Unsupervised learning of facial expression components

The face is one of the most important means of non-verbal communication. A lot of information can be gotten about the emotional state of a person just by merely observing their facial expression. This is relatively easy in face to face communication but not so in human computer interaction. Supervised learning has been widely used by researchers to train machines to recognise facial expressions just like humans. However, supervised learning has significant limitations one of which is the fact that it makes use of 'labelled' facial images to train models to identify facial actions. It is very expensive and time consuming to label face images. It takes about an hour to label a 5min video. In addition, more than 7000 distinct facial expressions can be created from a combination of different facial muscle actions [72]. The amount of labelled face images available for facial expression analysis is limited in supply and do not cover all the possible facial expressions. On the other hand, it is quite easy to collect or record a large amount of raw unlabelled data of people communicating without incurring so much cost. This research therefore used an unsupervised learning to identify facial action units. Similar appearance changes were identified from a large data set of face images without knowing beforehand which facial expressions they belong to and then the identified patterns or learned features were assigned to distinct facial expressions. For the purpose of comparison with the supervised learning approach, the learned features were used to train models to identify facial actions. Results showed that in 70% of the cases, the accuracy from the unsupervised learning model surpassed the performance of the supervised learning model