Regularized Near-Field Holographic Microwave Imaging Using Phaseless Data and Convolutional Neural Network

This paper proposes a method to incorporate a two-dimensional convolutional neural network (CNN) to obtain prior knowledge about the azimuthal position of the imaged objects in a near-field holographic microwave imaging (NH-MWI) system when limited data is acquired. By using the phaseless responses of the objects within a circular scanned region and considering their azimuthal locations as targets, a training set is created for the designed CNN. The azimuthal location predicted by CNN is then used to solve the relevant systems of equations for NH-MWI in the spectral domain as regularized cost function minimization problems. This approach drastically reduces the amount of data required for a satisfactory imaging process, and a much smaller number of measurement frequencies and receiver antennas are needed. Furthermore, phaseless data eliminates the need for vector network analyzer measurements. This significantly reduces the cost, size, complexity, and measurement time in a practical system. The proposed approach is validated using simulated data contaminated with artificial noise to provide a realistic study.