A Learning-based Eye Detector Coupled with Eye Candidate Filtering and PCA Features

In this work, we present a system based on a Neural Network classifier for eye detection in human face images. This classifier works on eye candidate regions extracted from a face image and represented by a reduced number of features, selected by Principal Component Analysis. The regions are determined considering that in an image window containing the eye, the grey level distribution will generally assume a pattern of adjacent light-dark-light horizontal and vertical stripes, corresponding to the eyelid, pupil and eyelid, respectively. For training, validation and testing, a database was built with a total of 4,400 images. Experimental results have shown that the proposed approach correctly detects more eyes than any of two existing systems (Rowley-Baluja-Kanade and Machine Perception Toolbox), for eye location error tolerances from 0 to 5 pixels. Considering an error tolerance of 9 pixels, the correct detection rate achieved was above 90%.