Video Compressive Sensing Reconstruction Using Unfolded LSTM.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Video compression sensing (VCS) aims to reconstruct videos from limited measurements.
• Exploiting interframe correlations is crucial for improving VCS reconstruction quality.
• Existing deep learning methods often overlook interframe information or use complex structures.

Purpose of the Study:

• To propose an efficient end-to-end deep learning network for video compression sensing.
• To overcome limitations of existing VCS methods regarding interframe information utilization and network complexity.

Main Methods:

• An integrated framework combining measurement and reconstruction was developed.
• A trained measurement matrix, optimized for video reconstruction, was employed.
• An unfolded Long Short-Term Memory (LSTM) network was utilized to fuse spatial-temporal information.

Main Results:

• The proposed method achieved higher reconstruction accuracy compared to existing VCS networks.
• The network demonstrated robust performance even at a low measurement ratio of 0.01.
• Deep fusion of intra- and interframe information significantly improved reconstruction quality.

Conclusions:

• The developed efficient end-to-end VCS network effectively leverages interframe correlations.
• The approach offers superior reconstruction accuracy and efficiency for video compression sensing.
• This method provides a promising solution for high-quality video reconstruction from sparse data.