In this paper we present a new approach to the problem of detecting the eyes from a still facial image. Features such as color, edges, intensity and shape are incorporated in different stages of our method. First, the face region is extracted from the image by skin-color filter and window growing. Second, eyes are approximately located by projection of gradient features in different directions. Third, the eyes could be extracted if it is required by some specific applications. Part of the AR Database has been used for experiments and benchmarking. For subjects not wearing glasses, the eyes could be correctly located in 91.56% of the images and in over 90% of these images they could be accurately extracted, while for subjects wearing glasses, the success rate is somewhat lower.