Fast Sparsity-Based 3D Near-Field MIMO Imaging Using b-SAGA with a Batch Size Selection Strategy

Near-field imaging with multiple-input multiple-output (MIMO) arrays suffers from high computational cost. This hinders the practicality of conventional optimization techniques for large-scale sparsity-based imaging. In this paper, we address this issue and reduce the total runtime using a stochastic gradient strategy with variance reduction. Specifically, we utilize the Batch-Stochastic-Average-Gradient-Ameliore algorithm, which retains the convergence rate of the full proximal gradient method while reducing per-iteration cost. To our knowledge, this is the first time stochastic gradients with variance reduction have been applied to near-field MIMO radar imaging. In parallel, we provide a practical batch size selection strategy to improve runtime, which can also be used with other batch-based optimization methods for various radar image formation problems. We demonstrate using real-world measurements and total variation regularization that, compared to full and standard stochastic proximal gradient methods, the developed approach achieves speedups of over 5x and 23x, respectively, reducing runtime from the order of minutes to tens of seconds while maintaining high image quality.