Multifrequency Trained Projection Nonlinear Framework for Electromagnetic Imaging With Contrast-Source Landweber–Kaczmarz
Abdulla Ali Desmal, Jawad Alsaei · IEEE Geoscience and Remote Sensing Letters · 2024
A multi-frequency trained projection framework based on the recently developed contrast-source Landweber-Kaczmarz (CSLWKZ) is proposed for electromagnetic imaging. In contrast to conventional frequency-hopping (FH) schemes that necessitate sequential initialization at each frequency transition, the proposed framework executes all frequency updates concurrently within each CSLWKZ iteration. The retrieved medium parameters are input into projection neural networks (NNs). The enhanced medium parameters extracted from the NNs after each CSLWKZ iteration are dispatched to all the frequencies in the next iteration step. To ensure convergence and stability, different projection NNs are assigned after each CSLWKZ iteration step, and the loss function involved in the training process aggregates the discrepancies of the enhanced medium parameters produced by each projection NN. When compared to FH schemes, the numerical results confirm a high quality, an accelerated rate of convergence, and a wider range of validity in predicting medium parameters under both free-space and half-space background mediums. The results from the half-space background illustrate a 23.03% mean relative norm error and a mean absolute error of 0.0493 over the testing set. Furthermore, the proposed approach has been experimentally validated using an example from the Fresnel database.