电子学报 ›› 2020, Vol. 48 ›› Issue (2): 249-257.DOI: 10.3969/j.issn.0372-2112.2020.02.004

• 学术论文 • 上一篇    下一篇

CVS中基于视频运动特征的多假设-双稀疏重构算法

郑学炜, 杨春玲, 禤韵怡   

  1. 华南理工大学电子与信息学院, 广东广州 510640
  • 收稿日期:2019-01-22 修回日期:2019-04-23 出版日期:2020-02-25
    • 通讯作者:
    • 杨春玲
    • 作者简介:
    • 郑学炜 男,1993年出生于广东潮州,华南理工大学电子与信息学院研究生.研究方向:视频压缩感知.E-mail:xuewei.zheng@foxmail.com;禤韵怡 女,1995年生于广东佛山,华南理工大学电子与信息学院研究生.研究方向:视频压缩感知.
    • 基金资助:
    • 广东省自然科学基金 (No.2017A030311028); 广东省自然科学基金 (No.2016A030313455)

Video Motion Features Based Multi-Hypothesis-Dual-Sparsity Reconstruction Algorithm in Compressed Video Sensing

ZHENG Xue-wei, YANG Chun-ling, XUAN Yun-yi   

  1. School of Electronic and Information Engineering, South China University of Technology, Guangzhou, Guangdong 510640, China
  • Received:2019-01-22 Revised:2019-04-23 Online:2020-02-25 Published:2020-02-25

摘要: 针对目前视频压缩感知重构算法对不同特征的视频序列重构质量参差不齐的问题,结合双稀疏对轮廓、细节的高清晰重构以及多假设算法对高频噪声有效抑制的优点,本文提出一种基于视频运动特征的多假设-双稀疏重构算法(VF-MH-DSR).基本思路是基于每个视频组(GOP)的运动特征,采取相应的多假设-双稀疏重构策略.首先给出一种观测域多维度参考帧的多假设重构算法(MD-MRF-MH)及其最优相似块个数设置方案;然后给出一种像素域多假设参考帧的重构算法(PD-MRF-MH)及一种高性能双匹配准则;最后介绍了视频信号运动特征判定方案及多假设-双稀疏重构的具体实现方案.仿真实验表明,本文所提多假设-双稀疏重构算法相对于目前较好的多假设预测重构算法2sMHR及组稀疏重构算法SSIM-InterF-GSR,重构性能平均提升了1.98dB和0.84dB.

关键词: 视频压缩感知, 双稀疏表示, 多假设预测, 视频运动特征, 相似块组, 匹配准则

Abstract: The existing approaches to reconstruct compressed video sensing achieve heavy quality fluctuation when reconstructing videos with different motion feature.To solve this problem, combing the merits of two CS(Compressed Sensing) methods:the clearly edges and fine details reconstruction of the dual sparsity representation and the effectively high frequency noise suppression of multi-hypothesis prediction, this paper proposes a video motion features based multi-hypothesis-dual-sparsity reconstruction algorithm(VF-MH-DSR) for compressed video sensing(CVS).The basic thinking of VF-MH-DSR is that adopting a corresponding MH-DSR method to each video group(GOP) based on their motion features.In our approach, we firstly develop a multi-hypothesis reconstruction algorithm based on multi-dimension reference frames in measurement domain(MD-MRF-MH) and a kind of setting scheme for optimal similar block.Then, multi-hypothesis reconstruction algorithm based on multi-dimension reference frames in pixel domain(PD-MRF-MH) and a double matching criterion to improve matching accuracy are proposed.Finally, we develop a strategy to determine the video motion feature and introduce the scheme of multi-hypothesis-dual-sparsity reconstruction.Simulation results show that the proposed VF-MH-DSR outperforms the existing state-of-art compressed video sensing reconstruction algorithms 2sMHR and SSIM-InterF-GSR by 1.98dB and 0.84dB respectively.

Key words: compressed video sensing, dual-sparsity representation, multi-hypothesis prediction, video motion features, the group of similar blocks, matching criterion

中图分类号: