Study of the Algorithm for Assessing a Person’s Emotional State by a Face Image Using Actions Units

The paper is devoted to the study of algorithm for basic emotions (fear, happiness, anger, sadness, disgust, surprise) recognition from a single face image in the feature space of facial action units. For classification k-nearest neighbors (k-NN) algorithm was applied as well as our own knowledge-based algorithm, which uses emotions description, presented in facial action coding system. Testing was performed on three publicly available databases and on our own database. The results for k-NN algorithm demonstrate comparatively high recognition level (98.2 % and 97.1 % for 6 and 7 classes respectively on Cohn-Kanade database, 10-fold cross-validation), but also its sensitivity to the homogeneity of training and test sets, which was shown during cross-database testing. In contrast, the proposed algorithm is not based on machine learning, demonstrates slightly weaker but stable results, does not require much space for storing data, and is suitable for real-time implementation.