Deep Unfolding Network for ISAR Imaging Based on Hypernetwork
Jianwen Guo, Hongyin Shi, Ting Yang, Liu Da, Zhijun Qiao
Abstract
Inverse synthetic aperture radar (ISAR) imaging can provide high-resolution images of targets. More recently, the unfolding algorithm has successfully realized fast and efficient ISAR imaging by providing a systematic connection between the traditional iterative algorithms widely used in ISAR imaging and data-based deep learning. However, once the unfolding framework is trained, the layer-dependent parameters are fixed and difficult to adapt to the variations in test scenarios, and usually, retraining is required. This article proposes an ISAR imaging framework based on a hypernetwork that can dynamically generate the unfolding network's internal parameters to accommodate the various scenarios. Specifically, the basis of the framework is an unfolding network based on the generalized expectation consistent (GEC) approximation phase recovery algorithm, where the damping factor is employed for data-driven learning. Instead of directly learning a set of optimal damping factors, the key is to develop a hypernetwork that trains an intelligent controller as the main unfolding network that can dynamically generate the damping factors according to testing scenarios. Thus, it exhibits strong robustness in various scenarios. Simulation and measurement experiments show that the proposed method exhibits excellent performance, robustness, and focusing accuracy.