A Hybrid Approach for Eye Localization in Video Ashish Phophalia DAIICT Gandhinagar, India 382007 Email: 201021014@daiict.ac.in Bena Kansara Assistant System Engineer- Trainee, TCS Banglore, India 560066 Email: bena.kansara@tcs.com Suman K. Mitra DAIICT Gandhinagar, India 382007 Email: suman mitra@daiict.ac.in Abstract—Eye Localization is a preprocessing step for opera- tions such as orientation correction and scaling required in face recognition problem. The success of facial feature analysis and face recognition system depends on eye position detected and the other facial features estimated on this. This paper presents a hybrid approach for eye localization in a video which integrates efficiency of a feature based approach and accuracy of an ap- pearance based approach. In particular, the proposed approach combines Line Edge Map (feature based) approach and Bayesian Classifier (appearance based) approach. The experiments have been carried out on videos of people in normal blink as well as in sleepy condition. The performance of proposed algorithm has been validated on various non-parametric tests. Index Terms—Bayesian Classifier approach, Eye Localization, Eye Line Edge Map approach, Face recognition, Facial feature analysis I. I NTRODUCTION Eye localization is an important preprocessing step in hu- man face recognition. The accuracy of face alignment affects the performance of a face recognition system [13]. Face align- ment is usually carried out using the eye positions; therefore, an accurate eye localization method is essential for accurate face recognition. As there are many face detection methods available today, it is assumed that face region in the input image has been roughly localized by a face detector. Here, we address the work of accurately localizing the eyes in a roughly localized face images extracted using one of the available face detectors. It could be useful in the applications such as to find drowsiness of a person driving a car [14] and controlling computer based systems through eyes for those having motor difficulties. Another problem is to improve accuracy of the face recognition system by providing the accurate eye localization. The eyes are the most important facial landmarks on hu- man face, for both human computer communication and face normalization. As eye localization is an important process for operations such as orientation correction and scaling, success of facial feature analysis and face recognition depends greatly on eye detection. Also, it is useful to detect eyes before any of the other features of face because localization of other facial features is easy once we know the eye positions. Other features can be located using eye positions and golden ratio [3]. The approaches proposed till now are categorized as feature based, template based and appearance based approaches. In feature based approach [8], [6], [4], certain characteristics of eyes are explored which may include edge and intensity infor- mation, color distribution etc. Also, some distinctive features around the eyes and symmetry between left and right eye are taken into consideration. Using these features and golden ratio, eyes are localized. In template based approach [8], [4], [12], a model same as that of the shape of the eyes is designed. And this template is matched to the entire face image pixel by pixel to localize the eyes. In appearance based approach [5], [10], [11], eyes are detected by using their photometric appearance. In this approach, a classifier is trained oveer a large number of training data which represent eyes of different individuals in different conditions. The parameters tunned during the training phase are used in testing phase to localize eyes. There are certain cases where feature based approach may fail. For example, if a person is having a mark on the face or due to shining hair, unusual lighting condition may also cause to fail the test. Also, dark spots on the face which are geometrically similar to the eyes may also come up as a potential eye pairs. Due to symmetry between inner or outer corners of the eyes, eyelids, regions below eyelids and eyebrows can also serve as erroneous pairs. Therefore, feature based approaches do not give accurate results [8]. While template based and appearance based approaches find exact match to eye pairs, they are very time consuming in contrast to feature based approach as they need to traverse entire face image pixel by pixel to localize eyes. The efficiency of eye localization technique could be strengthen if either a feature based method or a template based method is applied within a smaller area that has high potential of containing eyes. In particular, if potential eye regions are localized on which a template needs to be matched or a classifier needs to be applied, then it could result in an efficient eye localizer. With this aim in mind, we propose here, a hybrid approach for eye localization. In first step, potential eye regions are localized using Eye Line Edge Map (LEM). In second step, an appearance based Bayesian Classifier approach is employed for localizing eye regions accurately. The proposed method is implemented on static faces and then implemented on face video to localize eyes. The results are encouraging and could be considered as a first step to detect the drowsiness of a person driving a car. Note that the actual experiment towards this is not carried out. The LEM, appearance based Bayesian Classifier and their hybridization are discussed in next Section. The experimental results of the proposed method are presented in Section III followed by the conclusion. 2011 Third National Conference on Computer Vision, Pattern Recognition, Image Processing and Graphics 978-0-7695-4599-8/11 $26.00 © 2011 IEEE DOI 10.1109/NCVPRIPG.2011.30 98