Color computational ghost imaging by deep learning based on simulation data training.

Area of Science:

• Optics
• Computational Imaging
• Machine Learning

Background:

• Traditional ghost imaging requires extensive data acquisition.
• Color imaging presents challenges in computational ghost imaging due to spectral information handling.

Purpose of the Study:

• To develop a color computational ghost imaging strategy that minimizes experimental workload and sampling times.
• To leverage simulated datasets for training neural networks in ghost imaging applications.

Main Methods:

• Experimental detection of color channel responsibilities in the imaging device.
• Simulation of extensive datasets based on experimental response values for neural network training.
• Application of a trained neural network for image reconstruction from low-sampling ghost data.

Main Results:

• Successful recovery of image information and correction of color distortion from blurry ghost images at a 4.1% sampling rate.
• Improved reconstruction quality, including details and color fidelity, with increased sampling rates.
• Validation of the method's feasibility and stability across color objects of varying complexity.

Conclusions:

• The proposed simulated-data-trained computational ghost imaging method offers an efficient and effective approach for color image reconstruction.
• This strategy significantly reduces the need for experimental data acquisition, making ghost imaging more accessible.
• The method demonstrates robust performance and generalization capabilities for diverse imaging scenarios.