Optimizing level set initialization for satellite image segmentation

Obtaining segmentations of buildings from satellite images for telecommunication applications is a complex process due to the fact that satellite or aerial images are complicated scenes. The algorithm presented in this work uses level set Chan-Vese formulation, to establish the corresponding boundaries of the buildings. Level set segmentation tends to be sensitive on initialization, thus proper initialization can yield better segmentation results. In this work an effective procedure is performed using K-mean classifier in order to design and develop the initial level set contours. Morphological features are incorporated for refining the obtained outlines. Finally, the coordinates of each and every building are extracted along with additional information for the processed scene, like the number of buildings, as well as the center and area of each building. The optimization algorithm was evaluated qualitative and quantitative against the original Chan-Vese model and proved to provide more accurate results.