Genetic algorithm based framework for optimized sensing matrix design in compressed sensing

doi:10.1007/s11042-022-12894-0

2022-04-27

PRE

Irfan Ahmed *

Aftab Khan

DOI：10.1007/s11042-022-12894-0

原文链接

原文求助

摘要

En 中文

Sampling matrices used in compressed sensing framework are mostly randomly structured and thus inefficient in terms of memory utilization, reconstruction speed, and computational resources utilization. Recent use of learning based optimised sensing matrices are proven to be energy efficient and computationally feasible. However, such matrices are designed in isolation with reconstruction algorithm and thus could not facilitate efficient signal recovery. In this paper, a unique approach is devised to develop the optimized sampling matrix by using offline training mechanism of sparse matrix through Genetic Algorithm, Learning Based Compressive Subsampling, and combination of both strategies. The core idea of this paper lies in adopting GA for optimal sensing matrix design by using evolutionary mechanism. The designed matrices are evolved in number of iterations with respect to the desired accuracy. These matrices are of deterministic nature and aimed to provide computationally feasible and memory efficient solution by targeting maximal features of the training samples. In this work, we adopted two phase process for design and evaluation of the proposed framework which are called train and test phases. In first stage, the cumulative feature vector of all speech training samples is calculated and provided to the proposed techniques. In training phase, Genetic algorithm is tailored to be used as an optimization strategy to develop sensing matrix which supports faithful signal reconstruction. In test phase, the performance of the developed sampling matrices is evaluated by comparing signal-to-noise ratio and root mean squared error values obtained by sampling and reconstruction of test set through widely used modern signal reconstruction algorithms. The accuracy and precision analysis of acquired signal-to-noise ratio and root mean squared error values show that the GA based methods outperform the latest sampling matrix design method, when signal reconstruction is carried out with convex relaxation algorithm, greedy method, and linear reconstruction technique.

Keyword:

Genetic algorithm

Sensing matrix design

Compressive sensing

Learning based compressive subsampling

GA-LBCS