An method of automatic road extraction based on directional statistical characteristics

When roads appear as irregular linear targets, most of traditional methods which use local information often give a false alarm of linear feature. In order to overcome this disadvantage, we present an approach to extract main roads in aerial images with oriented filters. Firstly, the leading gradient direction of the local areas is estimated. Then the image is filtered along this direction. We use steerable filters to control the output and get the maximum output response along the road direction. Based on the steps above, we make much improvement in details such as the determination for the size of the sub-block image, the processing of image edge pixels and the processing of orientation angel areas and so on. Experiments on remote sensing images of different environmental settings show that our proposed scheme achieves high accuracy in the road extraction.