Road information construction is the key support to solve urban traffic problems. The technique of segmenting the road part from the image is called road extraction. Therefore, road extraction can be used as a technical support for traffic system to obtain a large amount of road information. Due to the advantages of high resolution and overhead shooting angle for remote sensing image, it is very suitable as the data source of road extraction. Although road extraction from remote sensing images becomes one of the important research directions and some achievements have been made, there are still many challenges. Different from common objects, roads have strong topological continuity and are also composed of many irregular curves. This special morphological feature is difficult to be mined by common convolution kernel. To solve the above problems, this paper proposes a novel feature extraction backbone network DICMNet, which is composed of dynamic irregular convolutional (DIC) module and multi-directional channel remapping (MDCR) combine with resnet. Experimental validation with the other methods is performed on the DeepGlobe road dataset. Our model has a significant improvement on most of the evaluation metrics with an average improvement of 2%.