(Left) Original image; (Middle) Corresponding spectral residual image; (Right) The original image convoluted with a horizontal-orientation and fifth-scale Gabor kernel.

The eyes are the most important facial landmarks on human face, for both human computer communication and face normalization. It is crucial to detect and locate the eyes accurately. This paper introduces a novel and simple method for eye localization. Based on the special gray lever distribution in the eye region, the spectral residual model is exploited to extract the rough eye region primarily. And then a coarse-to-fine searching strategy is adopted to locate the exact eye position. Compared with the Gabor transformation, the spectral residual transformation is lower in computational cost and more accurate in eye extraction. Some other experiments are carried out to investigate its robustness on variations of pose, orientation, scale, expression and illumination. More important, the proposed eye extraction method can be easily incorporated in other eye detection approaches as a combined processing module or employed in a real-time application.

Four face poses and their SRI (left to right): Turn right; Turn left; Tilt up; Tilt down.

Eye Localization results in the FERET faces

[1] F. Yang, J. Dai and D. Liu. A Novel Eye Localization Method Based on Spectral Residual Model. In Proc. of The 7th World Congress on Intelligent Control and Automation (WCICA'08), pages 6773-6777, 2008.