VGSR-Net to preliminarily enhance the initial reconstruction and then introduce STMC-Net as motion estimation, and the compensated frames are fed into the residual reconstruction network to further extract the missing detail and enhance the current frame. The second stage of reconstruction adopts a hybrid recursive structure with the aim of making full use of the existing better quality reconstructed frames. The simulation results show that the proposed algorithm improves the PSNR (Reak Signal to Noise Ratio) by 1.99dB compared with the existing state-of-art traditional compressed video sensing reconstruction algorithms SSIM-InterF-GSR, while improves the PSNR by 4.60dB with the comparation of the network-based algorithm CSVideoNet.

Key words: compressed video sensing, deep learning, group based sparse representation, hybrid recursive network, motion estimation, enhancement reconstruction