LFHRNet: Less-Forgetting High-Resolution Network for Through-Wall Radar Imaging

Through-wall radar (TWR) imaging is broadly applied in the detection of enclosed space, which plays an important role in security, rescue and military operations. However, the resolution of the existing TWR imaging method is not high enough to serve the practical application because of the physical constraints such as the limited antenna aperture. The deep learning method can improve the resolution due to its powerful reasoning capabilities, but it is challenging and costly to build comprehensive TWR datasets for training in practice. To address these challenges, this paper proposes a less-forgetting high-resolution network (LFHRNet) for through-wall radar imaging. Firstly, a cGAN network is trained on the source data as the pre-trained network to initialize the target network. The target network LFHRNet integrates two parallel branches by a router network. It forces a branch to focus on learning knowledge to solve target tasks, while another is frozen to alleviate the catastrophic forgetting. Finally, during the online phase, the low-resolution radar image in both source and target domains can be input into LFHRNet to get the high-resolution image. Simulation and practical experimental results show that LFHRNet can learn knowledge to reconstruct the shapes of the objects in target domain, while maintaining the source knowledge to reconstruct the shapes of the objects in source domain.