An improved eye detection method based on statistical moments

In this paper the problem of eye detection in 2D grayscale images is addressed. The proposed method analyses the input face images in topographic format. The reason is to alleviate sensitivity of the algorithm to illumination and contrast changes. Invariant moments are used as robust features describing eye shape. A new strategy to select robust features based on their variance among training images is proposed. Using several complementary features such as existing of nose between eyes, some non-eye candidates are removed. Finally, a Bayesian classifier is used to select the most probable locations of eyes. The eye detection results show a higher detection rate and robustness compared to the existing methods. The performance rate has increased comparing to our previous algorithm presented.