High-definition image transmission through dynamically perturbed multimode fiber by a self-attention based neural...

Area of Science:

• Optical Engineering
• Artificial Intelligence
• Image Processing

Background:

• Multimode fiber (MMF) optic systems face challenges in transmitting high-quality images due to signal degradation.
• Existing artificial neural networks (ANNs) like convolutional neural networks (CNNs) have limitations in fidelity and robustness for MMF image transmission.

Purpose of the Study:

• To develop a more effective and robust method for faithful image transmission through multimode fibers.
• To leverage self-attention mechanisms for enhanced image quality and reduced computational complexity.

Main Methods:

• Implementation of a self-attention-based neural network for image transmission over MMF.
• Comparative analysis against real-valued ANNs (CNNs) using metrics like Enhancement Measure (EME) and Structural Similarity (SSIM).
• Utilizing a simulation dataset and hybrid training to improve network robustness against MMF bending and disturbances.

Main Results:

• The self-attention network achieved higher image quality, with EME and SSIM improvements of 0.79 and 0.04, respectively.
• A reduction of up to 25% in the total number of network parameters was observed.
• Hybrid training improved SSIM by 0.18 on datasets with varying MMF disturbances, demonstrating enhanced robustness.

Conclusions:

• Self-attention neural networks offer a superior approach for faithful MMF image transmission compared to traditional CNN-based methods.
• Hybrid training significantly boosts the robustness of image transmission systems against physical disturbances like fiber bending.
• The proposed method presents a promising, simpler, and more robust solution for high-demand imaging applications, including endoscopy.