Auto-labeling algorithms on CA based interactive segmentation for high resolution remote sensing images

Abstract

Interactive image segmentation based on Cellular Automata (CA) has shown its effectivity of object extraction in photographs or video frames. But in high resolution remote sensing images it will be a heavy work to manually label out all of objects, especially the mixture-up objects, in large scale of map area. Three kinds of auto-labelling algorithms are studied to generate labels automatically just according to a few of artificial sample labels. These algorithms deal with the similarity between unlabeled area and the labeled samples in the aspects of spectrum, shapes, and mixture pattern respectively, and then make the use of maximum likelihood labelling, shape calculation by screening contours, and spatial clustering comprehensively to extract some feature pixels to construct new labels. The feasibility of the algorithms has been shown by experimental results of different types of high resolution remote sensing images retrieved from google earth.