Electromagnetic Imaging of Anisotropic Objects Using a Self-Attention Perceptual Generative Adversarial Network.

Area of Science:

• Electromagnetic imaging
• Applied physics
• Machine learning for geophysics

Background:

• Reconstructing high-resolution images of anisotropic targets in microwave imaging is challenging due to complex electromagnetic responses and nonlinear inverse scattering.
• Existing methods struggle with preserving structural details and capturing directional dependencies crucial for anisotropic materials.

Purpose of the Study:

• To develop a novel Perceptual Generative Adversarial Network (PGAN) integrated with a Self-Attention mechanism for enhanced anisotropic electromagnetic imaging.
• To improve the accuracy and quality of reconstructed images by refining permittivity estimates.

Main Methods:

• Proposed a Perceptual Generative Adversarial Network (PGAN) incorporating a Self-Attention module.
• Utilized perceptual loss for preserving high-level structural features.
• Trained the network to refine coarse permittivity estimates from Back-Propagation Schemes (BPSs).

Main Results:

• The PGAN with Self-Attention (SA) demonstrated superior performance compared to PGAN without SA and U-Net.
• Achieved a 15.1% reduction in Root Mean Square Error (RMSE).
• Improved the Structural Similarity Index Measure (SSIM) by 3.8%, indicating better recovery of fine-scale details.

Conclusions:

• The proposed PGAN with SA effectively enhances reconstruction accuracy and structural similarity in anisotropic electromagnetic imaging.
• The method shows significant potential for robust, high-resolution imaging in geophysical and remote sensing.
• The integration of Self-Attention is critical for capturing long-range dependencies in anisotropic targets.