03-02-2010, 08:39 PM
Automatic Eye Detection and Its Validation
The accuracy of face alignment affects the performance of
a face recognition system. Since face alignment is usually
conducted using eye positions, an accurate eye localization algorithm is therefore essential for accurate face recognition.
Automatic Eye Detection
The main uses of eye detection are:
-detect the existence of eyes
-accurately locate eye positions
: active and passive eye detection systems are available today.
- Passive methods directly detect eyes from images within
visual spectrum and normal illumination. Some early work
extracts distinct features from eyes localization.
-Active eye detection methods use special types of illumination. Under
IR illumination, pupils show physical properties which can
be utilized to localize eyes The advantages of active eye detection methods are that they are very accurate and
robust.But they need special lighting to work.
Eye Localization Algorithm
To better represent eyes the statistically learning of
discriminate features to characterize eye patterns is proposed. learning of probabilistic classi?ers to separate eyes and non-eyes is also studied. Multiple classi?ers are then combined in AdaBoost to form a robust and accurate eye detector.
Discriminant Features for Eye Detection
A training sample is containing the image intensity
data, and the sample label as the choosing criteria is taken. In this paper,
One criteria to extract a good feature for pattern classi-
?cation is that the feature can minimize the estimated
Bayes error function The Fisher discriminant analysis (FDA) is equiv-
alent to Bayesian classi?er if assuming Gaussian distribu-
tion and equivalent priors and covariance matrix for each class.
Feature Selection and Classi?er Construction with AdaBoost
The AdaBoost selects the critical features and train weak classi?ers as well
as updates the sample weights. The main task in the AdaBoost is the selection of features
to learn weak classi?ers. more powerful discrim-
inant features is used instead of rectangular Haar features to im-
prove eye detection accuracy. To train a robust eye detector, training
data was collected from various sources. 500 pairs of eyes were collected from a database for study. training. only a left eye detector is trained In application,
due to the symmetry of eyes.
Eye Localization
The eye localization method follows a hierarchical princi-
ple. a face is detected first, then eyes are located inside
the detected face. Ada boost is used here too. multiple eyes detected around the pupil center. The ?nal eye localization is the average of the multiple detection results.
Eye Detection Validation
In one kind of validation experiments, a set of manually labeled eye positions were used. The performance of our eye detector is characterized by the eye detection rate and
eye localization error. The localization error is measured as the Euclidean distance between the detected eye posi-
tions and manual eye positions. In the second experiment, performance of eye detection was measured based on
its in?uence on face recognition accuracy of two standard
baseline methods: PCA and PCA together with LDA.
full report:
[attachment=1711]
The accuracy of face alignment affects the performance of
a face recognition system. Since face alignment is usually
conducted using eye positions, an accurate eye localization algorithm is therefore essential for accurate face recognition.
Automatic Eye Detection
The main uses of eye detection are:
-detect the existence of eyes
-accurately locate eye positions
: active and passive eye detection systems are available today.
- Passive methods directly detect eyes from images within
visual spectrum and normal illumination. Some early work
extracts distinct features from eyes localization.
-Active eye detection methods use special types of illumination. Under
IR illumination, pupils show physical properties which can
be utilized to localize eyes The advantages of active eye detection methods are that they are very accurate and
robust.But they need special lighting to work.
Eye Localization Algorithm
To better represent eyes the statistically learning of
discriminate features to characterize eye patterns is proposed. learning of probabilistic classi?ers to separate eyes and non-eyes is also studied. Multiple classi?ers are then combined in AdaBoost to form a robust and accurate eye detector.
Discriminant Features for Eye Detection
A training sample is containing the image intensity
data, and the sample label as the choosing criteria is taken. In this paper,
One criteria to extract a good feature for pattern classi-
?cation is that the feature can minimize the estimated
Bayes error function The Fisher discriminant analysis (FDA) is equiv-
alent to Bayesian classi?er if assuming Gaussian distribu-
tion and equivalent priors and covariance matrix for each class.
Feature Selection and Classi?er Construction with AdaBoost
The AdaBoost selects the critical features and train weak classi?ers as well
as updates the sample weights. The main task in the AdaBoost is the selection of features
to learn weak classi?ers. more powerful discrim-
inant features is used instead of rectangular Haar features to im-
prove eye detection accuracy. To train a robust eye detector, training
data was collected from various sources. 500 pairs of eyes were collected from a database for study. training. only a left eye detector is trained In application,
due to the symmetry of eyes.
Eye Localization
The eye localization method follows a hierarchical princi-
ple. a face is detected first, then eyes are located inside
the detected face. Ada boost is used here too. multiple eyes detected around the pupil center. The ?nal eye localization is the average of the multiple detection results.
Eye Detection Validation
In one kind of validation experiments, a set of manually labeled eye positions were used. The performance of our eye detector is characterized by the eye detection rate and
eye localization error. The localization error is measured as the Euclidean distance between the detected eye posi-
tions and manual eye positions. In the second experiment, performance of eye detection was measured based on
its in?uence on face recognition accuracy of two standard
baseline methods: PCA and PCA together with LDA.
full report:
[attachment=1711]
