Abstract: Compressed sensing breaks through Nyquist Sampling Theory(NST), and greatly alleviates the pressure of data acquisition and transmission. In recent years, with the rapid development of deep learning, the application of deep neural network technology in the field of compressed sensing has effectively improved the reconstruction accuracy and efficiency of compressed sensing, and attracted extensive attention and research of scholars.In order to sort out and summarize the existing compressed sensing image reconstruction algorithms based on deep learning, first, introduces the basic mathematical knowledge of compressed sensing and two typical traditional reconstruction iterative optimization algorithms of compressed sensing: ISTA and ADMM. Then, we discuss in detail the deep network deployment framework of the two traditional algorithms as well as the improvement techniques of the benchmark framework: ISTA-Net⁺⁺and ADMM-Net, and contrastive analysis of five end-to-end deep neural network frameworks developed by unconventional algorithms including SDA, ReconNet and DR²-Net. After that, using Peak Signal-to-Noise Ratio (PSNR) as evaluation index, the reconstruction accuracy of representative network model on natural image dataset Train 400 and medical image dataset MICCAI was compared and analyzed. Finally, the research prospect of deep learning technology in the field of compressed sensing reconstruction is summarized and prospected.