An Electromagnetic Inversion Scheme Regularized by Deep Plug-and-Play Denoiser

A nonlinear electromagnetic inversion scheme incorporating a deep-learning-based plug-and-play (PNP) regularization approach is proposed. Unlike conventional regularization techniques that rely on fixed priors, the PNP method decouples the data fidelity and regularization terms, enabling the integration of a learned prior via a state-of-the-art denoiser. The recently introduced Swin-Conv-UNet (SCUNet), known for its superior image denoising capabilities by leveraging swin transformer blocks and residual convolutional blocks in its architecture, is employed as the pre-trained PNP denoiser within a multi-frequency Gauss-Newton nonlinear inversion framework. The resulting PNP-SCUNet inversion scheme is evaluated on both synthetic and experimental data, demonstrating superior performance compared to traditional inversion methods, including Tikhonov and total variation, and PNP denoisers using block-matching and 3D filtering (BM3D) and denoising convolutional neural network (DnCNN).