Fast Microwave Near-Field Phase Retrieval Based on Improved Diffraction Neural Networks

This paper presents a fast microwave near-field phase retrieval method based on an Improved Diffraction Neural Network (IDNN). The method first modifies the Rayleigh-Sommerfeld (RS) formula using the angular spectrum method to adapt it for the Phase Retrieval (PR) equation. By enhancing the output layer of the DNN to simultaneously provide amplitude and phase information and incorporating a backward phase iteration step, the accuracy of phase retrieval is significantly improved. The optimization process leverages a custom-designed loss function and the Adam optimizer for gradient descent, enabling the retrieval of phase information using only amplitude data from two plane measurements. Simulation results validate the effectiveness of this approach and demonstrate its advantages in speed compared to classical algorithms such as the Gerchberg-Saxton (GS) algorithm and the Truncated Amplitude Flow (TAF) algorithm.