Ghost imaging of blurred object based on deep-learning.

Area of Science:

• Computational imaging
• Deep learning for optics
• Ghost imaging techniques

Background:

• Traditional object reconstruction methods can be complex and require extensive data.
• Ghost imaging offers a unique approach using correlation measurements for imaging.
• Deep learning models show promise in enhancing image reconstruction quality.

Purpose of the Study:

• To introduce a novel neural network architecture for blurred object reconstruction.
• To leverage ghost imaging principles for efficient reconstruction using only bucket signals.
• To evaluate the network's generalization capabilities on unseen data.

Main Methods:

• A new neural network combining ResNetV2, residual dense block (RDB), and eHoloNet was developed.
• Object reconstruction was performed using ghost imaging theory, requiring only bucket signals.
• Training utilized the ENMNIST dataset with Airy convolution for blurring; multi-slit patterns served as testing sets.

Main Results:

• The proposed neural network demonstrated superior performance in reconstructing blurred objects.
• Both simulated and experimental results confirmed the network's effectiveness and generalization ability.
• The study also provided insights into the limitations of the developed reconstruction method.

Conclusions:

• The novel neural network effectively reconstructs blurred objects using ghost imaging and deep learning.
• The architecture exhibits strong generalization, outperforming existing methods on diverse test sets.
• This work advances computational imaging by offering an efficient and robust reconstruction solution.