Noise robust Zernike phase retrieval via learning based algorithm only with 2-step phase shift measurements.

Area of Science:

• Optical Engineering
• Computational Optics
• Machine Learning

Background:

• Aberration analysis is crucial for optical system performance.
• Existing methods often require multiple measurements and are sensitive to noise.
• Developing efficient and robust aberration analysis techniques is an ongoing challenge.

Purpose of the Study:

• To develop a noise-robust, deep learning-based aberration analysis method using 2-step phase shift data.
• To create a realistic synthetic dataset for training deep neural networks.
• To outperform existing 4-step iterative methods in accuracy and speed.

Main Methods:

• Proposed a realistic aberration pattern generation method using statistical distribution of Zernike coefficients.
• Synthesized a large dataset (200,000 aberrations) for training.
• Trained deep neural networks (GoogLeNet, ResNet101) on the synthetic dataset.
• Evaluated the 2-step method against 4-step methods (derivative fitting, TIE, robust TIE).

Main Results:

• The 2-step deep learning method significantly outperformed 4-step methods in noise robustness, RMSE, and inference time for numerically generated aberrations.
• Experimental validation showed comparable RMSE to 4-step methods for real-world aberrations.
• Qualitative analysis revealed superior reconstruction of fringe patterns and phase distributions compared to existing methods.

Conclusions:

• The proposed 2-step deep learning method offers a practical and efficient approach for comprehensive aberration analysis.
• The method exhibits excellent noise robustness and computational speed.
• Potential for extension to higher-order aberrations and further performance improvements.