A Novel Reconstruction Algorithm with High Performance for Compressed Ultrafast Imaging.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Imaging

Background:

• Compressed ultrafast photography (CUP) is a 2D imaging technique for observing ultrafast processes.
• Current reconstruction algorithms for CUP are slow due to reliance on image priors and complex parameter spaces, limiting practical applications.
• There is a need for efficient and high-quality reconstruction methods in CUP.

Purpose of the Study:

• To develop a novel, efficient, and high-quality reconstruction algorithm for compressed ultrafast photography (CUP).
• To accelerate the reconstruction process for CUP, enabling wider practical applications.
• To improve the imaging quality metrics (PSNR and SSIM) of CUP reconstructions.

Main Methods:

• Developed a forward model for CUP.
• Utilized the alternating direction multiplier method (ADMM) to derive three sub-optimization problems.
• Employed a neural network-based denoising algorithm (FFDNet) within the PnP-ADMM framework to solve a key sub-problem.

Main Results:

• The PnP-FFDNet algorithm achieved an average improvement of 3 dB in PSNR and 0.06 in SSIM on real CUP data compared to traditional algorithms.
• The proposed method reduced running time by approximately 96% compared to state-of-the-art algorithms on both benchmark and real CUP datasets.
• The algorithm demonstrated comparable visual results with significantly reduced computation time.

Conclusions:

• PnP-FFDNet offers a substantial improvement in both speed and quality for CUP image reconstruction.
• The algorithm effectively addresses the limitations of existing methods, making CUP more practical for ultrafast process observation.
• This work highlights the potential of integrating advanced deep learning denoising techniques with ADMM for efficient computational imaging.