Binary amplitude-only image reconstruction through a MMF based on an AE-SNN combined deep learning model.

Area of Science:

• Computational imaging
• Optical physics
• Machine learning applications

Background:

• Imaging through multimode fibers (MMFs) is hindered by mode dispersion and coupling, causing scattered light and distorted images.
• Speckle patterns at the MMF output prevent direct image formation.

Purpose of the Study:

• To propose and validate a deep learning model for reconstructing images transmitted through MMFs.
• To address the limitations of traditional imaging techniques in MMF environments.

Main Methods:

• A combined deep learning model utilizing an autoencoder (AE) for feature extraction and image reconstruction.
• Integration of self-normalizing neural networks (SNNs) for high-order feature representation within the AE-SNN architecture.

Main Results:

• The AE-SNN model successfully reconstructed images from binary amplitude targets passed through a 5-meter MMF in both simulations and experiments.
• The model demonstrated robustness against system noise in simulations.
• Experimental validation confirmed the method's efficacy for MMF image reconstruction.

Conclusions:

• The proposed AE-SNN deep learning model effectively reconstructs images through MMFs, overcoming scattering and distortion issues.
• The model's spatial variability and self-normalizing properties allow for generalization to other computational imaging problems.