A Mesh-Free 3-D Deep Learning Electromagnetic Inversion Method Based on Point Clouds

Deep learning has been successfully used in the 2-D-mesh-based electromagnetic inverse scattering (EMIS) problems in the past. However, for 3-D EMIS problems, the massive amount of meshes generated by the mesh-based inversion methods makes deep learning implementation very challenging. Addressing th...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on microwave theory and techniques Vol. 71; no. 8; pp. 3530 - 3539
Main Authors: Chen, Yanjin, Zhang, Hongrui, Cui, Tie Jun, Teixeira, Fernando L., Li, Lianlin
Format: Journal Article
Language:English
Published: New York IEEE 01.08.2023
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Subjects:
Artificial neural networks
Computing costs
Controllability
Deep learning
Electromagnetic interference
Electromagnetics
Forward scattering
Image reconstruction
Inverse problems
Inverse scattering
Machine learning
Mesh generation
mesh-free
Meshless methods
point cloud
Point cloud compression
Real-time systems
Three dimensional models
ISSN:0018-9480, 1557-9670
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Deep learning has been successfully used in the 2-D-mesh-based electromagnetic inverse scattering (EMIS) problems in the past. However, for 3-D EMIS problems, the massive amount of meshes generated by the mesh-based inversion methods makes deep learning implementation very challenging. Addressing this challenge will facilitate the further development of deep learning for 3-D EMIS problems. Inspired by the idea of point cloud, we propose a novel 3-D deep learning inversion methodology in this article. Different from previous work, the introduction of point clouds obviates the need to mesh the domain of interest (DOI) in this method, so we call it a mesh-free inversion method. The major advantage of this method compared with the mesh-based inversion methods is the low computational cost. In addition, the proposed method has other advantages. First, differently from prior deep learning inversion methods, the use of a reversible deep neural network (DNN) enables the proposed method to tackle both EMIS and electromagnetic forward scattering (EMFS) computations. Second, the inversion process becomes more controllable in this method, i.e., we can reconstruct merely a portion of the scatterer by simply shrinking the DOI. Third, the number of points used for the inversion can also be set as desired, without depending on the computational resources. Numerical and measured experiments demonstrate the feasibility and efficiency of the proposed method. This work opens a new vista on the application of deep learning toward 3-D EMIS problems. The code will be available at https://github.com/PKU-EMSensingLab/A-mesh-free-inversion-method-based-on-point-clouds .
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0018-9480
1557-9670
DOI:10.1109/TMTT.2023.3248174