The appearance and structure of retinal vessels is one of the diagnostic criterias for ophthalmic, cardiovascular and cerebrovascular diseases, which makes the segmentation of retinal vessels worthwhile in clinical medicine. Retinal images have low contrast, complex tissue structures, and diverse pathological conditions. Due to the difficulty to accurately and effectively identify these information, existing segmentation methods often have neglection on subtle vessels and have blurry segmentation on overlap area. This paper proposed a new vascular segmentation method based on deep learning. Considering the necessity of feature extraction from multi-aspect and multi-level, a cross-block dense connection was proposed to improve the feature identify and extract ability. It use additional mapping to connect and share features between different blocks of the network. Further, the residual connection is used inside the block to solve the degradation problem. Powerful extraction capability enables the network to obtain ideal segmentation results with less parameters, which immensely reduces the computing resource consumption of the network. Extensive experiments demonstrate that our approach achieves better performance based on multiple criteria with a much faster speed.