Implementation of Projection-Based Neural Networks on Landweber-Kaczmarz for Parallel Multifrequency Contrast-Source Electromagnetic Imaging

A multifrequency iterative neural projection framework, based on the recently proposed Contrast-Source Landweber-Kaczmarz (CSLK) scheme, is introduced for electro-magnetic imaging. Unlike traditional Frequency-Hopping (FH) techniques that require sequential initialization after each frequency update, this framework updates the medium parameters for all the frequencies in parallel during each CSLK iteration. The recovered medium parameters are then fed into projection Neural Networks (NNs), and the refined parameters generated by the NNs after each iteration are passed to the next CSLK iteration step. To achieve convergence and stability, different projection NNs are assigned after each iteration, with a loss function that combines the enhanced medium discrepancies from each NN. The numerical results section demonstrates the efficiency of the proposed approach.