Image-to-Image Autoencoders for Quantifying Confidence in Machine Learning Solutions to the Inverse Scattering Problem

An image-to-image autoencoder is used to assess, in part, the confidence of a deep-learning data-to-image architecture for solving the inverse scattering problem (ISP) for microwave imaging. The output image of the data-to-image model is passed through the autoencoder to assess if the target is in range of the common dataset used to train the image-to-image autoencoder and data-to-image model. The quality of the autoencoder reconstruction is assessed using the structural similarity between the input and output of the autoencoder. Results demonstrate that the image-to-image autoencoder may detect certain out-of-range targets, and serves as a partial tool for quantifying confidence in imaging results. Future work will seek to build data-to-image models from parts of pre-trained image-to-image autoencoders.