Quantitative phase imaging has the characteristics of being label-free, non-damaging, and capable of three-dimensional measurement, and has been applied in many fields. In recent years, artificial intelligence represented by deep learning has proven to be very effective in solving phase imaging. This article aims to introduce the research progress of phase recovery methods based on deep learning, briefly introduce the traditional methods of quantitative phase imaging, and comprehensively review the existing deep learning quantitative phase imaging technology. In addition to pure data-driven network methods to recover from intensity maps through training In addition to phase strategies, deep learning networks can also be combined with physical models for phase recovery, and the application of deep learning quantitative phase imaging technology in fields such as biomedicine and industrial measurement was discussed. The challenges faced by current deep learning-based phase imaging methods and the development directions of future research are discussed. This article summarizes the work of deep learning in the field of quantitative phase imaging and puts forward prospects for how to better utilize deep learning to improve the reliability and efficiency of quantitative phase imaging. First, traditional quantitative phase recovery methods are introduced. The traditional phase recovery method is to introduce additional information (sample operation, system adjustment and multiple acquisitions) to transform phase recovery into a well-posed/deterministic problem, such as holography or interferometry that introduces reference waves, Shack-Hartmann wavefront sensing that introduces microlens arrays in conjugate planes, and intensity transfer equations that require multiple defocus amplitudes. But the trade-off is increased system complexity or the time and effort of capturing multiple intensity patterns in exchange for a deterministic and straightforward solution. In this case, the phase recovery problem is well-posed or even overdetermined. In addition, prior knowledge can also be used to solve this ill-posed phase recovery problem by seeking the global optimal solution through optimization iteration from intensity measurements, which is the so-called phase retrieval.Then deep learning-based methods were introduced, including data-driven network methods and combine physical methods with deep learning technology for phase recovery. This data-driven network uses a large number of paired intensity map and phase ground truth datasets as implicit priors to iteratively train the initialized neural network. After training, the network is used as an end-to-end mapping. There are three main ways to combine physical methods with deep learning technology: physical cascade network, network embedded physics and embedded physical network. The networks combine with physical methods avoids the inherent problems of data-driven method such as data dependency, poor generalization, and lack of interpretability. It also does not require multiple holograms of traditional physical methods and can obtain faster results, reconstruct higher quality phase information and is more robust to noise. Due to its imaging principle, traditional phase recovery methods have some problems such as slow imaging speed and artifacts in reconstructed phase. Solving these problems using traditional physical methods or algorithms is often complicated and difficult to achieve the desired results. AI-QPI has greatly overcome various defects in phase recovery with its excellent ability to solve the inverse problem of image reconstruction. This article has conducted an extensive survey on the latest progress of deep learning phase recovery method, mainly discussing data-driven network recovery, and deep network methods combined with physical methods for recovery phase, and related improvements. Continuous optimization and improvement of phase recovery technology based on deep learning have great significance to the advancement of imaging and artificial-intelligent sensing technology, and has a positive role in promoting the development of biological microscopy, deformation detection, particle field analysis, micro-nano device detection and other fields.