Accurate extraction of human faces and their components from color digital images based on a hierarchical model

Many human face-based applications depend on accurate extraction of human faces and their components from complex color digital images. Most existing methods are only based on human skin regions and their adjacency relations, but do not utilize their sub-division and grouping at multiple levels. This paper presents a hierarchical model to address this. The model contains pixels, runs, regions, and their respective relations for a digital image. Human skin regions are identified by combining multiple color spaces. Adjacent human skin regions are grouped as a patch. Human face candidates are extracted from each refined human skin patch by un-linking adjacent regions of some regions based on a human facial shape model. Non-human skin patches in a human face candidate are identified as human facial internal component candidates according to the attributes and configurations of human facial components. Each human face candidate is further classified as a human face or not based on its component candidates. An extracted human face provides not only an accurate human facial region but also its components. Finally, this paper demonstrates experimental results of extracting various human faces and their components from three databases, showing that the proposed approach is more accurate and robust than other approaches.