Ghost translation: an end-to-end ghost imaging approach based on the transformer network

Area of Science:

• Computational Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep neural networks (DNNs) are increasingly utilized in computational imaging to enhance signal-to-noise ratio.
• Image quality is often compromised by low sampling ratios or noisy environments.

Purpose of the Study:

• To propose a novel computational imaging scheme using a transformer network for signal translation.
• To demonstrate high-quality 2D image retrieval from low-sampling-ratio data using a single-pixel detector.

Main Methods:

• A transformer network employing a sequence transduction mechanism was developed.
• A simulation database was used to train the network for signal translation.
• The method was tested with both designed and random speckle patterns for illumination.

Main Results:

• High-quality 2D images were retrieved with 2% sampling ratio and no background noise.
• The proposed method demonstrated robustness against noise interference.
• The system successfully translated single-pixel detector signals into images in an end-to-end manner.

Conclusions:

• The developed translation mechanism offers a new approach for DNN-assisted ghost imaging.
• This method is applicable to diverse computational imaging scenarios.
• The technique significantly improves image reconstruction quality under challenging conditions.