Phase-Retrieval with Incomplete Autocorrelations Using Deep Convolutional Autoencoders

TL;DR

This paper demonstrates that deep convolutional autoencoders can be trained to perform phase retrieval from incomplete autocorrelation data, offering a potential alternative to traditional iterative methods.

Contribution

The authors introduce a neural network-based approach for phase retrieval from incomplete autocorrelations, highlighting its advantages and limitations compared to existing techniques.

Findings

Neural networks can successfully recover signals from incomplete autocorrelation data.

The method shows potential advantages over iterative phase retrieval techniques.

Limitations include dependency on training data and incomplete information.

Abstract

Phase-retrieval techniques aim to recover the original signal from just the modulus of its Fourier transform, which is usually much easier to measure than its phase, but the standard iterative techniques tend to fail if only part of the modulus information is available. We show that a neural network can be trained to perform phase retrieval using only incomplete information, and we discuss advantages and limitations of this approach.