Transfer Learning-Based Method for Scattered-Field Data Enhancement

To enhance the performance of deep learning-based scattered field data augmentation under limited training samples, this article proposes a scattering field data enhancement model based on a teacher-student knowledge distillation framework. The methodology comprises three key phases: First, a computational method is employed to generate scattering field data containing rich shape features, which serves as the pre-training dataset for a teacher diffusion network. Subsequently, quasi-experimental simulation data generated by full-wave simulation software is utilized as the training set for a student U-Net CNN network. Finally, feature transfer between teacher and student networks is implemented, enabling the trained network to extract features not explicitly contained in the student training set. Experimental results demonstrate the superior performance of the model in imaging complex scatterer structures using quasi-experimental data.