To address the problem that a single polarization image cannot provide sufficient information in a certain scene, a polarization image fusion algorithm based on dense gradient generative adversarial networks are proposed by combining the advantageous features of intensity image and line polarization degree image. A dense gradient convolution module is constructed using a densely connected convolution network and gradient operator, and the module is applied to the generator to enhance the texture details of the fused images; a loss function combining multi-scale structural similarity and L1 parametrization are used to improve the overall performance of the network. Qualitative comparison and quantitative analysis were performed on the ZJU-RGB-P dataset, and the experimental results showed that the proposed algorithm has better subjective visual perception, while all evaluation indexes were significantly improved.