Gradient directional pattern: a robust feature descriptor for facial expression recognition

Presented is a novel local texture pattern, the gradient directional pattern (GDP), and an effective feature descriptor constructed with the GDP codes for facial expression recognition. The GDP operator encodes the texture information of a local region by quantising the gradient directional angles to form a binary pattern. The location and occurrence information of the GDP micro-patterns is then used as the facial feature descriptor. As the gradient operator can effectively enhance the edge information of an image, the resultant GDP features retain more information than grey-level based methods and describe the local image primitives in a more stable manner. Experiments with prototypic expression images from the Cohn-Kanade database shows the superiority of the GDP descriptor against some well-known appearance-based methods.