Abstract

We present a fast method for road network extraction in satellite images. It can be seen as a transposition of the segmentation scheme “watershed transform + region adjacency graph + Markov random fields” to the extraction of curvilinear objects. Many road extractors which are composed of two stages can be found in the literature.

The first one acts like a filter that can decide from a local analysis, at every image point, if there is a road or not. The second stage aims at obtaining the road network structure. In the method, we propose to rely on a “potential” image, that is, unstructured image data that can be derived from any road extractor filter. In such a potential image, the value assigned to a point is a measure of its likelihood to be located in the middle of a road. A filtering step applied on the potential image relies on the area closing operator followed by the watershed transform to obtain a connected line which encloses the road network. Then a graph describing adjacency relationships between watershed lines is built.

Defining Markov random fields upon this graph, associated with an energetic model of road networks, leads to the expression of road network extraction as a global energy minimization problem. This method can easily be adapted to other image processing fields, where the recognition of curvilinear structures is involved.