Deblurring Ghost Imaging Reconstruction Based on Underwater Dataset Generated by Few-Shot Learning.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Deep learning enhances underwater ghost imaging by mitigating scattering effects.
• Acquiring large paired underwater datasets for training is a significant challenge.
• Existing methods like Cycle-GAN generate limited blurring variations in synthetic datasets.

Purpose of the Study:

• To develop a few-shot learning method for generating diverse paired underwater image datasets.
• To propose an improved underwater ghost imaging reconstruction method incorporating deblurring.
• To enhance the quality and clarity of reconstructed underwater images, particularly under low sampling rates.

Main Methods:

• A novel few-shot underwater image generative network was developed.
• The proposed method generates paired underwater datasets with improved blurring diversity compared to Cycle-GAN.
• A two-part reconstruction method combining image reconstruction and deblurring was implemented.

Main Results:

• The few-shot generative method produced superior paired underwater datasets, especially with limited real data.
• The proposed deblurring ghost imaging method demonstrated enhanced performance at low sampling rates.
• Experimental and simulation results confirmed increased clarity of underwater targets using the new method.

Conclusions:

• The few-shot generative network effectively addresses the challenge of limited paired underwater datasets.
• The proposed underwater deblurring ghost imaging method significantly improves reconstruction quality at low sampling rates.
• This advancement promotes wider applications of underwater ghost imaging technology.