Use of local entropy changes as a measure for identification of facial expressions

Facial expression recognition systems have developed rapidly. Most of the current systems are based on complex measures such as motion parameters, or models of muscular activity. On the other hand, entropy is a simple, yet powerful tool in discriminating activity in subsequent frames. In this study, we examined the use of local entropy changes in the identification of facial expressions. Six basic emotional facial expressions are collected from subjects as video sequences. The face is partitioned into rectangular boxes blindly, without considering the location of facial features. A Bayesian classifier is used to identify the expressions by looking at the patterns of entropy changes in the individual boxes. The results are satisfactory for the three expressions happy, surprised and sad, and exhibit consistency with the behavioral pattern reported in psychology literature.