Phase imaging with an untrained neural network.

Area of Science:

• Computational Imaging
• Optical Physics
• Deep Learning

Background:

• Supervised learning in computational imaging requires extensive labeled datasets.
• Acquiring ground-truth data for training is often impractical due to time and stability constraints.

Purpose of the Study:

• To develop a novel deep neural network approach that eliminates the need for large training datasets in computational imaging.
• To integrate physical models into neural networks for enhanced performance and reduced data dependency.

Main Methods:

• Incorporation of a complete physical model into a conventional deep neural network architecture.
• Development of a physics-enhanced deep neural network (PhysenNet).
• Demonstration using single-beam phase imaging.

Main Results:

• PhysenNet operates without prior training, eliminating the need for extensive labeled data.
• A single diffraction pattern is sufficient to reconstruct the phase object.
• Successful experimental demonstration of phase object reconstruction.

Conclusions:

• PhysenNet offers a paradigm shift in neural network design for computational imaging.
• The approach generalizes to various computational imaging problems by incorporating physical models.
• Reduces data acquisition requirements and computational costs.