The retinal vessel segmentation in color fundus images is of great value for the clinical diagnosis and a retinal vessel segmentation method based on an improved convolutional neural network is proposed. First, the residual learning is combined with the densely connected network (DenseNet) to fully exploit the feature maps of each layer. The path from the low-level feature maps to the high-level ones via the addition of shortcuts is shortened and the feature propagation ability is strengthened. Second, as for the extraction of more fine vessels, the dilated convolutions are adopted in the encoder-decoder network to expand the receptive field without the increase of parameters. The experimental results show that the proposed network structure has less parameters, compared with the other existing deep learning methods. The average accuracy on the DRIVE datasets is up to 0.9556, the sensitivity is up to 0.8036, the specificity is up to 0.9778, the area under curve of receiver operating characteristic reaches 0.9800, better than the segmentation effects of the other existing deep learning methods.