High-quality and high-diversity conditionally generative ghost imaging based on denoising diffusion probabilistic...

Area of Science:

• Computational imaging
• Machine learning for optical systems

Background:

• Deep learning (DL) methods show promise for ghost imaging (GI) reconstructions.
• Current DL-GI methods often focus on pixel-level accuracy, neglecting reconstruction diversity.
• Existing approaches may not fully leverage conditional probability for improved GI.

Purpose of the Study:

• To develop a DL-based GI method that enhances reconstruction quality and diversity.
• To address limitations of current DL-GI techniques by incorporating probabilistic modeling.
• To validate the proposed method through simulations and physical experiments.

Main Methods:

• Utilized the denoising diffusion probabilistic model (DDPM) framework for GI.
• Developed a novel method named DDPMGI for image reconstruction.
• Evaluated performance at a low sampling rate (10%) and in color GI reconstruction.

Main Results:

• DDPMGI achieved superior image quality and reconstruction diversity compared to other methods.
• At 10% sampling rate, DDPMGI reached an average PSNR of 21.19 dB and SSIM of 0.64.
• Successful application in color GI reconstruction with average PSNR of 20.055 dB and SSIM of 0.723.

Conclusions:

• DDPMGI offers significant advancements in high-quality GI image reconstruction.
• The method demonstrates effectiveness in real-world scenarios and for color imaging.
• DDPMGI provides a powerful framework for diverse and accurate GI reconstructions.