Unrolled primal-dual networks for lensless cameras

Area of Science:

• Computational imaging
• Image reconstruction
• Optical system modeling

Background:

• Conventional lensless imaging models use fixed point-spread functions, failing to capture optical aberrations and depth variations.
• This limitation hinders accurate simulation and reconstruction of lensless camera data.
• Existing methods often require significant network capacity for effective reconstruction.

Purpose of the Study:

• To develop a supervised primal-dual reconstruction method for lensless imaging that improves image quality.
• To demonstrate that this method can achieve state-of-the-art results without large network requirements.
• To enhance lensless image reconstruction by incorporating learnable forward and adjoint models.

Main Methods:

• Implemented a supervised primal-dual reconstruction algorithm.
• Embedded learnable forward and adjoint models within the reconstruction framework.
• Evaluated reconstruction performance using Peak Signal-to-Noise Ratio (PSNR) metrics.

Main Results:

• The proposed method achieved image quality comparable to state-of-the-art techniques.
• Reconstruction quality improved by +5dB PSNR compared to methods using fixed point-spread functions.
• The approach demonstrated effectiveness without requiring extensive network capacity.

Conclusions:

• Learned supervised primal-dual reconstruction offers a robust solution for lensless imaging challenges.
• Learnable forward and adjoint models significantly enhance the fidelity of lensless image reconstruction.
• This method provides a more accurate and efficient approach to lensless camera imaging.