Learned reconstructions for practical mask-based lensless imaging.

Area of Science:

• Computational imaging
• Optics
• Computer vision

Background:

• Mask-based lensless imagers offer size and weight advantages over traditional cameras.
• Current model-based reconstruction methods are computationally intensive and require extensive calibration.
• Existing techniques often rely on heuristic denoisers, limiting performance.

Purpose of the Study:

• To address limitations of traditional mask-based lensless imaging reconstruction.
• To develop a faster and more accurate image reconstruction method.
• To explore a hybrid approach combining physical models and deep learning.

Main Methods:

• Unrolling a traditional model-based optimization algorithm into a trainable neural network.
• Optimizing network parameters using experimentally gathered ground-truth data.
• Optionally incorporating a jointly-trained denoiser for enhanced image quality.

Main Results:

• Achieved 20x faster reconstruction speeds compared to traditional methods.
• Demonstrated superior perceptual image quality.
• Successfully generalized to natural images captured with a prototype camera.

Conclusions:

• The proposed trainable neural network effectively reconstructs images from mask-based lensless imagers.
• The hybrid approach offers a balance between model-based and deep learning methods, yielding significant improvements.
• The method enables interactive scene previewing and shows promise for real-world applications.