Experimental Deep Learning Assisted Super-Resolution Radar Imaging

In recent radar applications, the demand for higher resolution imaging has been increased. Particularly, collision avoidance radars must recognize any objects of any size to prevent accidents and save lives. To address it, we design deep neural networks assisted MUSIC-based directional of arrival (DOA) estimators to enhance the resolution by increasing SNR of the radar array response. In fact, the SNR improvement is achieved by learning the noise structure using residual learning. To that end, we propose a modified residual deep network, which has higher performance than the conventional one. They are compared by the simulations in terms of eigenvalue distributions, SNR improvement, and the DOA resolution. Ultimately, in the experiment, we demonstrate the integration of the deep learning denoiser with the cumulant-based MUSIC giving more 3D details about the human shape as the radar target compared to the conventional MUSIC method.