On the Convenience of Using 32 Facial Expressions to Recognize the 6 Universal Emotions

Emotion and facial expression recognition are a common topic in artificial intelligence. In particular, main efforts focus on constructing models to classify within the six universal emotions. In this paper, we present the first attempt to classify within 33 different facial expressions. We define and train a simple convolutional neural network with a low number of intermediate layers, to recognize the 32 facial expressions (plus the neutral one) contained in an extension of UIBVFED, a virtual facial expression dataset. We obtained a global accuracy of 0.8, which is comparable to the 0.79 accuracy we got when training the neural network with only the six universal emotions. Taking advantage of this trained model, we explore the approach of classifying the images within the six universal emotions, translating the facial expression predicted by the model into its associated emotion. With this novel approach, we reach an accuracy level of 0.95, a value comparable to the best results present in the literature with the plus of using a very simple neural network.