Deep Feature Fusion for Facial Expression Recognition

Facial expression recognition (FER) is widely incorporated in many applications such as healthcare, entertainment, virtual reality and robotics. However, practical applicability of FER are still limited as existing models are incapable of exploiting relevant features from the facial expressions. In this work, a deep model for FER is proposed which performs fusion of high and low level features. We have exploited transfer learning through pre-trained ResNet18 network which extract high level features and a shallow convolution block is exploited to extract low level features, then, fusion of these features is performed for classification. We have evaluated performance of the proposed model in terms of classification accuracy on JAFFE, CK+ and MUG datasets. Since, JAFFE and MUG are very small datasets, we have exploited data augmentation to increase their size by applying random geometric transformations. Experiments performed on these datasets indicate that the proposed architecture outperforms several state-of-the-art models.