Image Compressed Sensing using Convolutional Neural Network.

Area of Science:

• Signal Processing
• Computer Vision
• Machine Learning

Background:

• Compressed Sensing (CS) faces challenges in sampling matrix design and reconstruction method development.
• Existing random sampling matrices are signal-independent, and state-of-the-art CS methods have high computational complexity.
• Current methods often ignore signal characteristics, leading to suboptimal reconstruction.

Purpose of the Study:

• To propose an efficient and effective image compressed sensing framework.
• To address the limitations of signal-independent sampling matrices and high computational complexity in CS.
• To develop a joint optimization approach for sampling and reconstruction networks.

Main Methods:

• Introduced CSNet, a convolutional neural network framework for image compressed sensing.
• Developed a joint framework with a sampling network and a reconstruction network, optimized together.
• Learned adaptive sampling matrices (floating-point, binary, bipolar) and an end-to-end reconstruction network.

Main Results:

• CSNet achieved state-of-the-art reconstruction quality with fast running speed.
• Learned binary and bipolar matrices showed comparable performance to deep learning CS methods and outperformed traditional ones.
• Learned sampling matrices significantly improved traditional image CS reconstruction methods.

Conclusions:

• CSNet offers a powerful solution for image compressed sensing by adaptively learning sampling matrices.
• The framework provides high reconstruction quality and computational efficiency.
• Learned sampling matrices hold potential for advancing both deep learning and traditional CS techniques.