PUP-Net: A Twofold Physical Model Embedded 3-D U-Net With Polarization Fusion for Solving Inverse Scattering Problems With a Sparse Planar Array

• Published in: IEEE transactions on microwave theory and techniques Vol. 73; no. 4; pp. 2123 - 2136
• Main Authors: Wang, Miao, Sun, Shilong, Zhang, Yongsheng, Dai, Dahai, Wu, Hao, Su, Yi
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.04.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: 3-D inverse scattering; Accuracy; Algorithms; Arrays; Artificial neural networks; coherence factor back-projection algorithm (CF-BPA); Computer vision; convolutional neural network (CNN); Convolutional neural networks; cross-correlated subspace-based optimization method (CC-SOM); Dielectrics; Geometry; Half spaces; Image reconstruction; Inverse problems; Inverse scattering; Iterative methods; Machine learning; Neural networks; Polarization
• ISSN: 0018-9480; 1557-9670
• DOI: 10.1109/TMTT.2024.3450684

In this article, a twofold physical model (PM) embedded 3-D U-Net with polarization fusion (PF), referred to as PUP-Net, is proposed for solving 3-D inverse scattering problems (ISPs) with a sparse planar array. To mitigate the high nonlinearity of ISPs, the twofold PM, consisting of the coherence factor back-projection algorithm (CF-BPA) and cross-correlated subspace-based optimization method (CC-SOM), first recovers the preliminary quantitative co-polarization images with high-fidelity geometry. The preliminary results are then input into the 3-D U-Net-based convolutional neural network (CNN), which outputs the fine contrast images of co-polarization (HH and VV). Finally, the upsampling PF achieves the co-polarization image fusion and generates higher accuracy and resolution results. PUP-Net follows the divide-and-conquer method: refine quantitative inversion based on good qualitative inversion. Synthetic and experimental inversion results demonstrate that PUP-Net achieves superior reconstruction accuracy and generalization ability compared with conventional iterative methods and current deep learning (DL) approaches in solving ISPs of half-space configurations.