Recognition of facial expressions using HMM with continuous output probabilities

Facial expression recognition is an important technology fundamental to realize intelligent image coding systems and advanced man-machine interfaces in visual communication systems. In the computer vision field, many techniques have been developed to recognize facial expressions. However, most of these techniques are based on static features extracted from one or two still images. Those techniques are not robust against noise and cannot recognize subtle changes in facial expressions. In this paper we use hidden Markov models (HMM) with continuous output probabilities to extract a temporal pattern of facial motion. In order to improve the recognition performance, we propose a new feature obtained from wavelet transform coefficients. For the evaluation, we use 180 image sequences taken from three male subjects. Using these image sequences, the recognition rate for user trained mode achieved 98% compared with 84% using our previous method. The recognition rate for user independent mode achieved 84% when the expressions were restricted to four expressions.