Facial expression analysis across databases

In this paper, we analyze the generalization ability of a facial expression recognition method and also study how to improve it across different databases. Since no common performance metric has been set up, people did experiments in different databases using different methods. So it is difficult for us to judge the performance and the generalization ability of an algorithm. And we also found that the performance is bad in general if we train a facial expression analysis algorithm on a database but test it on another database. So the problem is whether we could improve the generalization ability of a facial expression analysis system and how to improve it more. We extract Gabor features from face images and use SVM to classify the expressions. And we study the above problem using this algorithm on three databases (C-K+, JAFFE and TFEID). We mean to fuse these three databases to solve the problem. And three fusion methods are proposed respectively on the sample level, on the feature level and on the classifier level. Experiments show that the classifiers trained with the fusion methods are of better generalization ability. And the fusion method on the feature level reaches the highest accuracy.