As light propagation in water is subject to absorption and scattering effects， acquiring underwater images using conventional intensity cameras can result in low brightness of imaging results， blurred images， and loss of details. In this study， a deep fusion network was applied to underwater polarimetric images； the underwater polarimetric images were fused with light-intensity images using deep learning. First， the underwater active polarization imaging model was analyzed， an experimental setup was built to obtain underwater polarization images to construct a training dataset， and appropriate transformations were performed to expand the dataset. Second， an end-to-end learning framework was constructed based on unsupervised learning and guided by attention mechanism for fusing polarimetric and light intensity images and the loss function and weight parameters were elaborated. Finally， the produced dataset was used to train the network under different loss weight parameters and the fused images were evaluated based on different image evaluation metrics. The experimental results show that the fused underwater images are more detailed， with 24.48% higher information entropy and 139% higher standard deviation than light-intensity images. Unlike other traditional fusion algorithms， the method does not require manual weight parameter adjustment， has faster operation speed， strong robustness， and self-adaptability， which is important for ocean detection and underwater target recognition.