Microwave Imaging Using Cascaded Convolutional Neural Networks

As an efficient and fast way to solve inverse electromagnetic problems, a deep learning algorithm for microwave imaging of complex objects is proposed. It solves the scattered wide-band time domain signals using a cascaded structure of convolutional and U-net neural networks. The algorithm is trained and tested using 2000 sets of data covering the band 0.5-2 GHz generated from an imaging domain that includes complex shaped targets irradiated by 16 antennas. Mean values of Intersection over Union (IoU) are near 0.7 for all the tested cases, while 1.0 represents a perfect overlap between the reconstructed image and ground truth, and a value above 0.5 shows a satisfying shape reconstruction (i.e., more than half of the reconstructed image overlaps with the ground truth). More than 80% of tested cases have less than 50% relative errors compared to the ground truth. These results show great potential for the developed algorithm in localization, shape reconstruction, and classification of complex objects in microwave imaging.

Automated summary

A deep learning algorithm is proposed for microwave imaging of complex objects, which provides an efficient and fast solution to inverse electromagnetic problems. The algorithm uses a cascaded structure of convolutional and U-net neural networks to solve the scattered wide-band time domain signals. With training and testing using 2000 sets of data covering the band 0.5-2 GHz and complex shaped targets irradiated by 16 antennas, the algorithm achieves mean Intersection over Union (IoU) values near 0.7 for all tested cases and more than 80% of tested cases have less than 50% relative errors compared to the ground truth. These results demonstrate the great potential of the developed algorithm in localization, shape reconstruction, and classification of complex objects in microwave imaging.