A deep-learning architecture, dubbed as the 2D-ADMM-Net (2D-ADN), is proposed in this article. It provides effective high-resolution 2D inverse synthetic aperture radar (ISAR) imaging under scenarios of low SNRs and incomplete data, by combining model-based sparse reconstruction and data-driven deep learning. Firstly, mapping from ISAR images to their corresponding echoes in the wavenumber domain is derived. Then, a 2D alternating direction method of multipliers (ADMM) is unrolled and generalized to a deep network, where all adjustable parameters in the reconstruction layers, nonlinear transform layers, and multiplier update layers are learned by an end-to-end training through back-propagation. Since the optimal parameters of each layer are learned separately, 2D-ADN exhibits more representation flexibility and preferable reconstruction performance than model-driven methods. Simultaneously, it is able to better facilitate ISAR imaging with limited training samples than data-driven methods owing to its simple structure and small number of adjustable parameters. Additionally, benefiting from the good performance of 2D-ADN, a random phase error estimation method is proposed, through which well-focused imaging can be acquired. It is demonstrated by experiments that although trained by only a few simulated images, the 2D-ADN shows good adaptability to measured data and favorable imaging results with a clear background can be obtained in a short time.
We propose a deep learning architecture, dubbed Plug-and-play 2D ADMM-Net (PAN), by combining model-driven deep networks and data-driven deep networks for effective high-resolution 2D inverse synthetic aperture radar (ISAR) imaging with various signal-to-noise ratios (SNR) and incomplete data scenarios. First, a sparse observation model of 2D ISAR imaging is established, and a 2D ADMM algorithm is presented. On this basis, using the plug and play (PnP) technique, PnP 2D ADMM is proposed, by combining the 2D ADMM algorithm and the deep denoising network DnCNN. Then, we unroll and generalize the PnP 2D ADMM to the PAN architecture, in which all adjustable parameters in the reconstruction layers, denoising layers, and multiplier update layers are learned by end-to-end training through back-propagation. Experimental results showed that the PAN with a single parameter set can achieve noise-robust ISAR imaging with superior reconstruction performance on incomplete simulated and measured data under different SNRs.
27 January 2025 in IEEE Transactions on Geoscience and Remote Sensing
Due to the constraints of observation conditions, it is difficult to obtain a large amount of measured data for real targets in the inverse synthetic aperture radar (ISAR) system. Existing deep networks usually use the simulated data of random points for training, which will lead to the degradation of the imaging performance of measured data when the distribution of measured data is different from that of simulated data, i.e., poor generalization performance. To address this problem, a high-resolution ISAR imaging method based on Swin transformer based deep reinforcement learning (SwinRL) is proposed. The two-dimensional alternating direction method of multipliers (2D ADMM) is modeled as a sequential decision problem in this method. The internal adjustable parameters are modeled as actions, and the Swin Transformer is used as the backbone network of the policy network and value network. The optimization of the actions, i.e., the internal adjustable parameters of the 2D ADMM algorithm, is then guided through network training in a reinforcement learning framework. After that, trained agent can automatically give optimal internal parameters according to different input data, and then well-focused imaging results can be obtained by executing 2D ADMM algorithm with optimal parameters. Finally, experimental results based on simulated and measured data show the performance priority of the proposed method compared to existing deep unrolling networks with fixed parameters.
26 March 2025 in IEEE Transactions on Aerospace and Electronic Systems
Deep learning has significantly advanced the field of sparse inverse synthetic aperture radar (ISAR) imaging. Nonetheless, current methodologies often face constraints due to their reliance on fixed loss rates and types, which can limit practical applicability and result in elevated computational demands both spatially and temporally, as well as diminished generalizability. In tackle to these challenges, this paper introduces a resilient ISAR imaging architecture grounded in hypernetwork technology capable of accommodating diverse loss rates. Initially, we devise a sparse signal reconstruction technique that functions irrespective of the loss rate or matrix dimensions. Subsequently, a methodology for producing arbitrary loss rates is outlined. Building upon this foundation, we present a hypernetwork framework that harnesses hypernetworks to derive defocus characteristics from range-Doppler (RD) imagery across varying loss rates, subsequently transforming these insights into ideal parameters for the sparse signal reconstruction process. By executing the iterative algorithm with these optimized parameters, we achieve loss rate robust ISAR imaging. Comprehensive evaluations conducted on point simulated data, electromagnetic simulated data and measured data confirm the superiority of our proposed approach in terms of reconstruction accuracy and generalization capabilities. This advancement facilitates reliable ISAR imaging across a spectrum of loss rates.
Determining the attitude of a non-cooperative target in space is an important frontier issue in the aerospace field, and has important application value in the fields of malfunctioning satellite state assessment and non-cooperative target detection in space. This paper proposes a non-cooperative target attitude estimation method based on the deep learning of ground and space access (GSA) scene radar images to solve this problem. In GSA scenes, the observed target satellite can be imaged not only by inverse synthetic-aperture radar (ISAR), but also by space-based optical satellites, with space-based optical images providing more accurate attitude estimates for the target. The spatial orientation of the intersection of the orbital planes of the target and observation satellites can be changed by fine tuning the orbit of the observation satellite. The intersection of the orbital planes is controlled to ensure that it is collinear with the position vector of the target satellite when it is accessible to the radar. Thus, a series of GSA scenes are generated. In these GSA scenes, the high-precision attitude values of the target satellite can be estimated from the space-based optical images obtained by the observation satellite. Thus, the corresponding relationship between a series of ISAR images and the attitude estimation of the target at this moment can be obtained. Because the target attitude can be accurately estimated from the GSA scenes obtained by a space-based optical telescope, these attitude estimation values can be used as training datasets of ISAR images, and deep learning training can be performed on ISAR images of GSA scenes. This paper proposes an instantaneous attitude estimation method based on a deep network, which can achieve robust attitude estimation under different signal-to-noise ratio conditions. First, ISAR observation and imaging models were created, and the theoretical projection relationship from the three-dimensional point cloud to the ISAR imaging plane was constructed based on the radar line of sight. Under the premise that the ISAR imaging plane was fixed, the ISAR imaging results, theoretical projection map, and target attitude were in a one-to-one correspondence, which meant that the mapping relationship could be learned using a deep network. Specifically, in order to suppress noise interference, a UNet++ network with strong feature extraction ability was used to learn the mapping relationship between the ISAR imaging results and the theoretical projection map to achieve ISAR image enhancement. The shifted window (swin) transformer was then used to learn the mapping relationship between the enhanced ISAR images and target attitude to achieve instantaneous attitude estimation. Finally, the effectiveness of the proposed method was verified using electromagnetic simulation data, and it was found that the average attitude estimation error of the proposed method was less than 1°.