This paper presents an application of a neural network, namely the hierarchical cluster model (HCM) to intermediate-level image segmentation. The HCM forms a biological model of the brain for image region segmentation employing Gestalt rules. In particular, a three level HCM is proposed to hierarchically merge pixels into regions and methods are developed to quantify the Gestalt properties of similarity, continuity, closure and co-circularity as merging evidence between regions. Experiments have shown that the proposed algorithm produced more consistent results to manual segmentation than the well-known JSEG method.
Available at: http://works.bepress.com/wli/16/