A Marching-on-Frequency-Inspired Cascaded Network for Multi-frequency Microwave Imaging

A cascaded machine learning model designed to produce target images from multi-frequency microwave imaging data is presented. The architecture is inspired by a marching-on-frequency approach where lower frequency information is used to recover lower resolution images that are used as initial guesses or prior information at higher frequencies. The architecture consists of a sequence of U-net-like deep learning models, each taking single-frequency microwave measurement data and the reconstruction from the previous model in the sequence as input. For both synthetic and experimental imaging scenarios the cascaded network is shown to out-perform a data-to-image network where the same multi-frequency data is incorporated as a multi-channel input.