G-Images: Towards Multilevel Unsupervised Image Segmentation

We present two novel approaches to initiating unsupervised segmentation of digital images using an algorithm that utilises the concept of information theory. The first approach uses Information Gain and the second is based on the Gini Index. In the two approaches, Information Gain and the Gini Index are calculated locally, at a pixel level, resulting in a G-image where high G value occurs at contrasting boundaries and zero G value within homogeneous regions. Subsequently, a multi-level thresholding approach based on the G-image is used to obtain the optimal segmentation results. The segmentation is guided by both local and global parametric constraints. Comparative, visual, evaluation on real and artificial images shows promising results.