Facial Expression Recognition: A New Dataset and a Review of the Literature

Abstract

One of the biggest challenges in computer vision and deep learning is recognizing the emotion based on facial expression. Besides the challenge of model development in Deep Learning (DL), the dataset is considered one of the main factors that plays a crucial role in producing highly accurate deep learning frameworks for pattern classification, features extraction, and emotion recognition from images. Despite the existence of several dedicated datasets for the FER, the desired accuracy was not being achieved due to the enormous size of the available dataset, where some of such are unfeasible for regular computers or cloud computing services. In contrast, there is a lack of the other datasets’ size, generalization, and quality. The objectives of this paper are two‐fold. First, reviews the existing datasets. Second, presents a significant new dataset for facial expression called Emot-FE. This dataset contains 276,305 images of facial expressions formulated in a single file in the form of xlsx format. Emot-FE dataset has been filtered, pre-processed, labeled, and classified based on the seven labels (emotion’s expressions). Thus, this dataset will be the largest in such format with high-quality images. This new dataset has been evaluated using VGGNet Convolutional Neural Networks where the recognition accuracy outperforms the previously achieved accuracies of the previous studies. This dataset is available for distribution to the researchers..