A Sparse Aperture ISAR Imaging and Autofocusing Method Based on Meta-Learning Framework
Ruize Li, Shuanghui Zhang, Yongxiang Liu, Xiang Li
Abstract
The cross-range resolution of inverse synthetic aperture radar (ISAR) images is influenced by undersampled data under the sparse aperture (SA) condition. Recently, learning-based methods have been applied to SA-ISAR imaging and have achieved impressive performance. Learning-based methods can achieve satisfactory results by training on large datasets. However, these methods may fail to reconstruct high-quality images in practical applications due to training data limitations. In this article, we consider this problem within a meta-learning framework. In this framework, the SA-ISAR imaging network is trained by a learnable optimizer instead of a fixed stochastic gradient descent (SGD) optimizer. A fully connected network is designed as an optimizer for imaging network training; this network is also called a meta-learner. The whole training procedure in the proposed framework is divided into two parts. In the first part, a suitable meta-learner is trained. In the second part, the well-trained meta-learner is applied to train the ISAR imaging network. In this article, our previously proposed complex-valued alternating direction method of multipliers network (CV-ADMMN) is trained within this framework; this approach is called Meta-CV-ADMMN. The experimental results show that the proposed training framework can improve the imaging performance and data adaptability of CV-ADMMN, especially when the training data are limited.