Implicit Shape Biased Few-Shot Learning For 3D Object Generalization

Full waveform inversion (FWI) has been imple- mented using deep learning techniques as an analogue recurrent neural network for geophysics. However, the cycle-skipping issue, from which the conventional FWI suffers, troubles the deep- learning aided FWI as well if the least-square loss function is used to measure the misfit between observed and synthetic data. We propose to use a Wasserstein distance loss function from optimal transport for the inversion combined with a newly designed preprocessing transform, named integration affine scaling. This integration transform transfers the seismograms into probability densities, and significantly improves the results of optimal trans- port based methods. Numerical results show that the proposed method outperforms its counterparts in mitigating cycle-skipping, in comparison with other loss functions including the least-square loss, absolute loss, and the quadratic Wasserstein distance loss, with and without anisotropic total variational regularization.