High-Resolution Through-Wall Radar Imaging Based on Few Shot and Transfer Learning

Through-wall radar and its imaging play an important role in the detection of enclosed space. However, the resolution of the existing through-wall radar imaging method is not high enough to serve the practical application. The deep learning methods alleviate the problem but they suffer from the insufficient labeled data. To address these challenges, this paper proposes a high-resolution imaging method of through-wall radar based on few shot and transfer learning. Firstly, a cGAN network is trained on the source data and is used to initialize the parameters of the target cGAN as the pre-trained network. Secondly, the target cGAN is fine-tuned on target domain to obtain the transferred cGAN. Finally, during the online phase, the low-resolution radar image in target domain is input into the transferred generator to get the high-resolution optical image. Simulation results show that, when there is limited amount of target data, the quality of generated images can be improved in fewer iterations by using the knowledge of the pre-trained source network.