Quantitative phase imaging based on model transfer learning.

Area of Science:

• Optical Information Processing
• Computational Imaging
• Machine Learning Applications

Background:

• Convolutional neural networks (CNNs) are vital for optical information processing.
• CNN generalization relies heavily on large, diverse datasets, which are difficult to acquire and annotate.
• Existing data-driven methods face challenges due to data scale and representational limitations.

Purpose of the Study:

• To develop a model transfer-based quantitative phase imaging (QPI) method.
• To address the limitations of dataset acquisition and annotation in data-driven QPI.
• To enhance the generalization ability of CNNs in optical information processing.

Main Methods:

• Utilized transfer learning by fine-tuning pre-trained base models for QPI.
• Developed a feature fusion method based on moment reconstruction for training dataset generation.
• Introduced a feature distribution distance scoring (FDDS) rule to assess dataset rationality.

Main Results:

• The proposed method enables CNNs with good generalization ability for QPI.
• Generated rich, accurately annotated datasets covering diverse scenarios.
• Demonstrated suitability for various sample types, achieving fast and high-accuracy phase imaging.

Conclusions:

• The model transfer-based QPI method effectively overcomes data acquisition and annotation challenges.
• The feature fusion technique provides a robust solution for creating representative training datasets.
• This approach significantly reduces data-related pressures and improves generalization in data-driven QPI.