PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.
We report diffractive optical networks designed through a task-specific training process to classify and reconstruct spatially-overlapping phase images. Trained with ~550-million unique combinations of spatially-overlapping phase-encoded handwritten digits (MNIST), our blind testing achieves >85.8% accuracy for all-optical, simultaneous classification of two overlapping phase images of new/unseen handwritten digits. We also demonstrate the reconstruction of these phase images based on a shallow electronic neural network that uses as its input the highly-compressed optical signals synthesized by the diffractive network with ~20-65 times less number of pixels. This framework might find applications in computational imaging, on-chip microscopy and quantitative phase imaging fields.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.