Wentao Li

Sparse array synthesis is always a tough issue because of its nonlinear, non-convex characteristic, especially the synthesis with geometry-position constraints is particularly challenging where current methods struggle to effectively handle it. To address these issues, a novel alternating direction method of multipliers (ADMM) with relaxation-projection (RP) is proposed for sparse array synthesis in this work. By decomposing the original complex optimization into two computationally tractable subproblems of auxiliary variable updating and element position updating, the ADMM-RP method significantly improves processing efficiency through specifically designed approaches for each subproblem. Moreover, to address the computationally challenging subproblems, a novel relaxation-projection mechanism is designed, effectively eliminating the infeasible solutions in advance. Besides, geometric constraints are temporarily relaxed in the relaxation stage to obtain an initial feasible solution, significantly reducing computational complexity, whereas element position updating criteria is specifically designed in the projection stage to ensure that all geometric constraints are strictly satisfied. Numerical results for rectangular, symmetric, hybrid, and large-scale arrays under tough constraints demonstrate the superior performance, high computational efficiency, and robustness of the proposed ADMM-RP method, which is highly capable of handling geometry-position-constrained sparse array synthesis.

Deep learning has demonstrated its potential capability in object detection of synthetic aperture radar (SAR) images. However, the detection of small objects still faces unique challenges due to insufficient detailed information and complex background clutter distribution. Although existing methods alleviate these issues to some extent by extracting discriminative features, the detection performance is still highly unsatisfactory due to the inevitable information loss as neural networks deepen. In this work, a multiaspect compensation network (MACNet) is proposed to improve the detection accuracy of small objects, in which two key components interactive encoding enhancement (IEE) module and multi-scale feature enhancement guided (MFEG) network are introduced. Specifically, IEE initially encodes the low-level details and highlevel semantics from neighboring features to enhance the feature representation of small objects. Besides, MFEG performs global modeling using Mamba to extract high-frequency information in SAR images, then integrates it with higher-level predicted features to compensate for information loss in multiscale features. Moreover, since the commonly used loss functions often exhibit insensitivity to small objects, a more effective scale and position sensitive (SPS) loss function is designed to improve attention to small objects. Numerical results on the SAR datasets LS-SSDD-V1.0, HRSID, and SAR-AIRcraft-1.0 demonstrate the effectiveness and superiority of the proposed strategy. The code is available at https://github.com/CongLi-18/MACNet.

A new N -way reconfigurable power divider with path switching and frequency tuning functions is proposed. The divider successfully realizes flexible path switching and frequency adjustment by using the innovative design of PIN diodes, reconfigurable parallel coupling lines and reconfigurable short-circuit branches. By dynamically adjusting the state of the PIN diode, the port connection and disconnection functions can be realized, and an advanced isolation network is introduced to ensure excellent port isolation performance during port switching. In addition, by embedding a varactor diode ( C _{v 2} ) between the coupling lines of the input port, the impedance matching during port switching is effectively guaranteed. Furthermore, by loading a reconfigurable short-circuit branch containing a varactor diode ( C _{v 1} ) on one side of the input port coupling line, the frequency adjustment function is realized. Finally, N transmission modes with 2 ^N − 1 operating states can be realized. For verification, a four-way reconfigurable power divider is designed and fabricated, and 15 states of four transmission modes of the power divider are realized, each state enabling frequency tuning capability. The results facilitate the potential application of the proposed power divider in wireless communication systems.

Transformer-based methods have demonstrated potential capability in object detection of synthetic aperture radar (SAR) images. However, their reliance on global features and quadratic complexity hinders real-time and precise localization of small objects. To deal with these issues, in this work, a Mamba in Transformer DETR (MIT-DETR) framework is proposed. In MIT-DETR, the features are initially divided into visual sentences and processed by Mamba to extract global information with linear complexity. These visual sentences are then further divided into words, which are analyzed for local information through the transformer. In addition, a multiscale dilated attention (MSDA) module is designed with the purpose of obtaining rich multiscale information at a low cost. Numerical experiments show that MIT-DETR improves AP₅₀ by 5.01% and FPS by 5.1 on HRSID compared to the baseline. These results demonstrate the effectiveness and superiority of the proposed strategy. The code will be available at https://github.com/CongLi-18/MIT-DETR.

Deep learning has demonstrated its potential capability in object detection of synthetic aperture radar (SAR) images. However, the low detection accuracy for small and dense objects remains a critical issue. To address this issue, in this work, a feature focus enhanced YOLO (FFE-YOLO) architecture is proposed. In FFE-YOLO, a channel feature enhanced (CFE) module is introduced to extract richer information and reduce time consumption by integrating it into the backbone. Additionally, a feature selection fusion network (FSFN) is designed to enhance the feature representation of small and dense objects by fully utilizing channel information. Numerical results demonstrate that FFE-YOLO outperforms baseline by 3.12% and 3.06% on datasets HRSID and LS-SSDD-v1.0, respectively, but with less inference time. These results demonstrate the effectiveness and superiority of the proposed strategy. The code will be available at https://github.com/CongLi-18/FFE-YOLO.

In this letter, an efficient machine learning framework tailored for multi-objective antenna design is proposed. A hybrid neural network is designed as the decoder for training, possessing the ability of automatically determining the optimal number of network layers. This adaptability vastly speeds up electromagnetic (EM) response generation for antennas, eliminating the need for intensive EM simulations. Subsequently, encoders, mapping networks, and pre-trained decoders are sequentially connected to train the encoders and mapping network, with the purpose of learning features associated with multiple metrics. To this end, when the multi-objective changes, our proposed framework has the merit of facilitating swift retraining of the neural network. Simulation results of a SRR loaded patch antenna and an 8-element multiple-input multiple-output (MIMO) antenna, both for Internet of Things (IoT) terminals, are provided to verify the proposed framework. The two antennas respectively achieve bandwidths of 6 GHz (21.20 GHz to 27.20 GHz) and 2.76 GHz (3.26 GHz to 6.02 GHz). Numerical results reveal the superiority and effectiveness of the proposed framework.

Multiple reconfigurable intelligent surfaces (RISs) aided cell-free (CF) massive multiple input multiple output (MIMO) system has demonstrated its potential advantage in improving capacity and energy efficiency (EE). In this work, by deriving the upper and lower bounds of the asymptotic received power for double-hop signals, single-hop/multi-hop multi-RIS aided CF MIMO systems are theoretically discussed. On the basis of which, a phase modulation array (PMA) based hybrid precoding (HP) is proposed for single-hop RIS-aided CF massive MIMO system, in which PMA based RIS are employed to reduce energy consumption. Accordingly, a joint hybrid precoding and multi-RIS optimization problem is formulated to maximize the sum rate with regard to three variables (digital precoding, analog precoding, and RIS phase), and an alternating optimization algorithm is provided. To further reduce the computational complexity, a convolutional neural network (CNN) based method is proposed to seek the three-variable solution. Finally, simulation results verify that both schemes show superior performance in terms of sum rate and EE while the CNN based method exhibits fast convergence.