Natural scene segmentation based on information fusion and hierarchical self-organizing maps

This thesis focuses on an intuitive approach to natural scene segmentation. This research uses color and texture features in cooperation to provide comprehensive knowledge about every pixel in the image. A novel scheme for the collection of training samples, based on the notion of homogeneity, is proposed. Natural scene segmentation is carried out using a two-stage hierarchical self-organizing map. The first stage of the network employs a fixed-sized two-dimensional map that captures the dominant color and texture features of an image, in an unsupervised mode. The second stage combines a fixed-sized one-dimensional feature map and color merging, to control the number of color clusters formed as a result of the segmentation. The proposed method confirms that the self-learning ability and adaptability of the self-organizing map, coupled with the information fusion mechanism of the hierarchical network, leads to good segmentation results. These are further confirmed by extensive tests on a variety of natural scene images.