Eye localization for face recognition

Abstract

We present a novel eye localization method which can be used in face recognition applications. It is based on two SVM classifiers which localize the eyes at different resolution levels exploiting the Haar wavelet representation of the images. We present an extensive analysis of its performance on images of very different public databases, showing very good results.

Automatic Facial Feature Extraction for Face Recognition

“…Before the wavelet decomposition, each eye patch undergoes an illumination normalization process (a contrast stretching operation) and is then reduced to 16×16 pixels. 4 The decomposition is realized via an overcomplete bi-dimensional FWT (Fast Wavelet Transform) [Campadelli et al, 2006a] that produces almost four times as many coefficients with respect to the standard FWT. This redundancy is desirable as we want to increase the cardinality of the feature "vocabulary" before going through the selection procedure.…”

Automatic Eye Localization in Color Images

“…, max EL literature points out that d eye ≤ 0.25 roughly corresponds to a distance smaller than the eye width and therefore it can be used as criterion to claim eye localization [1,3,4,7]. However, this accuracy level may not be sufficient when localized positions are used for some face processing techniques, specially in the case of face recognition methods.…”

“…Automatic extraction of human face and facial features (eyes, nose and mouth, for example) is an essential task in various applications, including face and iris recognition, security, surveillance systems and human computer interfacing [1,2,3,4,5]. Facial feature detection and localization (FFDL) is a very important problem to be solved, because it provides meaningful input for most face processing algorithms.…”

“…There are several reasons for this outstanding importance of the eyes over other facial features [3,4,6,7]: eyes reveals information about the state of human beings; a face contains two eyes (if not occluded), so its position, scale and orientation can be estimated from the eye positions; and the appearance of eyes is less variant to face changes (than eyebrows, nose, ears and mouth, for example). Moreover, accurate EL provides means to identify all the other facial features of interest [1].…”

“…Although all research effort made in last years [1,3,6,7,8,9,10,11,12,13,14], EL remains an open problem due to increasing demand of accuracy and speed in eye localization [1,3,4]. It is well known that the behavior of face processing methods (face recognition, in particular) in real-life applications strongly depends on precise EL [1,13]. In this context, most approaches to EL are not accurate enough or perform very poorly in terms of efficiency, especially when cascaded AdaBoost classifiers [5] are used and the eyes are successfully localized [5,8] -the most time consuming case for this kind of methods.…”