Face Gesture Recognition Using Deep-Learning Models

This work compares face gesture recognition methods based on deep learning convolutional neural network (DCNN) architectures named DCNN1, DCNN2, DCNN3, DCNN4, and DCNN+Autoencoder, that maximize the classification performance on single and mixing datasets. We validated the proposed architectures on three different databases: Jaffe, CK+, and the combination of both databases (Jaffe & CK+) over a five-fold cross-validation strategy. The DCNN4, DCNN2, and DCNN+Autoencoder models achieved best performance mean accuracy scores of 95%, 94%, and 96% for the Jaffe, CK+, and Jaffe & CK+ databases, respectively. Moreover, according to the cross-entropy loss function, the selected models did not incur overfitting.