Eye detection based on rank order filter

Abstract

A novel eye detection algorithm based on rank order filter is proposed in this paper. Features such as eyeball is round and darker than surrounding pixels are widely used in eye detection. However quite often eyeball is distorted by iris reflection or other obstacles around eyeball. Rank order filter pair is designed to tackle these problems. One rank order filter is applied on central pixels to emphasize the darkness of the eyeball pixels and another rank order filter is applied on surrounding pixels to emphasize the brightness of the surrounding pixels. Then the difference of those two filter outputs is an important clue for eye detection, since regions near eyeball will yield large response. Those pixels of large response are selected as eyeball candidates, grouped and further verified by a series of geometric constraints. It is tested for 4095 face images with large variations in illuminations, hair styles, facial expressions and partial occlusions and detection rate as high as 98.97% is achieved.