Facial Expression Recognition In The Wild Using Bidirectional Convolutional Neural Network

The Static Facial Expressions In The Wild database (SFEW) contains unconstrained facial expressions close to the real world. In former research, current machine learning techniques are not robust enough for such an uncontrolled environment and it remains challenging nowadays. Coping with such task, we augment the state-of-art model which achieved the best performance for in the wild dataset and proposed two boosting algorithms of adding bidirectionality to convolution neural network based on the bidirectional neural network prototype, which is the first to integrate these two notions in literature. We also conducted experiments applying the decision fusion framework for classification, the proposed framework is trained simultaneously forward and backward, the final output is generated through voting mechanism. In this paper, two algorithms of adding bidirectionality to CNN are proposed, a framework for the facial expression recognition task (ensemble of HOG face detector and CNN with decision fusion and bidirectionality) is introduced and the classification result is listed, compared, and analyzed. The empirical results affirmed that the bidirectional boosting achieved good performance on the SFEW benchmark. Furthermore, some future works for precision improvement based on the existing deficiency of the current model are presented.