Deep Complex Convolutional Neural Networks for Subwavelength Microstructure Imaging

To take the advantages of a convolutional neural network (CNN), U-net, and a complex-valued CNN (complex-CNN), a new complex-valued U-net (CU-net) is proposed for deep learning (DL)-based methods to solve inverse scattering problem (ISP). With the proposed CU-net, the complex scattered data carrying rich information of object can be directly used for inversion without any preprocessing, which is very helpful for the accuracy improvement of the final result. To validate the performance of proposed method, a microstructure, consisting of a finite periodic set of circular cylindrical dielectric rods, is considered and detected for textural abnormalities, which contains the missing, flaw, and displacement of the rods. The distances between rods and diameters of rods are both subwavelength, well beyond the Rayleigh criterion, which causes this ISP extremely ill-posed. For comparison, both the conventional iterative method and DL-based method are used to solve this nonlinear problem. Numerical simulations demonstrate that the well-trained DL-based methods can successfully produce excellent results almost in real time and can greatly outperform the conventional iterative methods in terms of quality and efficiency.

Automated summary

A new complex-valued U-net (CU-net) is proposed for solving inverse scattering problem, which directly uses complex scattered data for inversion without preprocessing, leading to improved accuracy and efficiency of the final result.