Learned lensless 3D camera.

Area of Science:

• Computational Imaging
• Optics
• Machine Learning

Background:

• Single-shot 3D imaging faces challenges in device size, image quality, and processing speed.
• Mask-based lensless imagers offer portability but typically require extensive calibration and computational resources.
• Existing lensless imaging methods struggle with point spread function calibration and heavy computational demands.

Purpose of the Study:

• To demonstrate a compact and learnable lensless 3D camera for real-time, photorealistic imaging.
• To overcome the limitations of traditional lensless imaging, including calibration complexity and computational load.
• To enable advanced 3D imaging capabilities such as depth resolution and seeing through opaque obstacles.

Main Methods:

• Custom design and fabrication of an optical phase mask with optimized spatial frequency support and axial resolving ability.
• Development of a physics-aware deep learning model incorporating an adversarial learning module for reconstruction.
• Implementation of a single-shot imaging approach for rapid data acquisition.

Main Results:

• Demonstrated a compact lensless 3D camera capable of real-time photorealistic imaging.
• Achieved depth-resolved reconstructions without the need for point spread function calibration.
• Showcased the ability to image features blocked by opaque obstacles ('see-through' capability).

Conclusions:

• The developed lensless imager offers a practical solution for compact, high-quality, real-time 3D imaging.
• The physics-aware deep learning approach significantly reduces computational requirements and eliminates the need for calibration.
• This technology has broad potential applications in various fields requiring advanced computational imaging.