Full-Waveform Inversion With Denoising Priors Based on Graph Space Sinkhorn Distance

Full-waveform inversion (FWI) is a critical geophysical imaging technique, renowned for its ability to generate high-precision subsurface structural models. However, FWI is a highly nonlinear and ill-posed inverse problem, necessitating appropriate regularization to incorporate prior information. Recently, the plug-and-play (PnP) method has shown promise in addressing various inverse problems by leveraging pretrained deep learning (DL)-based denoisers as priors, thus eliminating the need for explicit regularization functions. Building on this approach, we employ a pretrained DRUNet denoiser from the field of computer vision to implement regularization constraints for FWI via the PnP method. We propose a novel FWI framework (FWISD-DRU) that integrates graph space Sinkhorn distance (GS-SD) with the DRUNet denoiser to enhance FWI quality. By utilizing a bias-free denoising network architecture and additional noise-level map inputs, our approach improves the adaptability of the pretrained network to various geological images. Numerical tests on typical geological models validate the universality and effectiveness of our method. The experimental results demonstrate that our approach, incorporating the DRUNet denoiser, produces more accurate and higher resolution inversion results than total variation (TV) regularization and the denoising convolutional neural network (DnCNN) and FFDNet denoisers. This advantage is particularly pronounced under challenging conditions such as poor initial models, noisy observed data, and missing low-frequency components.