Deep Compressed Imaging via Optimized-Pattern Scanning.

Area of Science:

• Computational imaging
• Optical sensing
• Machine learning for imaging

Background:

• High-speed imaging is crucial for applications in biomedicine, surveillance, and consumer electronics.
• While silicon focal plane array sensors advance, single-pixel detector imaging gains traction due to computational algorithms.
• Existing single-pixel camera methods face limitations in acquisition speed.

Purpose of the Study:

• To introduce a novel imaging modality, Deep Compressed Imaging via Optimized-Pattern Scanning (DeCIOPS), for significantly faster image acquisition.
• To demonstrate DeCIOPS's ability to reconstruct high-quality images from minimal samples at high frame rates.
• To validate the DeCIOPS method under different illumination conditions.

Main Methods:

• Developed an end-to-end optimized auto-encoder utilizing a deep neural network and compressed sensing algorithm.
• Optimized illumination patterns for projection and scanning across an object.
• Collected sampling signals using a single-pixel detector.
• Reconstructed images from a reduced number of samples.

Main Results:

• Achieved significantly higher imaging speeds compared to conventional switching-mask and point scanning systems.
• Maintained comparable imaging quality to existing methods.
• Successfully reconstructed high-quality images with a high compressed sampling rate under both continuous-wave and pulsed light illumination.
• Experimentally validated the DeCIOPS imaging modality.

Conclusions:

• DeCIOPS offers a substantial increase in acquisition speed for single-detector imaging systems.
• The deep learning-based optimized pattern scanning enables high-frame-rate imaging with minimal data.
• This compressed sensing modality holds potential for widespread application in high-speed imaging systems.