Single-Shot Recognition of 3D Phase Images With Deep Learning

Optical phase is a key information carrier in biomedical imaging and astronomical observation. However, it is often obscured by heterogeneous and scattering media, rendering optical phase imaging an utmost challenge. Limited by the memory effect or confinement, current methods have difficulties in retrieving phases at different depths. To address this challenge, a speckle three-dimensional reconstruction network (STRN) is developed to solve the inverse problem of the scattering process. STRN is featured by stacked convolutional filters to distinguish the depth of each image from the spatially overlapped speckles. From single-shot, reference-free, and scanning-free speckle patterns, STRN simultaneously extracts multiple phase objects at different depths with high fidelity. This new approach breaks the limitations of conventional approaches for phase reconstruction through scattering media and paves a novel avenue for optical communication and biological endoscopy applications.

Automated summary

Optical phase is crucial for biomedical imaging and astronomical observation, but it is often obscured by scattering media. Current methods are limited in retrieving phases at different depths. To address this challenge, a speckle three-dimensional reconstruction network (STRN) is developed to simultaneously extract multiple phase objects at different depths with high fidelity from speckle patterns.