基于结构特征约束两阶段重构的视频压缩感知

针对现有视频压缩感知多假设预测-残差重构方法重构精度不高的问题,提出一种基于结构特征先验约束两阶段重构的多假设预测视频压缩感知方法。该方法从相似图像块非局部相似性和梯度稀疏性出发,将第一阶段多假设预测重构后的当前帧直接作为第二阶段重构的初始重构帧,利用低秩正则化和全变差正则化再次进行重构,其中低秩正则化矩阵是通过欧氏距离-感知哈希算法获取的图像相似块集合,同时包含帧内和帧间的相似图像块,充分利用帧内帧间的结构相似性,有效提高重构性能,为后续残差重构打下基础。仿真实验表明,所提两阶段重构算法较现有几种优秀重构算法更好地保留了视频帧的细节,并具有更高的重构精度。

Video compressed sensing based on two-phase reconstruction of structural feature prior constraints

To solve the problem that the accuracy of existing multi-hypothesis prediction methods for compressed video sensing is not high, a multi-hypothesis prediction method based on structural feature priori constrained two-phase reconstruction is proposed. Starting from the non-local similarity and the gradient sparsity of the similar image blocks, the current frame reconstructed by the multi-hypothesis prediction is directly used as the initial frame of the second phase reconstruction. The low rank regularization and the total variation regularization are utilized to perform the second phase reconstruction, wherein the low rank regularization matrix includes the similar blocks of intra-frames and inter-frames, fully utilizing the structural similarity between intra-frames, and improving reconstruction performance effectively, laying the foundation for the subsequent reconstruction of residuals. Simulation experiments show that the proposed two-stage reconstruction algorithm preserves the details of video frames better and has a higher reconstruction accuracy than several excellent reconstruction algorithms.