Convolutional Neural Networks for Multifrequency Electromagnetic Inverse Problems
Hao Li, Lijia Chen, Jinghui Qiu
Abstract
In this letter, the multiple-channel scheme U-Net convolutional neural network (CNN) is introduced to solve the multifrequency electromagnetic inverse scattering problems. The U-Net CNN inversion method can achieve results with acceptable quality in a very short time, avoiding the drawbacks of the conventional iterative methods, such as ill conditions, heavy computational cost, time-consuming, etc. The training set is constructed by the multifrequency back propagation method. The inversion experiments based on synthetic and measured data show that the U-Net CNN inversion method has good performance in both single-and multifrequency cases. Compared with the single-frequency ones, the multifrequency U-Net CNN inversion results are more stable and accurate. This letter further shows that the multifrequency U-Net CNN work well in high contrast problems or more complex situations, and even can work in a different frequency band. It demonstrates that the multifrequency U-Net CNN suitable for solving actual inverse problems.