A Study on Facial Expression Recognition Model Using an Adaptive Learning Capability

Abstract

The study of facial expression recognition for the purpose of man-machine emotional communication is attracting attention lately (Akamatsu, 2002a; Akamatsu, 2002b; Akamatsu, 2002c; Akamatsu, 2003; Fasel & Luettin, 2003; Pantic & Rothkrantz, 2000; Tian et al., 2001). Most facial expression recognition models that have been proposed eventually create some classifier based on the expression images taken during a short period of time and using them as base data for learning (Pantic et al., 2005; Gross, 2005). However, because so many facial expression patterns exist that a human being cannot make representations of all of them, it is difficult to obtain and retain all available patterns and use them as learning data in a short time. The actual facial expressions that change from one time to another would show the other patterns at other times that are not contained in the learning data. For that reason, it is thought to be difficult to maintain and recognize those facial expressions just as they are without changing them continuously for a long time using the same classifier that was created at the initial stage. For a facial expression recognition model to retain its high robustness along the time axis continuously for a long time, the classifier created at the initial stage should be evolved to be adaptive gradually over time. In other words, what is necessary for the model is that it retains existing knowledge (i.e. past facial patterns) and simultaneously learns to keep adding newly available knowledge (i.e. new facial patterns) as it becomes available. As described in this chapter, we propose a method of creating a facial expression recognition model that can offer the adaptive learning capability described above. In addition, its degree of usefulness is described. We will show it from results of experiments made for evaluation of the incremental learning capability that the model has. Thereby, we will examine that point specifically.