Facial feature detection using compact vector-field canonical templates

Detecting and analyzing prominent facial regions forms the fundamental building block in most face recognition systems. Prominent regions such as left and right eyes, tip of the nose, mouth, etc. are localized to derive an overall representation of the face being recognized. In this paper, we present a method for deriving a set of compact translation-, scale- and rotation-invariant canonical templates which could be used on a large database. In contrast to conventional gray scale templates, these are of the 2D gradient field type. Facial feature detection is based on evidential reasoning from the measures of belief and disbelief estimations. The above method is demonstrated on a facial image database of size 137 using only 9-left, 9-right and 9-nose tip canonical templates