Real-time coherent diffraction inversion using deep generative networks.

Area of Science:

• Computational imaging
• Deep learning for scientific applications
• Coherent diffraction imaging (CDI)

Background:

• Phase retrieval is crucial for reconstructing images from intensity data in applications like CDI.
• Iterative phase retrieval algorithms are slow, computationally intensive, and struggle with complex structures, hindering real-time imaging.
• Current methods face challenges in achieving rapid and accurate image reconstruction.

Purpose of the Study:

• To develop a deep learning-based method for fast and accurate phase retrieval in 2D imaging.
• To overcome the computational and convergence limitations of traditional iterative phase retrieval algorithms.
• To enable real-time image reconstruction from far-field diffraction intensities.

Main Methods:

• Training a pair of deep deconvolutional networks (CDI NN) to predict object structure and phase.
• Utilizing far-field diffraction intensities as input for the neural networks.
• Testing the performance of CDI NN on 2D objects.

Main Results:

• CDI NN successfully predicts 2D object structure and phase from diffraction intensities.
• The trained neural network reconstructs images within milliseconds on standard hardware.
• Achieved significantly faster computation compared to iterative phase retrieval methods.

Conclusions:

• CDI NN offers a rapid and efficient solution for phase retrieval in coherent diffraction imaging.
• The deep learning approach facilitates real-time imaging by overcoming computational bottlenecks.
• This method opens new possibilities for dynamic imaging applications previously limited by reconstruction speed.