Untrained networks for compressive lensless photography

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning

Background:

• Compressive lensless imagers offer compact solutions for imaging applications like microscopy and hyperspectral imaging.
• Traditional reconstruction methods rely on convex optimization or deep learning, which often require extensive labeled training data.
• Acquiring large datasets of ground truth measurements for training can be challenging or impossible in many scenarios.

Purpose of the Study:

• To develop an unsupervised deep learning approach for compressive image recovery that eliminates the need for labeled training data.
• To demonstrate the efficacy of this untrained network method across various lensless imaging modalities.
• To improve image reconstruction quality compared to existing methods in compressive lensless imaging.

Main Methods:

• An unsupervised learning framework utilizing untrained neural networks for compressive image recovery.
• The network weights are updated using the measurement data itself, bypassing the need for ground truth pairs.
• Demonstration on lensless compressive 2D imaging, single-shot high-speed video (rolling shutter), and single-shot hyperspectral imaging.

Main Results:

• The proposed untrained network approach successfully reconstructs images from 2D measurements in lensless compressive imaging systems.
• Improved image quality was observed across simulations and experimental verifications for all tested imaging modalities.
• The method shows superior performance compared to existing reconstruction techniques, particularly where labeled data is scarce.

Conclusions:

• Untrained neural networks offer a viable and effective alternative for compressive image recovery in lensless imaging systems.
• This unsupervised method significantly advances the practicality of compressive lensless imaging by removing the bottleneck of data acquisition.
• The demonstrated improvements in image quality highlight the potential of this approach for future compact and high-performance imaging devices.