Underwater imaging is one of the most commonly used methods for ocean exploration, and a growing number of studies have shown that polarization is the key to the underwater creatures having vision in low illumination. In this paper, a multi-turbidity underwater image recovery method based on deep learning and polarization imaging is proposed. Multiple turbidity underwater polarization data sets are obtained by capturing images of clean water and underwater polarization images with different turbidity. A small size neural network is proposed to better learn the mapping relation between underwater polarization information under different turbidity and clear underwater images. A sliding window superposition method with different steps is proposed for different circumstances. The results show that the polarization method proposed in this paper can effectively recover the underwater image, and the peak signal to noise ratio recovered under different turbidity is 47.39% higher than that of the original image on average. The proposed method combining deep learning and polarization imaging technology can restore underwater images in multi-turbidity environment and overcome the problem of poor restoration effect of ordinary underwater images.