An optimal learning parameter for running Gaussian-based referenced compressive sensing

One of the approaches to exploit temporal redundancy in compressive sensing reconstruction of spatio-temporal signals is the Running Gaussian-based Referenced Compressive Sensing. It uses the weighted-average of all prior reconstructed instances as a reference to reconstruct the next instance with high accuracy. The performance of this approach depends on the weight called learning parameter. This work studies the relationship between the learning parameter and the reconstruction accuracy. We show that the small value of the learning parameter is more suitable for natural signals with dynamic sparse supports. We also propose a dynamic optimal learning parameter that provides good reconstruction accuracy for all signals. Out experimental results show that the proposed optimal learning parameter outperforms all fixed values of learning parameter in natural video sequences reconstruction.