Learning the parameters for a gradient-based approach to image segmentation using cultural algorithms

There are two basic approaches to image segmentation, region-based and neighborhood-based. Region-based approaches require less a priori knowledge about the scene than neighborhood-based approaches but are computationally more expensive. In cases where there is little prior knowledge about properties of the image, one is often forced to use region growing approaches. In this paper the authors use cultural algorithms, a form of evolutionary computation based upon principles of cultural evolution, as the basis for learning the parameters for a neighborhood-based approach to image segmentation from the results of a region-growing approach. Specifically, parameters for a differential gradient method utilizing the Sobel operator are learned from the results of a region growing approach. The prototype is applied to a sequence of real world images, taken from archaeological excavations of a prehistoric site in order to extract spatial activity areas in the site. A region-growing approach is applied first to the images, and then a cultural algorithm is used to extract the parameters for use by a gradient method for those images. The resulting performance of the gradient method produced a correspondence of over 95% with that of the original