Learning from simulation: An end-to-end deep-learning approach for computational ghost imaging.

Area of Science:

• Computational imaging
• Artificial intelligence
• Deep learning

Background:

• Deep learning (DL) models for computational imaging typically require extensive experimental data for training.
• Acquiring large labeled datasets can be time-consuming and resource-intensive.

Purpose of the Study:

• To demonstrate that a functional neural network for computational imaging can be trained using simulation data.
• To simplify the training process for DL-based imaging techniques.

Main Methods:

• Developed a one-step, end-to-end neural network.
• Trained the network using simulated data for computational ghost imaging (CGI).
• Reconstructed 2D images from 1D bucket signals without needing illumination patterns.

Main Results:

• Successfully trained a neural network using only simulation data.
• Achieved practical usability for computational imaging tasks.
• Demonstrated the method's effectiveness using computational ghost imaging as a case study.

Conclusions:

• Training neural networks with simulation data is a viable alternative to experimental data for computational imaging.
• This approach is particularly beneficial for imaging through scattering media.
• The concept is applicable to various DL-based computational imaging solvers.