Fig. 5

The trade-off between the complex field imaging performance and the output diffraction efficiency of diffractive complex field imagers. The PSNR on the y-axis reflects the mean value computed over the entire 5000 complex test objects derived from the EMNIST dataset. The data points with black borders correspond to the diffractive imager models trained exclusively using the structural fidelity loss function while disregarding diffraction efficiency, i.e., the ones shown in Supplementary Figs. S1a, S2a and Fig. 2a. The other data points originate from models trained with a diffraction efficiency-related penalty term, as defined in Eq. (12). These models were trained using varying target diffraction efficiency thresholds (\({\eta }_{{\rm{th}}}\)), specifically set at 0.1%, 0.2%, 0.4%, 0.8%, 1.6% and 3.2% corresponding to the data points from left to right on the plot, demonstrating the trade-off between the imaging performance and the output diffraction efficiency