Characteristic space inversion method and device based on supervised descent method

The invention relates to the technical field of electromagnetic inversion imaging, in particular to a feature space inversion method and device based on a supervised descent method.The method comprises the steps that in the offline training stage, a training image containing the potential spatial position and electromagnetic characteristics of a target is generated according to prior information, a variational auto-encoder is trained through the training image, and the target potential spatial position and the electromagnetic characteristics are obtained; obtaining a variational auto-encoder after training is completed; inputting the training image into an encoder of the trained variational auto-encoder to obtain a low-dimensional feature implicit code, the low-dimensional feature implicit code and the corresponding measurement data being used for training a supervised descent method as a model parameter to be inverted; and in an online prediction stage, inputting measurement data of an unknown target into th