Deep Learning For RADAR Signal Processing
E. T. Reehorst,Dr. Philip Schniter,Joe Castrilla,Michael Wharton,Logan McCarthy,Andrew Guazzerotti,Joe Wachob,Cole Holubeck,Brian Dillon,Megan Caroll,Dr. Emre Ertin,A. Thesis
Abstract:We address the current approaches to radar signal processing, which model radar signals with several assumptions (e.g., sparse or synchronized signals) that limit their performance and use in practical applications. We propose deep learning approaches to radar signal processing which do not make such assumptions. We present well-designed deep networks, detailed training procedures, and numerical results which show our deep networks outperform current approaches. In the first part of this thesis, we consider synthetic aperture radar (SAR) image recovery and classification from sub-Nyquist samples, i.e., compressive SAR. Our approach is to first apply back-projection and then use a deep convolutional neural network (CNN) to de-alias the result. Importantly, our CNN is trained to be agnostic to the subsampling pattern. Relative to the basis pursuit (i.e., sparsity-based) approach to compressive SAR recovery, our CNN-based approach is faster and more accurate, in terms of both image recovery MSE and downstream classification accuracy, on the MSTAR dataset. In the second part of this thesis, we consider the problem of classifying multiple overlapping phase-modulated radar waveforms given raw signal data. To do this, we design a complex-valued residual deep neural network and apply data augmentations during training to make our network robust to time synchronization, pulse width, and
Computer Science,Physics,Engineering