Ghost imaging based on asymmetric learning.

Area of Science:

• Optics and Photonics
• Machine Learning
• Image Reconstruction

Background:

• Ghost imaging (GI) is an unconventional optical imaging technique utilizing correlated beams.
• Deep learning has advanced GI (GIDL) for higher quality image reconstruction, trainable with simulation data.
• Existing GIDL methods show limited performance with random noise due to unbalanced sensitivity to estimation errors.

Purpose of the Study:

• To address the performance limitations of GIDL with random noise.
• To propose an asymmetric learning framework for GIDL to overcome unbalanced error sensitivity.
• To enhance the robustness and accuracy of GIDL in noisy environments.

Main Methods:

• Development of an asymmetric learning framework for GIDL.
• Training GIDL models with simulation data.
• Experimental validation of the proposed asymmetric framework against traditional GIDL.

Main Results:

• The asymmetric learning framework significantly improves reconstructed image quality compared to symmetric loss functions.
• The structural similarity index of ghost imaging is quadrupled for randomly selected objects.
• The proposed method demonstrates enhanced robustness against random noise patterns.

Conclusions:

• An asymmetric learning framework effectively tackles the unbalanced error sensitivity in GIDL.
• This approach leads to superior image reconstruction performance in ghost imaging under noisy conditions.
• The findings pave the way for more practical and robust deep learning-based ghost imaging applications.