Eye Localization

Having found a set of possible facial contours in the image, we proceed with the detection of the eyes within the face. When we refer to the eyes in this section, we are referring not only to the iris but rather the collection of contours forming the pupil, iris, eyelids, eyelashes, eyebrows and the shading around the eye orbit. This general eye region is a larger and more dominant structure as a whole than its individual subcomponents. Therefore, it is more stable and easier to detect as a whole. Reisfeld utilizes large operators that span regions larger than the eyelashes, iris and pupils to improve reliability in the eye detection [38]. Although the process of including the surrounding region improves robustness, it reduces accuracy since the contours of the eyebrows and eye orbit shading may have a center that does not coincide with the pupil's center. Some high quality, deformable model methods for detecting the iris and eyelids have been proposed by Yuille and others [45]. However, they can be computationally expensive and are not as robust as the large operators acting on the whole eye region. For example, if an individual in the image is squinting or if the image quality is poor, the iris will not be clearly visible and such high precision methods which search exclusively for an iris or eyelids might fail.

  

Figure 3.9: Generating eye search region from the facial contour

We shall use the knowledge acquired about the facial contour structure from the previous stage to constrain the search for eyes. The spatial search space will be restricted by a wide band perpendicular to the principal axis of the facial contour. Figure  shows the semi-elliptical model is composed of two axes intersecting at the model's center. Beginning from the center of the model, we move up the principal axis a distance dy₁ and down the principal axis a distance dy₂, where we form two parallel lines that contain the band of interest. The lengths dy₁ and dy₂ can be selected by analyzing several faces from a database and noting the relative placement of eyes with respect to the detected facial contour. The setting $dy_{1}=0.15 \times b$ and $dy_{2}=0.45 \times b$ generates a band that more than adequately covers the eyes. Although this setting seems rather conservative, a wide margin of error is needed since the facial contour might surround the face or the whole head. Either the inner hair-line or the top of the head could be traced by the boundary of our facial contour, so a narrow eye-band might be unsafe. Figure  shows the eye bands or eye spatial search space as brightened strips superimposed upon the original intensity images.

  

Figure 3.10: Isolating the eye search regions