GAN with Self-Attention for Electromagnetic Imaging of Biaxial Objects

Electromagnetic imaging of anisotropic objects is investigated using a deep learning-based framework combined with a Self-Attention (SA) mechanism. In this work, anisotropic scatterers are excited using both TM (Transverse Magnetic)- and TE (Transverse Electric)-polarized electromagnetic waves, and the resulting scattering responses are measured to reconstruct the electromagnetic image of the target. A Back Propagation Scheme (BPS) is implemented to generate a first-pass reconstruction of the dielectric distribution. Unlike our previous work on uniaxial objects, the present study extends the BPS formulation to biaxial anisotropic media, enabling the independent reconstruction of the permittivity components along the x-, y-, and z-directions. The reconstructed coarse profiles are subsequently refined through a Generative Adversarial Network with Self-Attention (GAN-SA) operations. Unlike conventional GAN approaches, the proposed framework enhances feature correlation across different spatial regions, allowing the network to preserve object boundaries and dielectric variations more effectively. Simulation results with 5% and 20% Gaussian noise reveal that the proposed framework achieves better reconstruction quality, improved image clarity, and stronger resistance to noise interference. In addition, the integration of the SA module significantly improves the reconstruction of anisotropic dielectric properties while reducing reconstruction artifacts. The proposed GAN-SA model contains approximately 5.2 million trainable parameters, requires approximately 20 min for training, and achieves an inference time of less than 1 s for each reconstruction. These results demonstrate the effectiveness and computational efficiency of the proposed framework for biaxial anisotropic electromagnetic imaging.