Learning spectral initialization for phase retrieval via deep neural networks.

Area of Science:

• Computational Imaging
• Diffractive Optics
• Machine Learning Applications

Background:

• Phase retrieval (PR) is crucial in optics due to missing phase information in sensor measurements.
• Phase masks and coded diffraction patterns (CDPs) introduce redundancy to resolve PR ambiguities.
• Deep neural networks (DNNs) are increasingly used for inverse problems in computational imaging.

Purpose of the Study:

• To propose an end-to-end (E2E) deep network approach for phase retrieval.
• To jointly learn spectral initialization and network parameters for improved PR performance.
• To overcome limitations of traditional PR algorithms requiring extensive iterations.

Main Methods:

• Developed an E2E deep network incorporating an optical layer for propagation simulation.
• Included an initialization layer to approximate the optical field from CDPs.
• Employed a double-branch DNN to refine phase and amplitude recovery.

Main Results:

• The proposed E2E approach requires fewer snapshots compared to state-of-the-art methods.
• Fewer iterations are needed for accurate optical field estimation with the E2E network.
• Simulation results demonstrate superior performance in phase retrieval tasks.

Conclusions:

• The E2E deep network offers an efficient solution for phase retrieval problems.
• Joint learning of initialization and network parameters enhances PR accuracy and speed.
• This approach advances computational imaging techniques in diffractive optical systems.