Genetic Algorithms in Machine Learning Applied to Computer Vision: Facial Emotion Recognition

In the literature, the recognition of human facial emotions has been highly improved. However, the great challenge is to perform the training of efficient and fast algorithms due to the large volume of input data. This article presents a new approach to emotion classification. Instead of using information vectors of all pixels extracted from the image, the pre-sent study works with 7 Euclidean distances considering key points of the face: corners of the mouth, eyes and tip of the nose, in addition to the mean and variance of the facial region, totaling 9 input attributes. Thus, we aim verify by neurogenetic algorithms if it is possible to use one-dimensional data as information carrier to facial emotion classification. Two classification algorithms were performed separately: Support Vector Machine and Multi-Layer Perceptron. Then, a genetic algorithm was applied to select the more appropriate input at-tributes for each architecture. By this we analyze which variables were really relevant for the classification. The database of facial emotions used was the Warsaw Set of Emotional Facial Expression Pictures (WSEFEP) containing 210 images of 30 individuals. Altogether, the database presents images of 7 facial emotions: angry, disgusted, happy, neutral, sad, surprised, and afraid. The best result was found for the MLP architecture with 6 input attributes. The global average accuracy found was 51%, being 88% and 83% for the prediction of happy and surprised emotions, respectively.