Facial expression recognition using deep learning

Facial expression recognition has become an increasingly important area of research in recent years. Neural network- based methods have made amazing progress in performing recognition-based tasks, winning competitions set up by various data science communities, and achieving high performance on many datasets. Miscellaneous regularization methods have been utilized by various researchers to help combat over-fitting, to reduce training time, and to generalize their models. In this paper, by applying the Haar Cascade classifier to crop faces and focus on the region of interest, we hypothesize that we would attain a fast convergence without using the whole image to analyze facial expressions. We also apply label smoothing and analyze its effect on the databases of CK+, KDEF, and RAF. The ResNet model has been employed as an example of a neural network model. Label smoothing has demonstrated an improvement of the recognition accuracy up to 0.5% considering CK+ and the KDEF databases. While the application of Haar Cascade has shown to decrease the achieved accuracy on KDEF and RAF databases with a small margin, fast convergence of the model has been observed.