Deep Learning for Compressive SAR Imaging with Train-Test Discrepancy (Preprint)
Abstract
We consider the problem of compressive synthetic aperture radar (SAR) imaging with the goal of reconstructing SAR imagery in the presence of under sampled phase history. While this problem is typically considered in compressive sensing (CS) literature, we consider a variety of deep learning approaches where a deep neural network (DNN) is trained to form SAR imagery from limited data. At the cost of computationally intensive offline training, on-line test-time DNN-SAR has demonstrated orders of magnitude faster reconstruction than standard CS algorithms. A limitation of the DNN approach is that any change to the operating conditions necessitates a costly retraining procedure. In this work, we consider development of DNN methods that are robust to discrepancies between training and testing conditions. We examine several approaches to this problem, including using input-layer dropout, augmented data support indicators, and DNN-based robust approximate message passing.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 30, 2021
- Accession Number
- AD1138245
Entities
People
- Morgan Mccamey
Organizations
- Wright State University