Space-time reconstruction for lensless imaging using implicit neural representations.

Area of Science:

• Computational Imaging
• Computer Vision
• Machine Learning

Background:

• Computational imaging inverse problems face challenges from noise and model imperfections, degrading reconstruction quality.
• Existing space-time algorithms leverage temporal redundancy for improved denoising and reconstruction but demand substantial computational resources.
• Developing effective and adaptable temporal priors remains a significant hurdle in space-time imaging.

Purpose of the Study:

• To propose an implicit neural representation as a flexible and computationally tractable space-time prior for video reconstruction.
• To enhance reconstruction quality and robustness to noise in sequential imaging data.
• To address the limitations of traditional frame-by-frame reconstruction methods.

Main Methods:

• Utilized an implicit neural representation to model temporal dynamics.
• Implemented the implicit neural representation as a space-time prior.
• Applied the method to video data acquired using a lensless imager (DiffuserCam).

Main Results:

• Achieved improved reconstruction results compared to frame-by-frame methods.
• Demonstrated enhanced robustness to noise in reconstructed video.
• Validated the computational tractability and flexibility of the proposed implicit neural representation.

Conclusions:

• Implicit neural representations offer a powerful and efficient approach for space-time priors in computational imaging.
• The proposed method significantly improves video reconstruction quality and noise resilience.
• This technique provides a flexible and computationally feasible solution for challenging imaging inverse problems.