Recently, with the development of remote sensing and computer techniques, automatic extraction and update of road information is becoming true. Remote sensing images can not only provide the surface information in large areas, but also save the costs in human resources, materials and time. Additionally, the update cycle of road network is now much shorter than traditional method. Nowadays, accurate extraction of road information from satellite data has become one of the most popular topics in both remotes sensing and transportation fields. However, as there is usually huge information provided by remote sensing data, an efficient and accurate method to refine the data and extract the road is thus important in real applications, such as the road update and emergency rescue decision making. By combining the deep convolution network and image segmentation approach, this paper proposed a new solution for extraction road network from high resolution images. For doing this, a road class table was built according to the road design and construction specifications that made by transportation industry. Following that, the ownership probability of different road classes should be predicted through deep convolutional network at pixel-level and the corresponding task was then regarded as a specific image segmentation. In addition, to make the extracted road segments more realistic, a post-processing approach was also used to modify, connect and smooth the road segments. Experiments in this paper showed that, the proposed solution successfully distinct multi-type roads from complex situations. Furthermore, a more realistic road map could also be provided with a total accuracy of more than 80% in discriminable areas.