SiSPRNet: end-to-end learning for single-shot phase retrieval.

Area of Science:

• Computational imaging
• Deep learning for image reconstruction

Background:

• Traditional phase retrieval methods are iterative and time-consuming.
• Recent deep learning methods offer real-time reconstruction but suffer from quality limitations due to domain discrepancies.

Purpose of the Study:

• To develop a novel deep neural network, SiSPRNet, for enhanced single-shot maskless phase retrieval.
• To improve the quality of reconstructed phase images by effectively utilizing spectral information and global correlations.

Main Methods:

• Designed SiSPRNet with a Multi-Layer Perceptron (MLP) for feature extraction and a self-attention mechanism in Up-sampling and Reconstruction (UR) blocks.
• Integrated UR blocks into a residual learning structure to address information flow and gradient issues.
• Evaluated performance on diverse datasets and a practical optical experimentation platform.

Main Results:

• SiSPRNet demonstrated superior performance over existing deep learning methods in single-shot maskless phase retrieval.
• The model effectively extracts representative features, reduces noise, and mitigates overfitting.
• Experimental validation confirmed the practical applicability and effectiveness of the proposed approach.

Conclusions:

• SiSPRNet offers a significant advancement in single-shot maskless phase retrieval, achieving high-quality reconstructions.
• The novel network architecture effectively leverages spectral information and global correlations for improved phase imaging.
• The developed method provides a robust solution for real-world phase retrieval applications.