The study of imaging through scattering media especially 3D imaging is of great significance in many fields such as biomedical imaging. Recently, deep learning has been widely used in the field of information processing with its remarkable performance. In this paper, we proposed a method of three - dimensional imaging through scattering media based on deep learning. This method uses the deep neural network to process the information captured by the light field imaging system based on the microlens array, recovering the no-scattering 4D light field information, and then realize three-dimensional reconstruction by using the processed light field information. Deep learning method requires a large number of samples. But in many environments, it is difficult to obtain a large number of three-dimensional samples through experiment. To solve this crucial problem, we use incoherent light propagation model to simulate the light field propagation and generate samples which contains three-dimensional information through simulation. In this paper, we simulated the propagation of radiation emitted from objects behind a single layer of weak scattering media, generated a large number of samples of 4D light field information by simulation, trained the neural network and processed the test data set generated by simulation, and we realized the deblurring of the light field information which contains information of multiple layers of flat semitransparent objects, which could be used to realize the 3D reconstruction.