Progressive compressive sensing of large images with multiscale deep learning reconstruction

Area of Science:

• Computational imaging
• Signal processing
• Machine learning for imaging

Background:

• Compressive sensing (CS) enhances imaging but faces challenges with large, high-resolution applications due to computational and acquisition demands.
• Current CS methods often require high-resolution sampling upfront, which is inefficient when only low-resolution data is needed.

Purpose of the Study:

• To introduce a multiscale progressive CS method for efficient high-resolution imaging.
• To enable fast deep learning reconstruction by leveraging multiscale properties of progressively acquired data.

Main Methods:

• Developed a progressive sampling strategy that refines image resolution while integrating previously acquired low-resolution information.
• Utilized the multiscale nature of the sensed samples for a computationally feasible deep learning (DL) reconstruction approach.

Main Results:

• Demonstrated 4-megapixel progressive compressive imaging using a single pixel camera.
• Achieved approximately 50% reduction in total samples compared to conventional CS.
• Reconstruction speed was improved by over an order of magnitude.
• Observed enhanced reconstruction quality.

Conclusions:

• The proposed multiscale progressive CS method offers an efficient solution for high-resolution imaging demands.
• This approach significantly reduces sampling requirements and accelerates reconstruction times, making high-resolution imaging more practical.
• The method shows superior performance over conventional CS techniques in terms of speed, sample efficiency, and image quality.