Recurrent neural network reveals transparent objects through scattering media.

Area of Science:

• Optics and Photonics
• Machine Learning
• Image Reconstruction

Background:

• Scattering significantly degrades image quality in inverse problems, posing challenges for imaging through complex media.
• Existing methods using static neural networks have limitations in handling arbitrary scattering conditions.
• Dynamical approaches are needed to improve correlation and reveal phase objects obscured by scattering.

Purpose of the Study:

• To develop a novel dynamical machine learning approach for imaging phase objects through arbitrary diffusers.
• To enhance pattern correlation during training for improved phase object retrieval.
• To overcome the limitations of static neural networks in scattering media imaging.

Main Methods:

• Utilized on-axis diffuser rotation to introduce dynamics for training data generation.
• Employed multiple speckle measurements from different angles to create image sequences.
• Implemented recurrent neural networks (RNNs) to process dynamic speckle patterns and extract quantitative phase information.

Main Results:

• The RNN effectively filtered useful information and discarded redundancies from dynamic scattering patterns.
• Quantitative phase information was successfully retrieved even in the presence of strong scattering.
• Transparent images were reconstructed by analyzing speckle correlations within image sequences.

Conclusions:

• The proposed dynamical machine learning approach enables robust imaging of phase objects through arbitrary scattering media.
• RNNs effectively average out dynamic scattering effects, learning the static underlying pattern.
• This method shows potential for various imaging applications involving spatiotemporal dynamics.