When polarized light propagates in the scattering medium, the polarization information is disturbed and lost due to multiple scattering of scattering particles. In order to ensure efficient and high-fidelity transmission of polarization information in scattering media, a polarization recognition method through scattering media based on deep learning is proposed. A convolutional neural network is constructed to extract the characteristics of the polarization information of incident light wave from the speckle light intensity information to realize the high resolution recognition of the polarization state of incident light wave, and the robustness of the convolutional neural network for polarization state recognition is verified by using polarized light with different initial phases. Experimental results show that the proposed method has the advantages of fast recognition speed and high accuracy, and the neural network can be trained with infinite data in theory. Therefore, the method has great application potential in polarization optical imaging and laser communication.