Generalization Capabilities of Deep Learning Schemes in Solving Inverse Scattering Problems

Inverse scattering problem (ISP) is devoted to retrieving permittivities of dielectric scatterers from the knowledge of measured scattering data, which is highly nonlinear due to multiple scattering that exists in the exact scattering model. Iterative algorithms with regularization are usually used to solve such problems, but they are often time consuming. Recently, deep learning schemes have been proposed to solve full-wave ISP, in which contrasts are directly trained and then estimated with the convolutional neural network (CNN). It was found that even for the tests well out of the range of the training databases, these learning schemes are able to obtain satisfying results since the training of network is conducted on natural pixel bases. In this paper, we compare the generalization capabilities, i.e., the ability to reconstruct unseen data, for different deep learning schemes. Further, methods to improve the generalization capabilities have also been discussed and verified by numerical examples.