Robust contrast-transfer-function phase retrieval via flexible deep learning networks.

Area of Science:

• Computational imaging
• Machine learning applications
• Diffractive imaging techniques

Background:

• Phase retrieval from coherent diffraction images is crucial for high-resolution imaging.
• Traditional methods like CTF-TV struggle with noise, limiting their application.
• Compressive sensing optimization is the basis for CTF-TV, but is sensitive to noise.

Purpose of the Study:

• To develop a robust and efficient phase retrieval algorithm for noisy coherent diffraction images.
• To improve the performance of contrast transfer function (CTF)-based phase retrieval using machine learning.
• To introduce a novel deep learning approach for phase map recovery.

Main Methods:

• Developed the CTF-Deep phase retrieval algorithm, integrating convolutional neural networks (CNNs) for regularization.
• Applied CNNs to regularize CTF-based phase retrieval problems, addressing limitations of traditional optimization.
• Utilized the sparsity of the investigated object within the CNN framework.

Main Results:

• The CTF-Deep algorithm demonstrates robustness against significant noise levels.
• Simulations and experimental results confirm improved phase map recovery.
• The method achieves high-resolution imaging capabilities, suitable for optical, x-ray, and terahertz applications.

Conclusions:

• Convolutional neural networks effectively regularize CTF-based phase retrieval, overcoming noise limitations.
• CTF-Deep offers a powerful and fast solution for phase retrieval in challenging imaging conditions.
• The proposed method significantly advances the applicability of phase retrieval in various scientific imaging domains.