Analysis of the effect of selecting statistically significant registered image pixels on individual face physiognomy recognition accuracy
Iliana V. Voynichka, Alila B. Megherbi · 2016
Research in facial recognition techniques that can identify individuals in harsh conditions, such as crowded places with limited visibility, is still in its infancy. Such research is becoming more imperative in many applications including in homeland security. In order to build better facial recognition systems there is an imperative and primary need to understand how different factors and various facial features affect the recognition accuracy of existing popular facial recognition algorithms. In our prior work we presented the effect artifacts such as image registration, number and type of training templates, presence of varying amount of partial facial information, and the presence of color have on facial recognition accuracies. In this paper, we continue our investigation into how certain factors affect facial recognition by looking into what are the most statistically significant pixel-features in an image that differentiate a given individual face from the rest of the individual faces in a given data set. In particular, we propose an algorithm to derive a mask of the pixels with the highest statistical significance levels to identify a given face based on this mask. Our investigation shows that creating a mask using the two-sample t-test, selects the pixels that are most representative of a given individual face physiognomy when compared to face images of the rest of the individuals in a given database. We also show that using the t-test based mask relatively improves the face recognition accuracy in certain cases when compared to using the original full pixel intensity images. Finally, we show that using statistically significant pixel-based masks that differentiate an individual from the rest of the individuals, yields better results than those obtained using pixels that are instead common among similar individual faces.