Self-Supervised Enhanced Imaging Network for Sparse Cross MIMO Arrays

In millimeter-wave (MMW) imaging, the focusing performance of sparse multiple-input-multiple-output (MIMO) arrays is susceptible to various noise interferences during the propagation of electromagnetic waves. In addition, the sparse array imaging based on the traditional compressive sensing (CS) algorithms suffers from low computational efficiency due to the numerous iterations, especially in 3-D imaging scenarios. To address these limitations, an interpretable self-supervised network with a parallel structure is proposed for fast enhanced imaging via sparse cross MIMO arrays. The designed interpolation-free MIMO operators are embedded into a generalized complex-valued framework unfolded by the alternating direction method of multipliers (ADMMs) to form the encoder of reconstruction branch, avoiding large-scale MIMO sensing matrix operations and allowing for the fast 3-D imaging processing of sparse MIMO data. The decoder is constructed to map the data from image domain to signal domain, thereby producing the pseudo-labels for self-supervised learning and eliminating the need for costly data annotation. The enhancement branch employs an optimization module based on the complex-valued total variation (CTV) and l₁ norms to guide the network training, thus improving the noise immunity and imaging performance of the network. The simulations and experiments show that the proposed network can obtain better-focused 3-D imaging results in a short time.