Deep Learning-Based Zero-Order Elimination for Enhanced Spatial Resolution in Single-Shot Imaging

Ultrafast events are common across various fields but often nonrepeatable, limiting traditional pump-probe methods. Single-shot ultrafast imaging techniques, particularly FRAME, have been developed to capture these events. FRAME uses four Ronchi gratings to modulate sub-pulses and separate them spatially. However, its spatial resolution is constrained by crosstalk from the zero-order term. This work proposes a deep learning approach using pix2pixGAN with a spatial attention mechanism to eliminate the zero-order term, enhancing FRAME’s spatial resolution. A simulated dataset of spatiotemporal multiplexed image pairs is used for training. Additionally, the K-means algorithm is introduced for adaptive filtering after zero-order term removal, improving reconstructed image quality and reducing filter coordinate selection errors.