Facial Expression Recognition with Recurrent Neural Networks

Abstract

Abstract — This paper presents a complete system for auto-matic facial expression recognition. The Candide-3 face model is used in conjunction with a learned objective function for face model fitting. The resulting sequence of model parameters is then presented to a recurrent neural network for classification. The advantage of using a recurrent network is that the temporal dependencies present in the image sequences can be taken into account during the classification. Since the entire process is automatic, and the recurrent networks can be used to make online predictions, the system would be ideal for real-time recognition. This would make it suitable for the CoTeSys ‘coffee break ’ scenario, where guests must be recognised and served by robot waiters. Promising experimental results are presented on the Cohn-Kanade database. I.