Segmentation of high resolution images based on the multifractal analysis

The edge content of very high resolution images, such as those from Ikonos, is very important due to the huge amount of details provided. Classical methods usually fail to achieve a good segmentation result on such images. We studied a new method for high resolution optical image segmentation which is based on the multifractal characterization of the image. Starting from the analysis of the Holder regularity at each point, we extract features leading to the segmentation of the image. Based on information from the high frequencies, we use a k-means clustering algorithm to perform the segmentation. The whole algorithm is described and results of the method applied to Ikonos image as well as a comparison with classical co-occurrence techniques are presented.