Since morphology of retinal blood vessels plays a key role in ophthalmological disease diagnosis, the automatic retinal blood segmentation method is essential for computer-aided diagnosis system. In this paper, a supervised method which is based on multi-level convolutional neural network is proposed to separate blood vessels from fundus image. By using both local and global feature extractors, small vessels can be well distinguished and global spatial consistency of the image can be ensured. Meanwhile, unsupervised pre-processing and postprocessing methods are applied to achieve better segmentation results. Experiment results on public database show that the proposed method outperforms the state-of-the-art performance (AUC up to >0.978) on DRIVE database.