Accurate eye localization in the Short Waved Infrared Spectrum through summation range filters

Abstract

The majority of facial recognition systems depend on the correct location of both the left and right eye centers in an effort to geometrically normalize face images. We propose a novel eye detection algorithm that efficiently locates the eye centers in five different bands of the SWIR spectrum, ranging from 1150 nm up to 1550 nm in increments of 100 nm. Our eye detection methodology utilizes a combination of algorithmic steps, including 2D normalized correlation coefficients as well as summation range filters to effectively find the eyes in the aforementioned SWIR wavelengths. We validate our method by comparing our approach with currently available eye detection algorithms including a commercial face recognition software in which one of its capabilities is the extraction of the eye locations and a state of the art academic approach. Eye detection results as well as face recognition studies show that our proposed approach outperforms all other approaches, including the state of the art (originally designed to work in the visible band), when operating in the SWIR spectrum. We also show that our approach is robust to typical image degradation factors including spatial resolution changes, image compression, and image blurring. This is an important achievement that has also practical value for biometric operators. It is impractical to manually annotate thousands to millions of eye centers, therefore, a quick and robust method for automatically determining the eye center locations is needed.