Um estudo da classificação de emoções em faces parcialmente oclusas

Abstract

With the increased use of intelligent technologies or machines by the average user, the need for the creation of human-centered computers and algorithms urges attention. In this sense, the inference of the emotional state of the human being with which a machine interacts with is fundamental for the establishment and regulation of affective connections, having applications in areas ranging from education to assistive technologies. The usage of intelligent algorithms, however, is starting to be regularized, and projects such as the General Data Protection Regulation guarantees for the user the right to an explanation of such automated inferences. This work is rooted in both topics mentioned: emotion estimation from facial images and the explicability of the intelligent algorithm in question. For this, we compared neural network architectures with and without attention mechanisms, trained separately in two datasets, with and without the presence of facial occlusions. At the end of the study, it is found that the addition of attention mechanisms contributes positively to emotion classification in most cases. Also, training with occluded images seems to expand the focus of a neural network on other input images. Regarding the CK+ dataset, the best model obtained an accuracy of 97.77% when trained and validated on images without occlusions and 90.01% when the data contained facial occlusions. In the cross-dataset evaluation, the best neural network model obtained 44.13% of accuracy for the JAFFE dataset.