Holographic optical field recovery using a regularized untrained deep decoder network.

Area of Science:

• Computational Imaging
• Optics
• Deep Learning

Background:

• Image reconstruction from limited data is challenging, especially in in-line holography.
• Existing methods like compressive sensing (CS) and supervised deep neural networks (DNNs) have limitations.
• CS methods struggle with insufficient image prior information, while DNNs require extensive training data.

Purpose of the Study:

• To develop a single-shot lensless in-line holographic reconstruction method.
• To overcome limitations of traditional CS and supervised DNN approaches.
• To enable accurate phase and amplitude recovery without large datasets.

Main Methods:

• Utilizing an untrained deep neural network integrated with a physical image formation model.
• Modifying a deep decoder network with regularization techniques.
• Employing a minimization process constrained by a deep image prior for Gabor hologram reconstruction.

Main Results:

• Accurate reconstruction of phase and amplitude images from Gabor holograms.
• Demonstrated effectiveness of the untrained neural network approach.
• Elimination of the need for training datasets or specific object/measurement assumptions.

Conclusions:

• The proposed method offers a robust solution for holographic reconstruction with minimal measurements.
• Untrained deep neural networks provide a powerful alternative for computational imaging tasks.
• This approach advances lensless in-line holography by removing data and assumption constraints.