Deep learning inversion with supervision: A rapid and cascaded imaging technique

•A machine learning algorithm, DLIS, which can effectively reduce the number of training samples, demands less computational resources and maintains equal reconstruction quality in solving inverse problems is proposed.•The proposed algorithm is fast in inversion speed, stable in performance and highly accurate, which is demonstrated by examples in corrosion mapping and makes it extremely promising in practical applications.•This research shows neural networks and descent direction matrix can be integrated together and have full potential in non-destructive evaluation (NDE), biomedical imaging and geophysical prospecting applications. Machine learning has been demonstrated to be extremely promising in solving inverse problems, but deep learning algorithms require enormous training samples to obtain reliable results. In this article, we propose a new solution, the deep learning inversion with supervision (DLIS) and applied it for corrosion mapping in guided wave tomography. The inversion results show that when dealing with multiple defects of complex shape on a plate-like structure, DLIS methods can reduce the scale of training set effectively compared with other deep learning algorithms in experiment because a good starting model is provided and the nonlinearity between the global minimum and observed wave field is greatly reduced. In terms of reconstruction accuracy using experimental data, the thickness maps produced by DLIS are reliable with high accuracy. With few modifications, this method can be conveniently extended to 3D cases. These results imply that DLIS is one of the promising methods to be applied in fields with similar physics like non-destructive evaluation (NDE), biomedical imaging and geophysical prospecting.