Noise-robust ptychography using unsupervised neural network

Abstract

Ptychography is a lensless phase retrieval technique that offers a large field of view and high resolution, making it highly suitable for bioimaging, semiconductor inspection, and material science. However, noise can lead to reconstruction failures. To address this, we propose an algorithm that leverages an untrained, physics-based neural network for ptychographic reconstruction, which effectively mitigates noise without priors or compromising computational efficiency. This algorithm is validated through both simulations and experiments. Our approach establishes a new optimization method for ptychography under severe noise environments such as low-dose imaging and wide-spectrum illumination conditions, and it can be adapted to other parameter optimization by adjusting the physical model and network structure.