基于深度学习的压缩感知重构及其应用现状与展望

Status and Trends in Deep-learning-assisted Compressed Sensing Reconstruction and Its Applications

经典的奈奎斯特采样定理要求采样频率大于两倍信号最大频率,才能重构出原始信号,但是当信号具有稀疏性时,压缩感知方法可突破奈奎斯特采样定理的限制,大大降低采样频率,显著提高数据采集和分析的效率。近年来,压缩感知作为一种全新的稀疏信号处理框架,已广泛使用于图像重建、稀疏信号重构等领域。随着深度学习技术的快速发展,将各种深度神经网络用于压缩感知领域中,既能提高重构的精度,又能减少重构的时间。通过对公开文献进行广泛调研,介绍了压缩感知重构的基本概念,梳理了常用的压缩感知算法,归纳了深度压缩感知的主要应用领域。最后探讨了当前仍面临的问题及未来的发展趋势。

The classic Nyquist sampling theorem requires a sampling frequency greater than twice the maximum signal frequency for reconstructing the original signal.Compressed sensing method can break through this limitation when coping with signals with sparsity,significantly reducing sampling frequency and improving efficiency of data collection and analysis,and therefore in recent years has been regarded as a new sparse signal processing framework and found wide applications in fields such as image reconstruction and sparse signal reconstruction.Moreover,rapid development of deep learning technologies enables the combination of compressed sensing with various types of neural networks,thus achieving both improved reconstruction accuracy and reduced reconstruction time consumption.This review conducts extensive literature survey,elaborates basic technological concepts,and summarizes prevailing compressed sensing algorithms and primary applicative aspects of deep compressed sensing,with respect to deep-learning-assisted compressed learning.It also critically discussed current problems and future development trends.