Less is more: Cropping to improve facial recognition with thermal images

Abstract

Numerous studies have investigated the functionality of the human facial recognition mechanisms with aims to improve computer vision methodologies. Understanding these mechanisms is important for applications in computer vision, as one goal of researchers is to create automatic facial recognition systems to equal the robustness of the human facial recognition system. Work in the field of human facial recognition shows that humans are able to recognize faces under very challenging scenarios, including partial occlusion. However, most human facial recognition experiments are conducted in the visible and near-infrared spectrums. During human facial recognition experiments with thermal images conducted in the lab, we found that removing the lower face yields better recognition results than when using the entire face. These tests included standard facial recognition algorithms and the results showed Rank 1 Recognition improvements of 10–25%, depending on the amount of occlusion derived from certain anatomical landmarks.