Improved Deep Learning-Based Microwave Inversion With Experimental Training Data

Deep learning-based inversion methods show great promise. The most common way to develop deep learning inversion techniques is to use synthetic (i.e., computationally-generated) data for training and initial testing. Later, the method can be used to image calibrated experimental data. However, it may be better to use experimental data in the training (not just testing) of these networks. In this paper, we (1) present a publicly available large-scale experimental dataset with 1638 measurements of 5 targets in a near-field imaging system that can be used for testing such deep learning inversion methods. A calibration MATLAB script is provided to assist users in processing and calibrating the dataset. (2) Using this dataset, we show that training a data-to-image deep learning-based inversion algorithm on either experimental data alone, or a mixture of experimental and synthetic data, leads to improved experimental imaging results for this data. The deep learning-based approaches are also compared against the gradient descent-based Multiplicative-Regularized Contrast Source Inversion Method.