Single photon counting compressive imaging using a generative model optimized via sampling and transfer learning.

Area of Science:

• Optics and photonics
• Computer vision
• Machine learning

Background:

• Single photon counting compressive imaging (SPCCI) offers low cost and high sensitivity.
• Traditional compressed sensing (CS) reconstruction algorithms lead to long imaging times in SPCCI.
• Deep learning offers efficient reconstruction by avoiding iterative computations.

Purpose of the Study:

• To develop an efficient deep learning model for SPCCI reconstruction.
• To improve the speed and image quality of SPCCI.
• To jointly optimize the sampling and reconstruction process.

Main Methods:

• A novel compressed reconstruction network (OGTM) based on a generative model was proposed.
• A sampling sub-network was integrated for joint optimization of sampling and generation.
• Initial weights were transferred from an autoencoder to accelerate training and avoid slow convergence.

Main Results:

• The OGTM network demonstrated significantly improved convergence speed and imaging quality.
• Experiments on specific and generalized targets showed OGTM's ability to generate images from few measurements.
• Reconstruction quality surpassed existing CS recovery algorithms, addressing generative model limitations.

Conclusions:

• The proposed OGTM network enhances SPCCI by enabling rapid, high-quality image reconstruction.
• Joint optimization of sampling and generation through deep learning is effective for SPCCI.
• OGTM offers a promising solution for overcoming the time limitations of traditional SPCCI methods.