Aiming at the problem that the traditional color transfer algorithm has color mis-transmission and unnaturalness as processing images, an image color transfer method based on dense connection generative adversarial network is proposed. During the training process, the training generation network generates the color transfer image. The coding layer in the generated network promotes the reuse of color features and speeds up the convergence rate of the network by using the cross-layer connection of the dense connection network. The conversion layer uses a three-layer residual module instead of the original two-layer residual module to combine different features of the image. The discriminating network is trained to distinguish the difference between the original image and the generated transfer image. The -log function is used to calculate the model loss in the network and speed up the initial update of the training. The experimental results show that the result image of this method retains more details compared with the similar model, can suppress some noise and the whole image is closer to the natural image.