Robust spatial phase prediction from paired intensities using multi-scale wavelets and aberration sensing network.

Area of Science:

• Optical Engineering
• Computational Imaging
• Machine Learning

Background:

• Spatial phase prediction from intensity images is challenging due to compressed phase information and system variability.
• Existing methods struggle with robustness, light intensity limits, and adaptability to different phase ranges.

Purpose of the Study:

• To develop a robust and efficient method for spatial phase prediction from intensity images.
• To create a feature map that accurately represents aberration characteristics.
• To overcome limitations of existing phase prediction techniques.

Main Methods:

• Utilized Fourier transform and multi-scale wavelet transform for feature mapping, creating an aberration feature sequence (AFS).
• Developed a lightweight Aberration Sensing Convolutional Neural Network (ASCNN) for supervised learning and efficient computation.
• Evaluated the method on scalar diffraction and polarized SLM systems.

Main Results:

• The proposed AFS improved performance by approximately 30% compared to baseline features.
• ASCNN achieved high accuracy in aberration retrieval, with RMSE better than 0.0143λ and SSIM exceeding 96%.
• Demonstrated robustness and broad applicability across different optical systems.

Conclusions:

• The ASCNN model, combined with AFS, offers a significant advancement in spatial phase prediction.
• The method provides accurate and robust phase retrieval, overcoming previous limitations.
• The approach shows strong potential for diverse applications in optical metrology and imaging.