Learning-based lens wavefront aberration recovery.

Area of Science:

• Optics and Optical Engineering
• Computational Imaging
• Machine Learning in Optics

Background:

• Wavefront aberration deviates light path, impacting imaging quality.
• Accurate measurement is vital for adaptive optics, microscopy, and ophthalmology.
• Existing wavefront sensors can be costly and have limited resolution.

Purpose of the Study:

• Introduce LWNet, a practical, learning-based method for wavefront aberration recovery.
• Enable wavefront aberration estimation from single intensity measurements.
• Provide an alternative to traditional, complex wavefront sensing techniques.

Main Methods:

• LWNet employs a two-stage deep learning network.
• Input is a measured point spread function (PSF).
• Supervised and self-supervised learning stages refine aberration estimation using Zernike decomposition.

Main Results:

• LWNet successfully recovers wavefront aberration from simulated and real imaging systems.
• The method demonstrates superior performance compared to prior learning-based approaches.
• A synthetic dataset generated via ray tracing supported supervised learning.

Conclusions:

• LWNet offers a lightweight and effective solution for wavefront aberration estimation.
• The method shows robustness even when trained on simulated data.
• LWNet advances practical applications in optical engineering and imaging.