Sub-Nyquist computational ghost imaging with deep learning.

Area of Science:

• Computational imaging
• Deep learning
• Image reconstruction

Background:

• Computational ghost imaging (CGI) traditionally requires numerous illumination patterns and one-dimensional light intensity sequences (LIS) for image reconstruction.
• Existing CGI methods, such as second-order-correlation and compressive-sensing CGI, face limitations in efficiency and data requirements.

Purpose of the Study:

• To propose a novel deep learning computational ghost imaging (CGI) scheme for sub-Nyquist and high-quality image reconstruction.
• To develop a deep neural network (DAttNet) capable of restoring target images using only a one-dimensional light intensity sequence (LIS).

Main Methods:

• A deep neural network, termed DAttNet, was designed and trained using simulation data.
• The DAttNet model was then utilized to retrieve target images from experimental data.
• The proposed deep learning CGI scheme was evaluated under sub-Nyquist sampling conditions.

Main Results:

• The DAttNet-based CGI scheme successfully reconstructed high-quality images at a sub-Nyquist sampling ratio.
• Experimental results demonstrated superior performance compared to conventional and compressive-sensing CGI methods under sub-Nyquist conditions (e.g., 5.45%).
• The proposed method achieved high-quality image reconstruction with significantly reduced data requirements.

Conclusions:

• The proposed deep learning CGI scheme offers a more efficient and effective approach to image reconstruction.
• This method shows significant potential for practical applications in challenging imaging scenarios like underwater, real-time, and dynamic CGI.
• The DAttNet model enables high-quality imaging with sub-Nyquist sampling, overcoming limitations of traditional CGI techniques.