Facial component detection in thermal imagery

Brais Martínez, Xavier Binefa, Maja Pantic

Abstract: This paper studies the problem of detecting facial components in thermal imagery (specifically eyes, nostrils and mouth). One of the immediate goals is to enable the automatic registration of facial thermal images. The detection of eyes and nostrils is performed using Haar features and the GentleBoost algorithm, which are shown to provide superior detection rates. The detection of the mouth is based on the detections of the eyes and the nostrils and is performed using measures of entropy and self similarity. The results show that reliable facial component detection is feasible using this methodology, getting a correct detection rate for both eyes and nostrils of 0.8. A correct eyes and nostrils detection enables a correct detection of the mouth in 65% of closed-mouth test images and in 73% of open-mouth test images.