Deep-learning-based multi-exposure image fusion for ptychographic imaging

Ptychography is a versatile coherent diffractive imaging (CDI) technique used in cell biology, materials science, and electronics. A persistent obstacle is the limited dynamic range of detectors, which degrades measurements under realistic, noisy conditions and thereby impairs phase retrieval. We address this limitation with an end-to-end deep learning framework that jointly performs multi-exposure image fusion (MEF) and phase retrieval. In contrast to earlier MEF pipelines that fuse exposures with task-agnostic heuristics and then run a separate reconstruction stage, our model learns an exposure-aware fusion inside the network and directly outputs amplitude and phase in real time. Using a real experimental dataset adapted for multi-exposure measurements, we benchmarked against a strong baseline and observed improved mean squared error (MSE) on both amplitude and phase; we further validated robustness under different exposure times. Across a wide range of exposure perturbations, the model maintains stable accuracy, demonstrating resilience to exposure-time variation and challenging experimental conditions.