Multimodal affect recognition in V-A-D space using deep learning

Now-a-days one of the most challenging problem is automatic affective state recognition. To identify emotions from physiological signals a knowledgeable and advanced learning algorithm is required that can represent high level abstraction of signals. This study presents the deployment of deep learning network (DLN) to determine unspecified feature interrelationship between input data. The DLN is implemented with two semi-supervised algorithms i.e.; Stack auto-encoder (SAE) and Deep Belief Network (DBN) individually on DEAP database for emotion state classification. A Bayesian inference classification based decision fusion method is used to improve the classification and the results has been improvised by almost 3% in case of valence and arousal of DBN and almost 15% and 6% in SAE than the individual classification results of SAE and DBN, however in case of dominance the result has been improved by almost 9% than SAE.