Journal of Intelligent Computing Francisco Edvan Chaves, Thelmo Pontes de Araujo, José Everardo Bessa Maia 10 1 2019 10.6025/jic/2019/10/1/34-45 http://www.dline.info/jic/fulltext/v10n1/jicv10n1_3.pdf Many approaches in both feature extraction and classification have been proposed in order to build a robust automatic facial expression recognition system. The chosen features and classifiers are usually compared in limited scenarios. In this paper, 5 feature extractors (3 variations of Gabor filters, Local Binary Patterns, and Discrete Cosine Transform) and 4 classifiers (K-Nearest Neighbors, Support Vector Machine, Radial Basis Function Neural Network, and Naive Bayes) were combined and applied to three different datasets: JAFFE, Yale, and CK+. The combinations of feature extractors and classifiers were compared in more robust settings, being evaluated in each dataset separately and in 3 cross-database settings, to verify the technique’s generalization power. All experiments were performed in two validation scenarios: in the first one, the system tries to recognize the facial expression of a person already known to it in the training phase, using an image not present in the training set (leave-one-out); in the other scenario, all images of a certain individual are in the testing set, so the system tries to recognize the facial expression of a person unknown to it in the training phase (leave-one-subject-out). Accuracy results, as well as computational times, are presented, suggesting the combination of feature extractors and classifiers more suited for generalization