Improving face segmentation in thermograms using image signatures

Abstract

The aim of this paper is to present a method for the automatic segmentation of face images captured in (LWIR), allowing for a large range of face rotations and expressions. The motivation behind this effort is to enable better performance of face recognition methods in the thermal (IR) images. The proposed method consists on the modelling of background and face pixels by two normal distributions each, followed by a post-processing step of face dilation for closing holes and delimitation based on vertical and horizontal images signatures. Our experiments were performed on images of the (UND) and (FSU) databases. The obtained results improve on previous existing methods from 2.8% to more than 25% depending on the method and database. © 2010 Springer-Verlag.