A Long Short-Term Memory Approach to Incorporating Multifrequency Data Into Deep-Learning-Based Microwave Imaging

Motivated by the benefits of using multifrequency data in traditional nonlinear iterative optimization approaches in microwave imaging (MWI), this work compares three different approaches to using multifrequency data in deep-learning-based MWI. Specifically, we evaluate the imaging capabilities of the following: 1) a multichannel simultaneous frequency data-to-image U-Net-like network; 2) a novel cascaded multifrequency network; and 3) a novel long short-term memory (LSTM)-based recurrent network. The cascaded and LSTM networks are motivated by marching-on-frequency approaches and attempt to leverage reconstructions at lower frequencies as additional input information at higher frequencies. Results on both synthetic and experimental data show that the LSTM-based approach significantly outperforms the other models.