Phase retrieval based on the distributed conditional generative adversarial network | JoVE Visualize

Area of Science:

Computational imaging
Digital signal processing
Machine learning for image reconstruction

Background:

Phase retrieval aims to reconstruct original images from their Fourier intensity measurements.
Existing deep learning methods often struggle to simultaneously enhance both phase and amplitude reconstruction quality.
There is a need for advanced algorithms to improve the fidelity of reconstructed images in phase retrieval applications.

Purpose of the Study:

To develop a deep learning framework capable of simultaneously reconstructing high-quality phase and amplitude images.
To introduce a novel network architecture, the distributed amplitude and phase conditional generative adversarial network (D-APUCGAN), for improved phase retrieval.
To enhance the reconstruction quality by incorporating a content loss function and cascade strategies.

Main Methods:

Development of the distributed amplitude and phase conditional generative adversarial network (D-APUCGAN), incorporating UCGAN, AUCGAN/PUCGAN, and APUCGAN modules.
Implementation of a content loss function using Frobenius norm and total variation modulus to ensure similarity between reconstructed and source images.
Utilizing cascade strategies to progressively refine image reconstruction quality, particularly for phase information.

Main Results:

The D-APUCGAN method demonstrated superior performance in simultaneously reconstructing high-quality phase and amplitude images.
The content loss function significantly improved phase image quality compared to methods relying solely on amplitude information.
Experimental results on diverse datasets (natural, unnatural, DIV2K, MNIST, realistic data) showed significant effectiveness of the proposed cascade strategies.

Conclusions:

The proposed D-APUCGAN framework effectively addresses the limitations of existing methods in simultaneous phase and amplitude reconstruction.
The integration of content loss and cascade strategies offers a robust approach for enhancing image reconstruction fidelity in phase retrieval.
The method achieved notable improvements in evaluation metrics, with PSNR and SSIM values enhanced by at least 2.25 dB and 0.18, respectively, compared to conventional neural network methods.