Igor Stankovic and Montri Karnjanadecha

pp. 461 - 468

Abstract

Studies of basic facial expression recognition have always shown different recognition rates for different emotional
expressions. Happiness and surprise detection easily exceed 90% recognition rates, while other basic emotions (i.e. sadness,
anger, fear, and disgust) produce much lower rates. In this paper we present a simple approach for reducing this gap,
increasing the recognition rates for the other four basic emotions, based on more closely analyzing the displacements of
extracted facial landmarks. With the use of a reference point in all the image sequences that represent an emotional expression,
calculations become more resistant to head movement errors, thereby reducing recognition errors. Also, emotions are
expressed differently across time so, besides first and peak frames, we also analyze the frames one third and two thirds along
each image sequence. Our results show great improvements in recognition, yielding total accuracies of 96.8%, and lowering
the recognition gap between facial expressions

Keywords

• facial expression recognition
• emotion recognition
• support vector machines
• the extended Cohn-Kanade database