Scalable Coding for High-Resolution, High-Compression Ratio Snapshot Compressive Video

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• High-speed cameras are essential for capturing rapid events but face storage, bandwidth, and cost limitations.
• Snapshot compressive video (SCV) uses compressed sensing to record dynamic scenes in a single image, enabling video recovery via inverse problem solving.
• Existing SCV methods struggle with scalability in spatial and temporal resolution for both data acquisition and video reconstruction.

Purpose of the Study:

• To introduce a versatile, scalable coded aperture approach for snapshot compressive video (SCV) that addresses limitations in spatio-temporal resolution.
• To enable on-the-fly, high-compression ratios with reduced computational and memory demands for SCV systems.
• To develop a universal sampling scheme compatible with various SCV reconstruction algorithms.

Main Methods:

• Leveraging time-division multiplexing to design a scalable coded aperture system for SCV.
• Implementing a universal sampling scheme applicable to diverse SCV acquisition and reconstruction strategies.
• Validating the approach through simulations and experimental results for high-resolution video compression.

Main Results:

• Achieved compression of up to 512 frames (2K x 2K resolution) into a single snapshot, a 0.2% compression ratio.
• Delivered reconstruction quality exceeding 30 dB PSNR with conventional algorithms.
• Surpassed 36 dB PSNR using state-of-the-art deep learning reconstruction algorithms.

Conclusions:

• The proposed time-division multiplexing coded aperture method offers significant spatio-temporal scalability for snapshot compressive video.
• This approach provides high compression ratios with low computational and memory overhead, making high-speed imaging more accessible.
• The method demonstrates the potential for high-quality video reconstruction from highly compressed single snapshots, advancing the field of compressive imaging.